Anti-cancer peptide prediction method based on bidirectional long short-term memory network and feature fusion

Abstract

The present application relates to the technical field of anticancer peptide prediction, and particularly relates to an anticancer peptide prediction method based on a bidirectional long short-term memory network and feature fusion, comprising: reading four reference peptide sequence datasets, and performing amino acid composition analysis on the datasets; performing feature extraction on the datasets by Bi-LSTM to generate Bi-LSTM feature vectors; performing feature extraction on five amino acid feature vectors by a fully connected neural network; performing feature fusion on the feature vectors by a Concatenate algorithm, obtaining a probability score through a fully connected layer with 1 unit and a Sigmoid activation function, and distinguishing anticancer peptides and non-anticancer peptides through the score. The present application realizes anticancer peptide prediction with high accuracy, high Matthews correlation coefficient, high sensitivity, high specificity and high area under the ROC curve.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of anti-cancer peptide prediction, and particularly relates to an anti-cancer peptide prediction method based on a bidirectional long short-term memory network and feature fusion. BACKGROUND

[0002] The discovery of anti-cancer peptides (ACP) broadens people's vision of the road to fight cancer, and its specificity and tumor's inability to develop resistance to it solve the side effects brought by some traditional anticancer treatments, and it is expected to become an alternative treatment for cancer; an anti-cancer peptide is usually composed of 5-40 amino acids; in order to further understand the mechanism of action of anti-cancer peptides, there are many biological experimental methods for identifying anti-cancer peptides. For example, Vidal et al. determined the peptide cocktail against intracellular tumor proteins through a yeast two-hybrid system, and Peelle et al. found a new type of localization peptide that is not cell type specific through mammalian cell screening; however, these identification methods are time-consuming, expensive and very complex, and it is difficult to achieve them in a high-throughput manner, so fast and effective identification of anti-cancer peptides is particularly important.

[0003] Wu et al. proposed a PTPD model, which uses k-mer and Word2vec (word vector) extracted feature vectors to input a convolutional neural network (CNN) to predict peptides; Rao et al. applied a graph convolution network (GCN) to the prediction of anti-cancer peptides and proposed an ACP-GCN model; however, these deep learning methods only consider the original sequence information and physical and chemical properties of amino acids, ignoring the long-term correlation information of anti-cancer peptides at the time level, and cannot identify anti-cancer peptides at low cost, quickly and efficiently. SUMMARY

[0004] In view of the shortcomings of the existing algorithm, the present application realizes high accuracy, high Matthew correlation coefficient, high sensitivity, high specificity and high area under the ROC curve for anti-cancer peptide prediction.

[0005] The technical solution adopted by the present application is: an anti-cancer peptide prediction method based on a bidirectional long short-term memory network and feature fusion includes the following steps:

[0006] Step 1, read four benchmark peptide sequence datasets, and analyze the amino acid composition of the datasets;

[0007] Step 2, extract features from the dataset through a bidirectional long short-term memory network (Bi-LSTM) to generate a Bi-LSTM feature vector;

[0008] Further, step 2 includes:

[0009] Step 2.1, in order to input the peptide sequence into Bi-LSTM, first, the primary letter sequence of the peptide is digitally coded according to the amino acid alphabet, that is, 20 basic amino acids are assigned numbers 1-20, and the peptide sequence with insufficient length is filled with 0 to ensure that all peptide sequences have the same length;

[0010] Step 2.2, the input digital code is converted into a 64-dimensional vector representation by the embedding layer of Bi-LSTM;

[0011] Step 2.3, Bi-LSTM extracts features from the input 64-dimensional vector, and the specific composition of Bi-LSTM is: input x t , cell state C t , temporary cell state hidden state h t , forget gate f t , memory gate i t , output gate O t ;

[0012] Bi-LSTM is composed of forward and backward long short-term memory network layers, each layer is composed of a memory unit and a 64-dimensional hidden unit;

[0013] Forget gate (select information to be forgotten):

[0014] f t =σ(W f ·[h t-1 ,x t ]+b f ) (1)

[0015] Memory gate (select information to be remembered):

[0016] i t =σ(W i ·[h t-1 ,x t ]+b i ) (2)

[0017]

[0018] Current cell state:

[0019]

[0020] Output gate and current hidden state:

[0021] o t =σ(W o [h t-1 ,x t ]+b o ) (5)

[0022] h t =O t ·tanh(C t ) (6)

[0023] where W and b represent the Bi-LSTM network learning weights and bias, respectively;

[0024] Step 3, feature extraction of the five amino acid feature vectors is performed by a fully connected neural network;

[0025] Step 3.1, according to the five amino acid features: binary (BPF), dipeptide composition (DPC), k-spaced amino acid group pair composition (CKSAAGP), amino acid composition (AAC) and sequence order coupling number (SOCNumber), the primary letter sequence of the peptide is feature encoded, and the feature encoding converts the peptide sequence into a 770-dimensional feature vector;

[0026] where the five feature encodings include: BPF feature encoding, DPC feature encoding, CKSAAGP feature encoding, AAC feature encoding and SOCNumber feature encoding;

[0027] The BPF feature encoding is represented as:

[0028] In the binary, each amino acid letter is represented by a 20-dimensional vector composed of 0 / 1, for example, the first amino acid letter A is represented as f(A) = (1, 0,..., 0), the second amino acid letter C is represented as f(C) = (0, 1,..., 0), and so on. For a peptide sequence P, its binary feature can be represented as:

[0029] B(P) = [f(p1), f(p2),..., f(p n )] (7)

[0030] where P is the peptide sequence, f(p n ) represents a certain amino acid letter;

[0031] The DPC feature encoding is represented as:

[0032] The DPC composition consists of 400 descriptors, which are defined as: the number of dipeptide combinations in a given peptide sequence, represented as:

[0033]

[0034] where N ab is the number of dipeptides represented by amino acid types a and b;

[0035] The CKSAAGP feature encoding is represented as:

[0036] In the k-spaced amino acid pair composition, based on different groups composed of physicochemical properties, the frequency of amino acid pairs separated by any k residues is calculated. Taking k=0 as an example, there will be 25 zero-spaced pairs (g1g1, g1g2, ..., g5g5). The eigenvector is defined as follows:

[0037]

[0038] The value of each descriptor represents the composition of the corresponding residue pair in the peptide sequence. For a peptide sequence of length n, when k = 0, 1, 2, 3..., N all = n-1, n-2, n-3, n-4...

[0039] AAC feature encoding is represented as:

[0040] AAC composition calculations show the frequency of each amino acid type in a peptide sequence. The frequencies of the 20 amino acids (A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W, Y) can be expressed as:

[0041]

[0042] Where N(a) represents the number of times the amino acid appears in the peptide sequence, and N represents the length of the peptide sequence;

[0043] The SOCNumber feature encoding is represented as follows:

[0044]

[0045] Where, d i,i+d This describes the distance between two amino acids at positions i and i+d, where nlag represents the maximum lag value and N is the length of the peptide sequence.

[0046] Step 4: The feature vectors generated in Step 2 and Step 3 are fused using the Concatenate algorithm. The feature vectors are input into a fully connected layer with 512 units and a ReLU activation function, and then passed through a fully connected layer with 1 unit and a Sigmoid activation function to obtain a probability score between 0 and 1. A score greater than 0.5 is considered an anticancer peptide, and a score less than 0.5 is considered a non-anticancer peptide.

[0047] The beneficial effects of this invention are:

[0048] 1. It enables rapid and efficient identification of anticancer peptides; through feature fusion, it improves the accuracy of anticancer peptide identification. Attached Figure Description

[0049] Figure 1A schematic diagram of an anticancer peptide prediction model architecture based on a bidirectional long short-term memory network and feature fusion according to the present application;

[0050] Figure 2 An amino acid alphabet according to the present application;

[0051] Figure 3 An amino acid composition analysis of the data set according to the present application;

[0052] Figure 4 A comparison of the method and a single feature model according to the present application;

[0053] Figure 5 A comparison of different model methods of the ACpred-Fuse and ACpred-FL data sets according to the present application;

[0054] Figure 6 A comparison of different model methods of the ACP240 and ACP740 data sets according to the present application. DETAILED DESCRIPTION

[0055] The present application will be further described below with reference to the accompanying drawings and examples, which are simplified schematic diagrams and only schematically illustrate the basic structure of the present application, and thus only show the components related to the present application.

[0056] As shown in Figure 1 , the anticancer peptide prediction method based on a bidirectional long short-term memory network and feature fusion comprises the following steps:

[0057] Step 1, read four reference peptide sequence data sets, analyze the amino acid composition of the data sets, data set table 1, data set analysis as shown in Figure 3 ;

[0058] Table 1 Four reference peptide sequence data sets

[0059]

[0060]

[0061] The amino acid alphabet encodes the primary letter sequence of the peptide, that is, 20 basic amino acids are assigned numbers 1-20, and 0 is filled for peptide sequences with insufficient length to ensure that all peptide sequences have the same length, and the amino acid alphabet is shown in Figure 2 ;

[0062] Step 2, feature extraction of the data set by a bidirectional long short-term memory network (Bi-LSTM), to generate a Bi-LSTM feature vector;

[0063] Further, step 2 includes:

[0064] Step 2.1, the primary letter sequence of the peptide is digitally coded according to the amino acid alphabet, that is, 20 basic amino acids are assigned numbers 1-20, and the peptide sequence is filled with 0 if the length is not enough, so as to ensure that all the peptide sequences have the same length;

[0065] Step 2.2, the input number is converted into a 64-dimensional vector representation by the embedding layer of Bi-LSTM;

[0066] Step 2.3, the input 64-dimensional vector is extracted by Bi-LSTM, and the specific composition of Bi-LSTM is: the input x t at time t, the cell state C t , the temporary cell state the hidden layer state h t , the forget gate f t , the memory gate i t , the output gate O t ; the bidirectional long short-term memory network layer is composed of forward and backward long short-term memory network layers, and each layer is composed of a memory unit and a 64-dimensional hidden unit;

[0067] Step 3, the five amino acid feature vectors are extracted by the fully connected neural network;

[0068] The amino acid feature coding encodes the primary letter sequence of the peptide according to five amino acid features: binary (BPF), dipeptide composition (DPC), k-interval amino acid group pair composition (CKSAAGP), amino acid composition (AAC) and sequence order coupling number (SOCNumber). The feature coding will convert the peptide sequence into a 770-dimensional feature vector, and the five feature codings are as follows:

[0069] (1) Binary (BPF):

[0070] In the binary, each amino acid letter is represented by a 20-dimensional vector composed of 0 / 1; for example, the first amino acid letter A is represented as f(A) = (1, 0,..., 0), the second amino acid letter C is represented as f(C) = (0, 1,..., 0), and so on; for a peptide sequence P, its binary feature can be represented as:

[0071] B(P) = [f(p1), f(p2),..., f(p n )] (7)

[0072] (2) Dipeptide composition (DPC):

[0073] The dipeptide composition consists of 400 descriptors, which are defined as: the number of dipeptide combinations in a given peptide sequence, which can be represented as:

[0074]

[0075] where N ab is the number of dipeptides represented by amino acid types a and b.

[0076] (3) k-spaced amino acid pair composition (CKSAAGP):

[0077] In k-spaced amino acid pair composition, the frequency of amino acid pairs separated by any k residues is calculated according to different groups composed of physicochemical properties. Taking k = 0 as an example, there will be 25 zero-spaced group pairs (g1g1, g1g2,..., g5g5), and the feature vector will be defined as:

[0078]

[0079] where the value of each descriptor represents the composition of the corresponding residue pair in the peptide sequence. For a peptide sequence of length n, when k = 0, 1, 2, 3,..., N all = n-1, n-2, n-3, n-4,...

[0080] (4) Amino acid composition (AAC):

[0081] Amino acid composition calculates the frequency of each amino acid type in the peptide sequence. The frequency of 20 amino acids (A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W, Y) can be represented as:

[0082]

[0083] where N(a) represents the number of times the amino acid appears in the peptide sequence, and N represents the length of the peptide sequence.

[0084] (5) Sequence order coupling number (SOCNumber):

[0085] The sequence order coupling number can be defined as:

[0086]

[0087] where, d i,i+d describes the distance between two amino acids at positions i and i+d, nlag represents the maximum value of lag, and N is the length of the peptide sequence.

[0088] Step 4, the feature vectors generated in steps 2 and 3 are fused by the Concatenate algorithm, the feature vectors are input into a fully connected layer with 512 units and relu activation function, and through a fully connected layer with 1 unit and Sigmoid activation function, a probability score between 0 and 1 is obtained, greater than 0.5 is considered as an anticancer peptide, and less than 0.5 is considered as a non-anticancer peptide.

[0089] In order to prove the effectiveness of the fusion features, the BLSTM-ACP is compared with the model of single feature, and the results are shown in Table 2. Figure 4

[0090] According to Table 2, the accuracy, the Matthews correlation coefficient, the sensitivity, the specificity and the area under the ROC curve of the BLSTM-ACP of the application are obviously better than those of other methods, and the intuitive comparison is shown in Table 2. Figure 5 6

[0091] Table 2 Comparison of BLSTM-ACP and different anti-cancer peptide prediction methods on four benchmark data sets

[0092]

[0093]

[0094] Note: The maximum value is marked in bold; ACC: accuracy; MCC: Matthews correlation coefficient; SE: sensitivity; SP: specificity; AUC: area under the ROC curve

[0095] In summary, the application can realize high accuracy, high Matthews correlation coefficient, high sensitivity, high specificity and high area under the ROC curve of the anti-cancer peptide prediction.

[0096] Based on the above ideal embodiments according to the application, through the above description, relevant personnel can make various changes and modifications without deviating from the technical idea of the application. The technical scope of the application is not limited to the contents of the specification, and must be determined according to the scope of the claims.​​​

Claims

1. A method for predicting anticancer peptides based on bidirectional long short-term memory networks (Bi-LSTM) and feature fusion, characterized in that, Includes the following steps: Step 1: Read the four benchmark peptide sequence datasets ACPred-Fuse, ACPred-FL, ACP240, and ACP740, and perform amino acid composition analysis on the datasets; Step 2: Extract features from the dataset using Bi-LSTM to generate Bi-LSTM feature vectors; include: Step 2.1: Encode the primary letter sequence of the peptide according to the amino acid alphabet; Step 2.2: Transform the input digit encoding into a 64-dimensional vector using the embedding layer of Bi-LSTM; Step 2.3: Bi-LSTM extracts features from the input 64-dimensional vector. Bi-LSTM includes: Input of time Cell state Temporary cell state Hidden state Forgotten Gate Memory Gate Output gate Bi-LSTM consists of forward and backward long short-term memory network layers, each layer consisting of one memory unit and 64-dimensional hidden units; among which, (1) (2) (3) (4) (5) (6) in, W and b These represent the weights and biases learned by the Bi-LSTM network, respectively. Step 3: Extract features from the five amino acid feature vectors using a fully connected neural network; Step 3 includes: Based on the five amino acid characteristics of BPF, DPC, CKSAAGP, AAC, and SOCNumber, the primary letter sequence of the peptide is feature-encoded, and the feature encoding converts the peptide sequence into a 770-dimensional feature vector. Feature encoding includes: In BPF features, each amino acid letter is represented by a 20-dimensional vector consisting of 0s and 1s; BPF feature encoding is represented as follows: (7) in, P It is a peptide sequence. Indicates amino acids The 20-dimensional 0 / 1 binary representation; DPC feature encoding is represented as: (8) in, It is determined by amino acid type a and b The number of dipeptides represented. The length of the peptide sequence; CKSAAGP feature encoding is represented as: (9) in, - There are 25 pairs with zero gaps. , n The length of the peptide sequence. This represents the number of amino acid intervals. AAC feature encoding is represented as: (10) in, Indicates amino acids The number of times it appears in the peptide sequence, Indicates the length of the peptide sequence; The SOCNumber feature encoding is represented as follows: (11) in, Describes the position of two amino acids and The distance between them This indicates the maximum value of the lag. It is the length of the peptide sequence; Step 4: The feature vectors generated in Step 2 and Step 3 are fused using Concatenate feature fusion, and then input into a fully connected layer with 512 units and a ReLU activation function, and then into a fully connected layer with 1 unit and a Sigmoid activation function to obtain a probability score from 0 to 1. A score greater than 0.5 indicates an anticancer peptide, and a score less than or equal to 0.5 indicates a non-anticancer peptide.

Images