A method and related device for predicting anticancer peptides

Abstract

The present invention discloses an anticancer peptide prediction method and related devices. The method obtains the anticancer peptide sequence data to be identified and inputs it into a trained prediction model to obtain the similarity information, physicochemical characteristic information, graphic and statistical characteristic information and ProtBERT input format information of the anticancer peptide sequence data to be identified; a convolutional neural network is used to process the graphic and statistical characteristic information to obtain the first eigenvector, a bidirectional long short-term memory network is used to process the concatenated matrix of the similarity information and the physicochemical characteristic information to obtain the second eigenvector, and a ProtBERT model is used to process the ProtBERT input format information using an attention mechanism to obtain the third eigenvector; the first, second and third eigenvectors are concatenated, and the recognition result is obtained under the action of a fully connected layer, a ReLU activation function, a dropout and a sogmoid function. The present invention can improve the recognition and prediction of anticancer peptides.

Description

Technical Field

[0001] The present invention relates to the technical field of peptide recognition, and in particular to an anticancer peptide prediction method and related devices. Background Art

[0002] Anticancer peptides are a class of peptide substances with anticancer activity. They can directly act on cancer cells, inhibit their growth, spread or induce cancer cell death. These peptide substances can be derived from nature or obtained through synthesis or bioengineering technology. Therefore, the prediction and classification of anticancer peptides is a crucial research topic.

[0003] Existing technologies often fail to distinguish between different types of feature encoding methods, and usually simply concatenate different types of features without considering using specific neural networks to process the corresponding feature encoding. Secondly, the feature encoding of manually converted sequences has certain limitations in obtaining more complex and deeper information. In addition, the existing anticancer peptide recognition methods have low recognition accuracy and poor generalization ability. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes an anticancer peptide prediction method and related devices. In the method and related devices, different processing methods are used for different types of feature encodings, which also ensures the independence between them and the effectiveness of the features, and adds the ProtBERT model in the prediction of anticancer peptides, thereby improving the correct recognition and prediction of anticancer peptide sequences.

[0005] In order to achieve the above object, the technical solution of the present invention is as follows:

[0006] A method for predicting anticancer peptides, characterized in that it comprises the following steps:

[0007] Acquiring sequence data of anticancer peptides to be identified;

[0008] The anticancer peptide sequence data to be identified is input into a trained prediction model, wherein the prediction model includes a sequence processing module, a multi-module processing feature encoding module and a feature fusion classification module, wherein the sequence processing module is used to process the anticancer peptide sequence data to be identified separately to obtain similarity information, physicochemical feature information, graphic and statistical feature information and ProtBERT input format information; the multi-module processing feature encoding module includes three channels, the first channel uses a convolutional neural network to process the graphic and statistical feature information to obtain a 512-dimensional first feature vector; the second channel uses a bidirectional long short-term memory network to process the concatenated matrix of similarity information and physicochemical feature information, and reduces the dimension of the processed data to obtain a 512-dimensional second feature vector; the third channel uses the ProtBERT model and uses the attention mechanism to process the ProtBERT input format information to obtain a 512-dimensional third feature vector; the feature fusion classification module is used to concatenate the first feature vector, the second feature vector and the third feature vector, and obtain the probability of whether the anticancer peptide sequence data to be identified is an anticancer peptide under the joint action of the fully connected layer, the ReLU activation function, the dropout and the sogmoid function.

[0009] Preferably, the process of constructing the trained prediction model comprises the following steps:

[0010] Peptide data were collected to construct a peptide data set, wherein the peptide data set included sequences and labels of samples, wherein the sequence of each sample consisted of twenty kinds of amino acids, which were represented by the English letters A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W and Y, respectively; the labels included positive sample labels and anticancer peptides as 1, and negative sample labels as 0;

[0011] The prediction model is trained using a peptide data set, and the model parameters are updated using a back-propagation algorithm until the training end condition is met to obtain a trained prediction model.

[0012] Preferably, the process of acquiring the physicochemical characteristic information comprises the following steps:

[0013] Use 12 physicochemical property data of 20 amino acids selected from the AAindex database;

[0014] The sequence length is fixed to 50 characters. If the sequence exceeds 50 characters, the first 50 data are selected; if the sequence is less than 50 characters, the character X is added to the end of the sequence;

[0015] Based on the attribute feature values ​​corresponding to the AAindex database and the encoding representation of the X character corresponding to 0, the sequence is encoded into a 50*12 matrix.

[0016] Preferably, the process of obtaining the similarity information comprises the following steps:

[0017] The sequence length is fixed to 50 characters. If the sequence exceeds 50 characters, the first 50 data are selected; if the sequence is less than 50 characters, the character X is added to the end of the sequence;

[0018] Based on the alignment of amino acid sequences, the identity between the two sequences did not exceed 62%, and the sequences were encoded into a 50*20 matrix based on the substitution scores of 20 amino acids in BLOSUM62.

[0019] Preferably, the process of acquiring the graphic and statistical feature information comprises the following steps:

[0020] 158 physicochemical properties of 20 amino acids were selected from the AAindex database, and the amino acids were ranked using the data of each set of physicochemical properties;

[0021] Under each physicochemical property data: Use formula (1) to determine the positions of the 20 amino acids at the bottom of a right cone with a height of 1. The formula is as follows:

[0022]

[0023] Where Ωi represents each amino acid represented by the order of physicochemical properties;

[0024] Since the number of basic amino acids is 20, the number of amino acid pairs is 20x20=400. The 400 amino acid pairs are mapped to the cone using formula (2), as follows:

[0025]

[0026] Where ΩiΩj represents each amino acid pair;

[0027] In the sequence data, the position of the first amino acid is P0 (0, 0, 0), and the position of the second amino acid is P1 (x1, y1, z1). Using formula (3), the subsequent amino acids are placed in the positions of Pi (xi, yi, zi) in sequence; a 3D path is made on the surface of the cone; 158 different cones are drawn for 158 physical and chemical properties, and 158 3D curves are generated. The formula is as follows:

[0028]

[0029] Where Ni represents the amino acid at position i in the sequence data, and f(Ω1Ω2) represents the frequency of the amino acid pairs of the first i amino acids in the sequence;

[0030] For each 3D curve, a non-negative symmetric matrix of size L*L is constructed, where L represents the length of the sequence, and the off-diagonal elements M(i, j) (i≠j) are calculated as the ratio of the geometric distance between point Pi and point Pj along the curve to their Euclidean distance; all diagonal elements are set to 0; in order to eliminate the deviation caused by different sequence lengths, the eigenvalues ​​of each matrix are normalized relative to the sequence length, so the 158 matrices of each sequence produce 158 different values;

[0031] AAC was used to represent the frequency of each amino acid in the sequence, generating 20 values; DPC was used to represent the frequency of each amino acid pair, obtaining 400 values; the 158 different values ​​were combined with the numerical results of AAC and DPC, so each sequence data was represented by a 578-dimensional feature vector.

[0032] Preferably, the process of obtaining the ProtBERT input format information comprises the following steps:

[0033] Fix the sequence length to 50 characters;

[0034] The vectorized text tool Tokenizer is used to encode, add the CLS start symbol and fill the end with PAD characters when the anticancer peptide sequence is less than 50 characters, and obtain the input format code. The input format code includes input_ids, which represents the index of each amino acid in the sequence in the vocabulary; token_type_ids, which represents the number of the current sentence in the input, and all 0s are assigned to the tuple; attention_mask, which represents whether the current position corresponds to a valid character, which is 1 if valid, otherwise 0.

[0035] Preferably, the processing process of the multi-module processing feature encoding module includes the following steps:

[0036] Input the graph-based and statistical feature information into the convolutional neural network, use one-dimensional convolution to move in only one direction to obtain the convolution feature; use the activation function to increase the nonlinear factor after the convolution output to enhance the channel's expressive power; use the flatten layer to expand the multidimensional data to one dimension and pass it through the fully connected layer and Dropout function to output the first 512-dimensional feature vector;

[0037] The similarity information and the physical and chemical characteristic information are concatenated and input into the bidirectional long short-term memory network, and the output of the last time step of the forward direction is concatenated with the output of the first time step of the backward direction; LayNorm is used for normalization, and a 512-dimensional second eigenvector is output;

[0038] The ProtBERT input format information is fed into the pre-trained ProtBERT model and the attention mechanism is applied to output the third eigenvector of 512 dimensions.

[0039] Based on the above content, the present invention also discloses an anticancer peptide prediction system, including: a data acquisition module, a sequence processing module, a multi-module processing feature encoding module and a feature fusion classification module, wherein:

[0040] The data acquisition module is used to obtain the anticancer peptide sequence data to be identified;

[0041] The sequence processing module is used to encode the anticancer peptide sequence data to be identified to obtain similarity information, physical and chemical characteristic information, graphic and statistical characteristic information and ProtBERT input format information;

[0042] The multi-module processing feature encoding module includes three channels. The first channel uses a convolutional neural network to process the graphic and statistical feature information to obtain a 512-dimensional first feature vector; the second channel uses a bidirectional long short-term memory network to process the concatenated matrix of similarity information and physical and chemical feature information, and reduces the dimension of the processed data to obtain a 512-dimensional second feature vector; the third channel uses the ProtBERT model and uses the attention mechanism to process the ProtBERT input format information to obtain a 512-dimensional third feature vector;

[0043] The feature fusion classification module is used to concatenate the first feature vector, the second feature vector and the third feature vector, and obtain the probability of whether the anticancer peptide sequence data to be identified is an anticancer peptide under the joint action of the fully connected layer, the ReLU activation function, the dropout and the sogmoid function.

[0044] Based on the above content, the present invention further discloses a computer device, including: a memory for storing a computer program; and a processor for implementing any of the above methods when executing the computer program.

[0045] Based on the above content, the present invention further discloses a readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, any of the methods described above is implemented.

[0046] Based on the above technical solution, the beneficial effects of the present invention are as follows: the present invention discloses an anticancer peptide prediction method and related devices. The method obtains the anticancer peptide sequence data to be identified; the anticancer peptide sequence data to be identified is input into a trained prediction model, the prediction model includes a sequence processing module, a multi-module processing feature encoding module and a feature fusion classification module, wherein the sequence processing module is used to process the anticancer peptide sequence data to be identified separately to obtain similarity information, physicochemical feature information, graphic and statistical feature information and ProtBERT input format information; the multi-module processing feature encoding module includes three channels, the first channel uses a convolutional neural network to process the graphic and statistical feature information to obtain a 512-dimensional first feature vector; the second channel A bidirectional long short-term memory network is used to process the concatenated matrix of similarity information and physical and chemical characteristic information, and the processed data is reduced in dimension to obtain a second eigenvector of 512 dimensions; the third channel uses the ProtBERT model and uses the attention mechanism to process the ProtBERT input format information to obtain a third eigenvector of 512 dimensions; the feature fusion classification module is used to concatenate the first eigenvector, the second eigenvector and the third eigenvector, and under the joint action of the fully connected layer, the ReLU activation function, the dropout and the sogmoid function, the probability of whether the anticancer peptide sequence data to be identified is an anticancer peptide is obtained. The present invention uses different processing strategies for different types of feature encoding methods, and also ensures their independence and the validity of the features. The use of a pre-trained ProtBERT model enables the prediction model to learn more complex and deeper features, effectively improving the overall performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of a method for predicting anticancer peptides in one embodiment;

[0048] Figure 2 is a schematic diagram of sequence length processing in an embodiment;

[0049] Figure 3 is a schematic diagram of a first channel processing process of a multi-module processing feature encoding module in an embodiment;

[0050] Figure 4 is a schematic diagram of a second channel processing process of a multi-module processing feature encoding module in an embodiment;

[0051] Figure 5 The diagram is a schematic diagram of the third channel processing process of a multi-module feature encoding module in an embodiment. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0053] like Figure 1 As shown, this embodiment provides an anticancer peptide prediction method, and the specific implementation steps are as follows: obtaining anticancer peptide sequence data to be identified; inputting the anticancer peptide sequence data to be identified into a trained prediction model, wherein the prediction model includes a sequence processing module, a multi-module processing feature encoding module and a feature fusion classification module, wherein the sequence processing module is used to process the anticancer peptide sequence data to be identified separately to obtain similarity information, physicochemical feature information, graphic and statistical feature information and ProtBERT input format information; the multi-module processing feature encoding module includes three channels, the first channel uses a convolutional neural network to process the graphic and statistical feature information to obtain 512 The first eigenvector of dimension 512 is obtained by using a bidirectional long short-term memory network to process the concatenation matrix of similarity information and physical and chemical characteristic information in the second channel, and the processed data is subjected to dimensionality reduction to obtain a second eigenvector of dimension 512; the third channel adopts the ProtBERT model and uses the attention mechanism to process the ProtBERT input format information to obtain a third eigenvector of dimension 512; the feature fusion classification module is used to concatenate the first eigenvector, the second eigenvector and the third eigenvector, and under the joint action of the fully connected layer, the ReLU activation function, the dropout and the sogmoid function, obtain the probability of whether the anticancer peptide sequence data to be identified is an anticancer peptide.

[0054] In one embodiment, a method for predicting anticancer peptides further includes a process for constructing a trained prediction model, including the following steps:

[0055] Peptide data were collected to construct a peptide data set, wherein the peptide data set included sample sequences and labels, wherein the sequence of each sample consisted of twenty kinds of amino acids, represented by the English letters A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W and Y, respectively; the labels included positive sample labels and anticancer peptides as 1, and negative sample labels as 0.

[0056] In this embodiment, an anticancer peptide-related data set is collected, and the data set contains the sequence of each sample (composed of 20 naturally occurring amino acids, each represented by 20 English letters (ACDEFGHIKLMNPQRSTVWY) (for example, FVQWFSKFLGRIL is an anticancer peptide sequence)) and a label (where the positive sample label is 1 (anticancer peptide) and the negative sample label is 0)

[0057] The prediction model is trained using a peptide data set, and the model parameters are updated using a back-propagation algorithm until the training end condition is met to obtain a trained prediction model.

[0058] In one embodiment, a method for predicting anticancer peptides further includes a processing process of a sequence processing module, including the following steps:

[0059] (1) Sequence feature encoding method - AAindex to obtain physicochemical feature information: 12 physicochemical properties are selected from AAindex. Each amino acid is encoded by 12 data. First, the sequence length is fixed to 50 characters. If the sequence exceeds 50 characters, the first 50 data are selected; if the sequence is less than 50 characters, the character X is added to the end of the sequence. See Figure 2 ; Then, based on the attribute feature values ​​corresponding to the AAindex database and the encoding representation of the X character corresponding to 0, the sequence is encoded into a 50*12 matrix.

[0060]

[0061] (2) Sequence feature encoding method - BLOSUM62 obtains similarity information: The encoding matrix is ​​obtained by using the comparison scoring mechanism between protein sequences. Similar to AAindex, each amino acid is encoded by 20 data in the similarity encoding. The sequence length is fixed to 50 characters. If the sequence exceeds 50 characters, the first 50 data are selected; if the sequence is less than 50 characters, the character X is added to the end of the sequence; based on the comparison of amino acid sequences, the identity between the two sequences does not exceed 62%. Based on the substitution scores of 20 amino acids in BLOSUM62, each sequence is encoded as a 50*20 feature matrix.

[0062]

[0063] (3) Sequence feature encoding method - FEGS obtains information based on graphics and statistical features. The specific encoding steps are as follows:

[0064] In the AAindex database, 158 physicochemical properties of 20 amino acids were selected, and the amino acids were sorted using the data of each set of physicochemical properties. Under each set of physicochemical property data: using formula (1), the positions of the 20 amino acids were determined at the bottom of a right cone with a height of 1, and the formula is as follows:

[0065]

[0066] Where Ωi represents each amino acid represented by the amino acid ranking under a physicochemical property.

[0067] Since the number of basic amino acids is 20, the number of amino acid pairs is 20x20=400. The 400 amino acid pairs are mapped to the cone using formula (2), as follows:

[0068]

[0069] where ΩiΩj represents each amino acid pair.

[0070] In the original sequence of the anticancer peptide, the position of the first amino acid is set to P0 (0, 0, 0), and the position of the second amino acid is set to P1 (x1, y1, z1). Using formula (3), the subsequent amino acids are placed in the Pi (xi, yi, zi) positions in sequence. After calculating each amino acid in the sequence, it is wound around a cone surface to form a 3D path, as shown in the figure below, which represents a graph generated by an anticancer peptide sequence. Each sequence generates 158 3D curves based on 158 physical and chemical property values, and the formula is as follows:

[0071]

[0072] Where Ni represents the amino acid at position i in the anticancer peptide sequence, and f(Ω1Ω2) represents the frequency of the amino acid pairs of the first i amino acids in the sequence.

[0073] For each 3D curve, a non-negative symmetric matrix of size L*L is constructed, where L represents the length of the sequence. The off-diagonal element M(i, j) (i≠j) is calculated as the ratio of the geometric distance between point i and point j along the curve to their Euclidean distance. All diagonal elements are set to 0. In order to eliminate the deviation caused by different sequence lengths, each matrix is ​​reduced in dimension to obtain the main eigenvalue. Finally, using the above three steps, each anticancer peptide sequence is represented by 158 numerical values.

[0074] In addition, we also combined AAC and DPC. AAC represents the frequency of each amino acid in the anticancer peptide sequence, so 20 values ​​are generated. DPC represents the frequency of each amino acid pair in the anticancer peptide sequence, so 400 values ​​are obtained. Therefore, under this sequence feature encoding, each anticancer peptide sequence is represented by 578 values, forming a 578-dimensional feature vector.

[0075] (4) Sequence conversion to ProtBERT input format. After the original sequence enters the tokenizer function, a dictionary is obtained, which is the input format

[0076] {

[0077] "input_id": [represents the index of each amino acid in the sequence in the vocabulary, and adds special symbols, such as the decimal point [CLS] and the filler [PAD]]

[0078] "token_type-ids": [indicates the number of the current sentence in the input. In this article, we treat each sequence as a separate sentence input, so all the tuples are 0]

[0079] "attention_mask": [indicates whether the input_id corresponding to the current position is a valid character, 1 if valid, otherwise 0]

[0080] }

[0081] Tokenizer function, for example, the input sequence is (length set to 50): GLFDIVKKIAGHIAGSI

[0082] The sequence is converted to [CLS][G][F][D][I][V][K][K][I][A][G][H][I][A][G][s][I][PAD][PAD][PAD][PAD][PAD][PAD]......x50The output of this function is {

[0083] "input_id": [2, 7, 19, 14, 11, 8, 12, 12, 11, 6, 7, 22, 11, 6, 7, 10, 11, 0, 0, 0, 0, 0, 0...x50]

[0084] "token_type_ids": [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...x50]

[0085] "attention_mask": [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0...x50]

[0086] }

[0087] In one embodiment, a method for predicting anticancer peptides further includes a processing process of a multi-module feature coding module, wherein the multi-module feature coding module includes three channels, including the following steps:

[0088] (1) See Figure 3 In the CNN channel module of the first channel: input the graph-based and statistical feature information into the convolutional neural network, use a one-dimensional convolutional network to process the manually extracted graph-based and statistical feature information, use a one-dimensional convolution kernel to move in only one direction, use an activation function after the convolution output to increase the nonlinear factor to enhance the channel's expressiveness, and then use the flatten layer to expand the multi-dimensional data to one dimension. Finally, through a fully connected layer, the output is converted into a 512-dimensional first feature vector.

[0089] (2) See Figure 4, in the BiLSTM module of the second channel: a bidirectional long short-term memory network (BiLSTM) is used to process the combination of the two feature matrices of similarity information and physical and chemical feature information, and the output of the last time step of the BiLSTM output forward is concatenated with the output of the first time step of the backward. LayNorm is used for normalization at the same time, and finally the module outputs a 512-dimensional second feature vector.

[0090] (3) See Figure 5 , the ProtBERT input format information after sequence conversion is used as the input of the ProtBERT channel module of the third channel. The pre-trained ProtBERT model is used. The model has a 30-layer network architecture, and each layer contains a multi-head attention mechanism and a fully connected layer. The multi-head attention mechanism allows the model to focus on different parts of the sequence when processing the sequence, thereby capturing the complex features and patterns in the protein sequence. In order to use the ProtBERT model, the sequence is converted into the input format of the model. During the entire model training process, the weight parameters of the last 4 layers of ProtBERT are opened, and the weight parameters of other layers are locked. During the back propagation process, only the weight parameters of the last 4 layers of this module are modified. This can reduce the number of model parameters, prevent overfitting, and make training more effective. After fine-tuning on the peptide dataset, a 512-dimensional third feature vector is obtained after the convolution layer and attention mechanism.

[0091] In one embodiment, a method for predicting anticancer peptides also includes a processing process of a feature fusion classification prediction module, including the following steps: concatenating the first eigenvector, the second eigenvector and the third eigenvector, using a fully connected layer for classification processing, adding a ReLU activation function and dropout, and under the probability prediction of binary cross entropy, treating values ​​greater than a threshold as positive samples and values ​​less than a threshold as negative samples.

[0092] Experimental description: Under the same test set, the comparison results of the proposed prediction model with other models (ACP-MHCNN model, ACP-DL model, AntiCP 2.0 model, TriNet model) are as follows:

[0093]

[0094] Through experiments, we can see that the proposed prediction model has certain improvements in accuracy, Matthews coefficient, specificity and sensitivity indicators compared with existing models (ACP-MHCNN model, ACP-DL model, AntiCP 2.0 model, TriNet model). This also further verifies that the use of pre-trained ProtBERT model can enable the model to learn more complex and deeper features, effectively improving the overall performance of the model. At the same time, in order to give full play to the potential of these feature encodings, different processing strategies are used for different types of feature encoding methods to ensure their independence and the effectiveness of the features.

[0095] The above are only preferred embodiments of the embodiments of the present application and are not intended to limit the embodiments of the present application. For those skilled in the art, the embodiments of the present application may have various modifications and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the embodiments of the present application shall be included in the protection scope of the embodiments of the present application.

Claims

1. A method for predicting anticancer peptides, characterized in that: The steps include: Acquiring sequence data of anticancer peptides to be identified; The anticancer peptide sequence data to be identified is input into a trained prediction model, wherein the prediction model includes a sequence processing module, a multi-module processing feature encoding module and a feature fusion classification module, wherein the sequence processing module is used to process the anticancer peptide sequence data to be identified separately to obtain similarity information, physicochemical feature information, graphic and statistical feature information and ProtBERT input format information; the multi-module processing feature encoding module includes three channels, the first channel uses a convolutional neural network to process the graphic and statistical feature information to obtain a 512-dimensional first feature vector; the second channel uses a bidirectional long short-term memory network to process the concatenated matrix of similarity information and physicochemical feature information, and reduces the dimension of the processed data to obtain a 512-dimensional second feature vector; the third channel uses the ProtBERT model and uses the attention mechanism to process the ProtBERT input format information to obtain a 512-dimensional third feature vector; the feature fusion classification module is used to concatenate the first feature vector, the second feature vector and the third feature vector, and obtain the probability of whether the anticancer peptide sequence data to be identified is an anticancer peptide under the joint action of the fully connected layer, the ReLU activation function, the dropout and the sogmoid function.

2. The method for predicting anticancer peptides according to claim 1, characterized in that: The process of constructing the trained prediction model includes the following steps: Peptide data were collected to construct a peptide data set, wherein the peptide data set included sequences and labels of samples, wherein the sequence of each sample consisted of twenty kinds of amino acids, which were represented by the English letters A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W and Y, respectively; the labels included positive sample labels and anticancer peptides as 1, and negative sample labels as 0; The prediction model is trained using a peptide data set, and the model parameters are updated using a back-propagation algorithm until the training end condition is met to obtain a trained prediction model.

3. The method for predicting anticancer peptides according to claim 2, characterized in that: The process of obtaining the physicochemical characteristic information comprises the following steps: Use 12 physicochemical property data of 20 amino acids selected from the AAindex database; The sequence length is fixed to 50 characters. If the sequence exceeds 50 characters, the first 50 data are selected; if the sequence is less than 50 characters, the character X is added to the end of the sequence; Based on the attribute feature values ​​corresponding to the AAindex database and the encoding representation of the X character corresponding to 0, the sequence is encoded into a 50*12 matrix.

4. The method for predicting anticancer peptides according to claim 3, characterized in that: The process of obtaining the similarity information includes the following steps: The sequence length is fixed to 50 characters. If the sequence exceeds 50 characters, the first 50 data are selected; if the sequence is less than 50 characters, the character X is added to the end of the sequence; Based on the alignment of amino acid sequences, the identity between the two sequences did not exceed 62%, and the sequences were encoded into a 50*20 matrix based on the substitution scores of 20 amino acids in BLOSUM62.

5. The method for predicting anticancer peptides according to claim 4, characterized in that: The acquisition process based on graphics and statistical feature information includes the following steps: 158 physicochemical properties of 20 amino acids were selected from the AAindex database, and the amino acids were ranked using the data of each set of physicochemical properties; Under each physicochemical property data: Use formula (1) to determine the positions of the 20 amino acids at the bottom of a right cone with a height of 1. The formula is as follows: Where Ωi represents each amino acid represented by the order of physicochemical properties; Since the number of basic amino acids is 20, the number of amino acid pairs is 20x20=400. The 400 amino acid pairs are mapped to the cone using formula (2), as follows: Where ΩiΩj represents each amino acid pair; In the sequence data, the position of the first amino acid is P0 (0, 0, 0), and the position of the second amino acid is P1 (x1, y1, z1). Using formula (3), the subsequent amino acids are placed in the positions of Pi (xi, yi, zi) in sequence; a 3D path is made on the surface of the cone; 158 different cones are drawn for 158 physical and chemical properties, and 158 3D curves are generated. The formula is as follows: Where Ni represents the amino acid at position i in the sequence data, and f(Ω1Ω2) represents the frequency of the amino acid pairs of the first i amino acids in the sequence; For each 3D curve, a non-negative symmetric matrix of size L*L is constructed, where L represents the length of the sequence, and the off-diagonal elements M(i, j) (i≠j) are calculated as the ratio of the geometric distance between point Pi and point Pj along the curve to their Euclidean distance; all diagonal elements are set to 0; in order to eliminate the deviation caused by different sequence lengths, the eigenvalues ​​of each matrix are normalized relative to the sequence length, so the 158 matrices of each sequence produce 158 different values; AAC was used to represent the frequency of each amino acid in the sequence, generating 20 values; DPC was used to represent the frequency of each amino acid pair, obtaining 400 values; the 158 different values ​​were combined with the numerical results of AAC and DPC, so each sequence data was represented by a 578-dimensional feature vector.

6. The method for predicting anticancer peptides according to claim 5, characterized in that: The process of obtaining the ProtBERT input format information includes the following steps: Fix the sequence length to 50 characters; The vectorized text tool Tokenizer is used to encode, add the CLS start symbol, and fill the end with PAD characters when the anticancer peptide sequence is less than 50 characters, to obtain the input format code, wherein the input format code includes input_ids indicating the index of each amino acid in the sequence in the vocabulary; token_type_ids indicating the number of the current sentence in the input, and all the token_type_ids tuples are assigned 0; attention_mask indicates whether the current position corresponds to a valid character, 1 if valid, otherwise 0.

7. The method for predicting anticancer peptides according to claim 1, characterized in that: The processing process of the multi-module processing feature encoding module includes the following steps: Input the graph-based and statistical feature information into the convolutional neural network, use one-dimensional convolution to move in only one direction to obtain the convolution feature; use the activation function to increase the nonlinear factor after the convolution output to enhance the channel's expressive power; use the flatten layer to expand the multidimensional data to one dimension and pass it through the fully connected layer and Dropout function to output the first 512-dimensional feature vector; The similarity information and the physical and chemical characteristic information are concatenated and input into the bidirectional long short-term memory network, and the output of the last time step of the forward direction is concatenated with the output of the first time step of the backward direction; LayNorm is used for normalization, and a 512-dimensional second eigenvector is output; The ProtBERT input format information is fed into the pre-trained ProtBERT model and the attention mechanism is applied to output the third eigenvector of 512 dimensions.

8. An anticancer peptide prediction system, characterized in that: include: Data acquisition module, sequence processing module, multi-module processing feature encoding module and feature fusion classification module, among which, The data acquisition module is used to obtain the anticancer peptide sequence data to be identified; The sequence processing module is used to encode the anticancer peptide sequence data to be identified to obtain similarity information, physical and chemical characteristic information, graphic and statistical characteristic information and ProtBERT input format information; The multi-module processing feature encoding module includes three channels, the first channel uses a convolutional neural network to process the graphic and statistical feature information to obtain a 512-dimensional first feature vector; The second channel uses a bidirectional long short-term memory network to process the concatenated matrix of similarity information and physical and chemical characteristic information, and reduces the dimension of the processed data to obtain a 512-dimensional second eigenvector; the third channel uses the ProtBERT model and uses the attention mechanism to process the ProtBERT input format information to obtain a 512-dimensional third eigenvector; The feature fusion classification module is used to concatenate the first feature vector, the second feature vector and the third feature vector, and obtain the probability of whether the anticancer peptide sequence data to be identified is an anticancer peptide under the joint action of the fully connected layer, the ReLU activation function, the dropout and the sogmoid function.

9. A computer device, characterized in that: The device comprises: a memory for storing a computer program; A processor, configured to implement the method according to any one of claims 1 to 7 when executing the computer program.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images