A method and system for identifying anticancer peptides

Abstract

The present invention is applicable to the field of peptide recognition technology and provides an anticancer peptide recognition method and system. The recognition method includes the following steps: inputting a given peptide sequence, performing data enhancement on the peptide sequence to obtain an original sequence; the first channel uses Bi-LSTM to extract features from the original sequence; the second channel converts the original sequence into a chemical formula, and uses SMILES to simplify the chemical formula, and inputs the sequence represented by SMILES into a pre-trained BERT model to obtain deep abstract features; the third channel fuses four features, BPF, DPC, PAAC, and K-mer, to jointly extract features at different levels of the original sequence; splicing the features extracted by the three channels, and classifying the input peptide sequence through a fully connected layer. The ACP-BC model in the present invention has strong robustness and versatility, and has great potential in predicting ACP and non-ACP.

Description

Technical Field

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

[0002] Cancer is a serious disease and one of the leading causes of death in humans. Anticancer peptides (ACPs) have been shown to possess potent anticancer activity while being well tolerated and relatively safe. However, wet-lab experimental methods to identify ACPs are expensive, laborious, and time-consuming.

[0003] Numerous computational techniques in the field of bioinformatics are being used to solve a wide range of problems. Machine learning has shown significant advantages and promise in effectively identifying ACPs. Over the past few years, a range of traditional machine learning methods have been proposed for ACP identification. These methods require manually designed features to construct protein sequences for classification. Consequently, various methods for extracting effective features have been proposed. However, with the rapid development of the big data era and the surge in biological big data in recent years, traditional machine learning algorithms are no longer able to adapt to complex and volatile data. Deep learning methods have been widely used in bioinformatics due to their ability to quickly and efficiently process unstructured data. An increasing number of deep neural network models are being applied to ACP identification. While these studies have achieved acceptable results, there is still room for improvement. For example, these studies only consider information derived from the raw amino acid sequence and fail to consider the spatial structure of amino acids, resulting in increased accuracy. To this end, we propose a method and system for identifying anticancer peptides. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for identifying anticancer peptides, aiming to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for identifying anticancer peptides, comprising the following steps:

[0007] Step S1: input a given peptide sequence, perform data enhancement on the peptide sequence, and obtain the original sequence;

[0008] Step S21: extract features from the original sequence using a bidirectional long short-term memory network (Bi-LSTM);

[0009] Step S22: convert the original sequence into a chemical formula, simplify the chemical formula using Simplified Molecular Linear Input Specification (SMILES), and input the SMILES-represented sequence into a pre-trained Bidirectional Encoder Representation from Transformer (BERT) model to obtain deep abstract features;

[0010] Step S23, fusing four features, namely, binary profile feature (BPF), dipeptide composition (DPC), pseudo amino acid composition (PAAC), and k-mer sparse matrix (K-mer), to jointly extract features at different levels of the original sequence;

[0011] Step S3: Concatenate the features extracted in steps S21, S22, and S23, and classify the input peptide sequence through a fully connected layer.

[0012] Furthermore, the peptide sequence data enhancement method is a combination of any two of replacement, inversion, local random shuffling and combination enhancement.

[0013] Furthermore, the peptide sequence encoding and embedding is represented as:

[0014] P=p1p2p3…pn

[0015] Where p1 is the first residue of peptide P, p2 is the second residue in peptide P, and so on, pn is the nth residue of peptide P, and n is the length of the peptide; each residue pi is represented by an element of the standard amino acid alphabet; the standard amino acid alphabet consists of 20 different amino acids, including A, C, D, E, G, H, I, K, L, M, N, P, Q, R, T, V, W and Y.

[0016] Furthermore, the BPF is expressed as:

[0017] BPF(K)=[f(p1),f(p1),f(p1),…,f(pK)]

[0018] Where f(p1) is the code for the first amino acid letter in the alphabet, f(p2) is the code for the second amino acid letter in the alphabet, and so on, f(pk) is the code for the Kth amino acid letter in the alphabet; where K represents the N-terminal length of peptide P, and the dimension of BPF(k) is 1*20.

[0019] Furthermore, the DPC is expressed as:

[0020]

[0021] Where f(a,b) is the dipeptide represented by type a and b amino acids in the 400 dipeptides, Nab is the number of dipeptides represented by type a and b amino acids, and N is the length of the peptide.

[0022] Furthermore, the PAAC is expressed as:

[0023]

[0024] Where λ is a comprehensive parameter set by the user, with a recommended value of 10; fi is the frequency of each amino acid in the protein; w is a weight factor set by the user, with a default value of 0.05; τ k is the k-th layer correlation factor, reflecting the sequence order correlation between all k-th adjacent residues.

[0025] Furthermore, the K-mer is represented as:

[0026] M=(a ij )7 k ×(L-K+1)

[0027]

[0028] Assuming the length of a peptide sequence is L, there are 7 k There are different possible K-mers, L-K+1 steps in the sequence; converting a peptide sequence into 7 k ×(L-K+1)K-mer sparse matrix M, which regards any K consecutive amino acids as a unit.

[0029] An anticancer peptide recognition system, comprising:

[0030] ACP-BC model, the ACP-BC model includes:

[0031] The data enhancement module is used to perform data enhancement on the input peptide sequence to obtain the original sequence;

[0032] The first channel is used to extract features from the original sequence using Bi-LSTM;

[0033] The second channel is used to convert the original sequence into the form of chemical molecular formula, and use SMILES to simplify the chemical molecular formula. The sequence represented by SMILES is input into the pre-trained BERT model to obtain deep abstract features;

[0034] The third channel is used to fuse the four features of BPF, DPC, PAAC and KMER to jointly extract features at different levels of the original sequence;

[0035] The classification module concatenates the features extracted by the first, second and third channels and classifies the input peptide sequence through a fully connected layer.

[0036] Furthermore, in the first channel, the original sequence is first mapped into a 256-dimensional vector through the Embedding layer integrated into the model, and then input into the Bi-LSTM for feature extraction.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. Data augmentation technology further demonstrates that the present invention is more effective on small datasets; the three-channel end-to-end model (ACP-BC model) that utilizes a combination of various data augmentation techniques has strong robustness and versatility, and is suitable for the identification of unknown anticancer peptide sequences;

[0039] 2. The present invention provides a new method for identifying anticancer peptides, which solves the problems of expensive, laborious and time-consuming wet laboratory experimental methods for identifying ACPs;

[0040] 3. On ACP240, the present invention was compared with multiple advanced methods. Compared with the best ACP-Check method, the ACC was improved by 1.3%, and the MCC and AUC were improved by 0.01 respectively. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Flowchart of the present invention.

[0042] Figure 2 Schematic diagram of the process of anticancer peptide recognition in the present invention.

[0043] Figure 3 This is a preference map of amino acid molecular positions in the anticancer peptides and non-anticancer peptides in the benchmark data set of Example 1 of the present invention.

[0044] Figure 4 This is a comparison chart of fusion features on the benchmark dataset of Example 1 of the present invention.

[0045] Figure 5 2 is a comparison chart of the ROC curves of the ACP740 and ACP240 datasets in Example 1 of the present invention. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0047] The specific implementation of the present invention is described in detail below with reference to specific embodiments.

[0048] like Figure 1 and Figure 2 As shown, a method for identifying anticancer peptides provided by one embodiment of the present invention comprises the following steps:

[0049] Data collection (Data Collection) is performed on the peptide sequence, and data augmentation (DataAugmentation) is performed on the peptide sequence to obtain the original sequence;

[0050] Create an ACP-BC model and perform feature construction. The ACP-BC model includes:

[0051] In the first channel, Bi-LSTM is used to extract features from the original sequence. The original sequence is first mapped into a 256-dimensional vector through the Embedding layer integrated into the model, and then input into the Bi-LSTM for feature extraction.

[0052] In the second channel, the original sequence is converted into a chemical formula and simplified using SMILES (a structured language for representing chemical molecular structures, used to simplify the representation of chemical formulas). The sequence represented by SMILES is input into the pre-trained BERT model for fine-tuning to obtain deep abstract features.

[0053] The third channel integrates some handcrafted features with relatively good results, including binary profile features (BPF), dipeptide composition (DPC), pseudo amino acid composition (PAAC) and k-mer sparse matrix (k-mer) features. These four features are combined to jointly extract features at different levels of amino acid sequences.

[0054] The features extracted by the first, second, and third channels are combined, and the input peptide sequence is classified (Classification) through a fully connected layer to train the model. The last part is to evaluate the trained model.

[0055] As a preferred embodiment of the present invention, the peptide sequence data enhancement method is a combination of any two of replacement, inversion, local random shuffling and combination enhancement.

[0056] As a preferred embodiment of the present invention, the peptide sequence encoding and embedding is expressed as:

[0057] P=p1p2p3…pn

[0058] Where p1 is the first residue of peptide P, p2 is the second residue in peptide P, and so on, pn is the nth residue of peptide P, and n is the length of the peptide; each residue pi is represented by an element of the standard amino acid alphabet; the standard amino acid alphabet consists of 20 different amino acids, including A, C, D, E, G, H, I, K, L, M, N, P, Q, R, T, V, W and Y.

[0059] As a preferred embodiment of the present invention, the BPF is expressed as:

[0060] BPF(K)=[f(p1),f(p1),f(p1),…,f(pK)]

[0061] In the formula, f(p1) is the encoding of the first amino acid letter in the alphabet, f(p2) is the encoding of the second amino acid letter in the alphabet, and so on. f(pk) is the encoding of the k-th amino acid letter in the alphabet. Here, K represents the N-terminal length of the peptide P, and the dimension of BPF(k) is 1*20.

[0062] As a preferred embodiment of the present invention, the DPC is expressed as:

[0063]

[0064] In the formula, f(a,b) is the dipeptide represented by amino acids of type a and b among 400 dipeptides, Nab is the number of dipeptides represented by amino acids of type a and b, and N is the length of the peptide.

[0065] As a preferred embodiment of the present invention, the PAAC is expressed as:

[0066]

[0067] PAAC can be represented by P = [P1, p2,..., p20, p20+1,..., p20+λ] T (λ < L). In the formula, λ is a comprehensive parameter set by the user, and the recommended value is 10; fi is the frequency of each amino acid in the protein; w is a weight factor set by the user, and the default value is 0.05; τ k is the correlation factor of the k-th layer, reflecting the sequence order correlation between all k-th adjacent residues.

[0068] As a preferred embodiment of the present invention, the K-mer is expressed as:

[0069] M = (a ij )7 k × (L - K + 1)

[0070]

[0071] K-mer means scanning a sequence from left to right and taking out K consecutive elements each time through a sliding window. It not only considers the properties of one amino acid and K - 1 consecutive amino acids. Assuming the length of a peptide sequence is L, there are 7 k different possible K-mers, and there are L - K + 1 steps in the sequence; a peptide sequence is transformed into a 7 k × (L - K + 1) K-mer sparse matrix M, and any K consecutive amino acids are regarded as a unit.

[0072] An anti-cancer peptide recognition system provided by an embodiment of the present invention includes:

[0073] ACP-BC model, the ACP-BC model includes:

[0074] The data enhancement module is used to perform data enhancement on the input peptide sequence to obtain the original sequence;

[0075] The first channel is used to extract features from the original sequence using Bi-LSTM;

[0076] The second channel is used to convert the original sequence into the form of chemical molecular formula, and use SMILES to simplify the chemical molecular formula. The sequence represented by SMILES is input into the pre-trained BERT model to obtain deep abstract features;

[0077] The third channel is used to fuse the four features of BPF, DPC, PAAC and KMER to jointly extract features at different levels of the original sequence;

[0078] The classification module concatenates the features extracted by the first, second and third channels and classifies the input peptide sequence through a fully connected layer.

[0079] Example 1: Two benchmark datasets, ACP740 and ACP240, and an independent validation dataset, ACP164 and ACP20, are used. These four datasets are non-redundant and non-overlapping, laying a solid foundation for the research of the present invention.

[0080] To better train the model, the present invention divides the dataset into training and test sets. The benchmark datasets ACP740 and ACP240 are randomly shuffled and split, with 80% used for training and 20% for testing. ACP164 and ACP20 are used as an independent validation set to further verify the model's generalization ability. It is important to note that all existing methods are evaluated on the test set to ensure a fair assessment.

[0081] 1. Amino acid composition and position preference analysis

[0082] Anticancer peptides are small molecule peptides, usually composed of 5-40 amino acids. In order to investigate whether there is a preference for amino acid molecules in anticancer peptides and non-anticancer peptides in terms of position, the present invention created a probability logo map (such as the 15 amino acid residues taken from the N-terminus of the two benchmark data sets ACP740 and ACP240) Figure 3As shown), larger letters indicate that the amino acid is more frequently present at that position. Preliminary observations indicate that the letter F, representing phenylalanine, is more frequently present at the N-terminus of ACPs, while the letters G, A, and D, representing glycine, alanine, and aspartic acid, respectively, also occupy the majority of the N-terminus, but their physical and chemical properties vary greatly. At the second position, the letter L represents leucine, and the letter F is the most common. They have similar properties. At other positions, the letter K represents lysine, and the letter L appears frequently, but also appears in non-ACPs. Although there are differences between ACPs and non-ACPs, there are also large differences between ACPs. In terms of the positional preferences of amino acid residues, ACPs and non-ACPs are not completely separated, and distinguishing ACPs from non-ACPs remains a challenging problem.

[0083] 2. Hyperparameter Selection

[0084] In the experiment, there are many hyperparameters that need to be set manually. The present invention determines the selection of several key hyperparameters through experiments on the independent validation set ACP164. These hyperparameters have a significant impact on the results. For example, the data expansion ratio R (0.5, 1.0, 2.0), the number of neurons C in the LSTM hidden layer (128, 256, 512), and the number of neurons D in the embedding layer (128, 256, 512). R represents the multiple by which the data set is expanded, where R = 1.0 means that the data set is expanded to twice its original size. The number of neurons D in the embedding layer is also a parameter that has a significant impact on the results. It represents the encoding of each amino acid residue in the original peptide chain as a vector of length D.

[0085] The present invention selects the final hyperparameters primarily based on accuracy and the Matthews correlation coefficient. The effects of different hyperparameter combinations revealed that when R is set to 1.0, C is set to 256, and D is set to 512, the corresponding percentage increases in accuracy, Matthews correlation coefficient, sensitivity, and specificity are 1.2%-9.3%, 1.3%-18%, 1.3%-9%, and 1.4%-8%, respectively.

[0086] 3. Ablation Experiment

[0087] In the Chemical BERT of the feature channel, it is very important to choose a suitable BERT pre-training model. We compared two different BERT models for feature extraction of SMILES format data, namely ChemBerta based on RoBerta and BERT-base. Among them, ChemBERTa is pre-trained on a large amount of SMILES format data, while BERT-base is pre-trained based on human natural language text. For ChemBERTa, the present invention also compared two different word segmentation methods, namely SMILES-tokenizer (ST) and BPE tokenizer (BT). On the ACP740 dataset, ChemBerta+ST improved the accuracy, Matthews correlation coefficient, sensitivity and specificity percentages by 1.2%-2.4%, 5.1%-6.3%, 2.1%-2.4% and 2.2%-3.1% respectively compared with ChemBerta+BT and BERT-base. On the ACP240 dataset, ChemBerta+ST achieved improvements of 1.3%-4.2% in accuracy, 2.4%-6.2% in Matthews correlation coefficient, 2.1%-3.3% in sensitivity, and 1.2%-4.2% in specificity compared to ChemBerta+BT and BERT-base, respectively. Experimental results show that while the performance difference with the BERT model is minimal, the ChemBerta model, based on the SMILES tokenizer, achieves superior results.

[0088] 4. Comparison of fusion features

[0089] The third channel of the ACP-BC model is the four manually selected features. This paper selects four features, which respectively contain composition information, local information, long-range relationship information and local information calculated based on the physical and chemical properties of proteins. In order to investigate the role of these features, this paper designed a series of ablation experiments, and conducted experiments on ACP740 and ACP240 respectively to investigate the effect of these fusions. First, a separate experiment was conducted on each feature representation method. The experimental results are shown in the figure below. Figure 4 As shown in Figure 2, PAAC performs better than other features. However, other features come from different perspectives and cannot be simply evaluated. Therefore, we use BERT+LSTM and other features to design experiments. The specific designs are BERT+LSTM+DPC, BERT+LSTM+BPF, BERT+LSTM+DPC+BPF, BERT+

[0090] LSTM+BPF+DPC+KMER, BERT+LSTM+BPF+DPC+KMER+PAAC. from Figure 4It can be seen that these features are relatively representative features, and they extract features at different levels, so better results are achieved when they are combined together.

[0091] 5. Comparative Experiments on Advanced Methods

[0092] In order to better demonstrate the effectiveness of the model, the present invention compared the ACP-BC model with other advanced methods. The present invention evaluated the performance of methods such as ACP-DA, ACP-DL, ACPNet, ACP-CHECK, DEEPACPred, GRCI-Net, and ACP-MHCNN on the same ACP740 and ACP240 datasets. Compared with ACP-ADA, the two feature enhancement methods adopted by the present invention achieved better results. In terms of feature engineering, the features selected by the present invention are more representative and can extract the features of amino acid sequences from multiple angles. In addition, compared with other existing methods, the present invention proposed for the first time the application of SMILES sequences and BERT models to the recognition of ACP. According to the experimental results, the present invention achieved satisfactory results.

[0093] The present invention outperformed ACP-DA, ACP-DL, GRCI-Net, DeepACPred, ACP-MHCNN, and ACP-CHeck on the ACP740 and ACP240 datasets. The present invention achieved accuracies of 0.87 and 0.90, respectively, representing percentage improvements of 1.1%-6.2% and 1.3%-7.3% over the six aforementioned methods. In terms of the MCC metric, the ACP-BC model achieved a 0.75 on the ACP740 dataset, improving the Matthews correlation coefficient of the six aforementioned methods by 1.2%-17%. It also achieved the best result of 0.78 on the ACP240 dataset, improving the Matthews correlation coefficient by 1.1%-12%. These preliminary results provide preliminary evidence for the effectiveness of data augmentation methods. Specifically, on the ACP240 dataset, the present method demonstrated improvements of 1.3% ACC, 0.01 MCC, and 0.01 AUC compared to ACP-Check. This shows that the ACP-BC model of the present invention performs better on small datasets. At the same time, on the ACP740 dataset, the ACP-BC model of the present invention was compared with the above six methods in terms of sensitivity, specificity, and area under the curve, with improvement percentages of 1.1%-7.2%, 1.3%-8.2%, and 1.2%-16%, respectively. On the ACP240 dataset, the ACP-BC model of the present invention was compared with the above six methods in terms of sensitivity, specificity, and area under the curve, with improvement percentages of 1.3%-2.4%, 1.2%-13%, and 1.4%-13%, respectively. Overall, the ACP-BC model successfully extracted features at different levels and showed stronger compatibility with small datasets. Therefore, the ACP-BC model has great potential in predicting ACP and non-ACP.

[0094] In addition, the present invention draws a ROC curve diagram to further prove the performance of ACP-BC, such as Figure 5 As shown in Figure 2, on the ACP740 and ACP240 datasets, the AUCs are 0.90 and 0.93, respectively, showing better performance.

[0095] 6. Independent Verification

[0096] In order to further verify the effectiveness and robustness of ACP-BC, the present invention was independently validated using the ACP20 dataset, which consists of 10 ACP and 10 non-ACP samples. The model of the present invention was not trained on this dataset, which enables us to better evaluate the robustness of the model. Through experiments, when the model trained on the ACP740 dataset was applied to the ACP20 dataset, if the prediction score was greater than 0.5, the corresponding peptide was considered to be an anticancer peptide (ACP), otherwise it was considered to be a non-anticancer peptide (Non-ACP). The prediction scores of the 10 ACPs were 0.85, 0.88, 0.82, 0.68, 0.93, 0.73, 0.78, 0.78, 0.92, and 0.67, respectively. The prediction scores of the 10 non-ACP sequences were 0.16, 0.18, 0.10, 0.03, 0.05, 0.36, 0.25, 0.40, 0.12, and 0.29, respectively. Experimental results show that all ACP and non-ACP samples were accurately identified. The ACP-BC model is robust and versatile and is suitable for the identification of unknown anticancer peptide sequences.

[0097] The above are only preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, several variations and improvements can be made without departing from the concept of the present invention. These should also be regarded as the scope of protection of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A method for identifying anticancer peptides, characterized in that: The following steps are involved: Step S1, performing data enhancement on the input peptide sequence to obtain the original sequence; the data enhancement method of the peptide sequence is a combination of any two methods of replacement, inversion, local random shuffling and combination enhancement; Step S21: Use Bi-LSTM to extract features from the original sequence; The original sequence is first mapped into a 256-dimensional vector through the Embedding layer integrated into the model, and then input into the Bi-LSTM for feature extraction; Step S22: convert the original sequence into a chemical formula, simplify the chemical formula using SMILES, and input the sequence represented by SMILES into the pre-trained BERT model to obtain deep abstract features; Step S23: Fusing BPF, DPC, PAAC, and K-mer features to jointly extract features at different levels of the original sequence; Step S3: splicing the features extracted in steps S21, S22, and S23, and classifying the input peptide sequence through a fully connected layer; The BPF is expressed as: ； Where f(p1) is the code of the first amino acid letter in the alphabet, f(p2) is the code of the second amino acid letter in the alphabet, and so on, f(pk) is the code of the Kth amino acid letter in the alphabet; where K represents the N-terminal length of peptide P, and the dimension of BPF(k) is 1*20; The DPC is expressed as: ； Where, It is a dipeptide represented by a and b type amino acids among the 400 dipeptides. is the number of dipeptides represented by a- and b-type amino acids, and N is the length of the peptide; The PAAC is represented by: ； Where λ is a comprehensive parameter set by the user, with a recommended value of 10; fi is the frequency of each amino acid in the protein; w is a weight factor set by the user, with a default value of 0.05; τ k is the k-th layer correlation factor, reflecting the sequence order correlation between all k-th adjacent residues; The K-mer is represented as: ； ； Assuming the length of a peptide sequence is L, then different possible K-mers, appearing in the sequence Step 1: Convert a peptide sequence into The K-mer sparse matrix M treats any K consecutive amino acids as a unit.

2. The anticancer peptide identification method according to claim 1, characterized in that The peptide sequence is encoded and embedded as: ； Where p1 is the first residue of peptide P, p2 is the second residue in peptide P, and so on, pn is the nth residue of peptide P, and n is the length of the peptide; each residue pi is represented by an element of the standard amino acid alphabet; the standard amino acid alphabet consists of 20 different amino acids, including A, C, D, E, G, H, I, K, L, M, N, P, Q, R, T, V, W and Y.

3. An anticancer peptide recognition system, characterized in that: include: ACP-BC model, the ACP-BC model includes: The data enhancement module is used to perform data enhancement on the input peptide sequence to obtain the original sequence; The first channel is used to extract features from the original sequence using Bi-LSTM; The second channel is used to convert the original sequence into the form of chemical molecular formula, and use SMILES to simplify the chemical molecular formula. The sequence represented by SMILES is input into the pre-trained BERT model to obtain deep abstract features; The third channel is used to fuse the four features of BPF, DPC, PAAC and K-mer to jointly extract features at different levels of the original sequence; The classification module combines the features extracted by the first, second, and third channels and classifies the input peptide sequence through a fully connected layer; The BPF is expressed as: ； Where f(p1) is the code of the first amino acid letter in the alphabet, f(p2) is the code of the second amino acid letter in the alphabet, and so on, f(pk) is the code of the Kth amino acid letter in the alphabet; where K represents the N-terminal length of peptide P, and the dimension of BPF(k) is 1*20; The DPC is expressed as: ； Where, It is a dipeptide represented by a and b type amino acids among the 400 dipeptides. is the number of dipeptides represented by a- and b-type amino acids, and N is the length of the peptide; The PAAC is represented by: ； Where λ is a comprehensive parameter set by the user, with a recommended value of 10; fi is the frequency of each amino acid in the protein; w is a weight factor set by the user, with a default value of 0.05; τ k is the k-th layer correlation factor, reflecting the sequence order correlation between all k-th adjacent residues; The K-mer is represented as: ； ； Assuming the length of a peptide sequence is L, then different possible K-mers, appearing in the sequence Step 1: Convert a peptide sequence into The K-mer sparse matrix M treats any K consecutive amino acids as a unit.

4. The anticancer peptide recognition system according to claim 3, characterized in that In the first channel, the original sequence is first mapped into a 256-dimensional vector through the Embedding layer integrated into the model, and then input into the Bi-LSTM for feature extraction.

Images