Anti-cancer peptide prediction method and system based on feature fusion and cross attention mechanism

Abstract

The application relates to an anticancer peptide prediction method and system based on feature fusion and cross attention mechanism. The method comprises the following steps: constructing a data set containing various protein sequences; inputting the protein sequence into a protein language model ESM-2 to extract peptide structure features; inputting the protein sequence into a feature extraction model to extract peptide physical and chemical features; performing dimension transformation processing on the peptide structure features, using BiLSTM to continuously process discrete peptide physical and chemical features; using a cross attention mechanism for feature fusion to obtain target features input into a multilayer perception machine (MLP) to obtain an anticancer peptide prediction result. By using a protein language model to extract peptide structure features, using a traditional feature extraction model to extract peptide physical and chemical features, and using a cross attention mechanism for feature fusion, time-consuming and high cost can be avoided, and the extracted features are related to each other, so that anticancer peptide prediction can be quickly, efficiently and accurately performed.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of bioinformatics, and in particular to an anti-cancer peptide prediction method and system based on feature fusion and cross attention mechanism. BACKGROUND

[0002] Cancer remains a great challenge for modern medicine, with a wide variety of types and strong metastatic properties, making it a malignant disease that has not yet been completely conquered by humans. Despite significant progress in the field of cancer treatment and numerous attempts at treatment, no universally effective and patient-friendly treatment has been found. Current cancer treatment methods, including chemotherapy, radiotherapy, surgery, and targeted therapy, each have limitations. Chemotherapy and radiotherapy place a heavy burden on the body and are often accompanied by serious side effects such as hair loss and vomiting. Surgical intervention often fails to remove cancer cells that have metastasized to other parts of the body; although the primary tumor can be removed, isolated tumor cells remaining in the surrounding tissue may not be detected. In addition, targeted therapy is expensive and only effective for specific types of cancer.

[0003] In recent years, anti-cancer peptides (ACPs) have attracted widespread attention from researchers. Anti-cancer peptides are widely present in various organisms, including mammals, amphibians, insects, plants, and microorganisms, and can also be obtained synthetically. They interact with the phospholipid bilayer of cancer cell membranes, alter the permeability of the cell membrane, cause the leakage of cell contents, and ultimately lead to cell death. Anti-cancer peptides have many advantages in tumor treatment: they have low molecular weight, simple structure, strong anti-cancer activity, and high selectivity; they have few side effects, can be administered through multiple routes, and are not prone to induce multi-drug resistance.

[0004] However, the identification and discovery of anti-cancer peptides in a large number of extracted peptides often relies on traditional methods such as in vitro cell experiments or animal experiments. These methods are not only time-consuming, but also require careful experimental design, selection of appropriate control groups, and a large amount of financial support. In addition, animal experiments are increasingly questioned due to ethical and moral issues involved. With the rapid development of artificial intelligence, a large number of methods based on machine learning or deep learning have been proposed, but these methods rely heavily on traditional feature encoding techniques, require complex feature engineering steps, and the extracted features have little correlation and are one-sided in focus.

[0005] Therefore, the traditional way of discovering anti-cancer peptides has the problems of low efficiency, low accuracy, and high cost. SUMMARY

[0006] Based on this, in order to solve the above technical problems, a cancer peptide prediction method and system based on feature fusion and cross attention mechanism are provided, which can quickly, efficiently and accurately predict cancer peptides.

[0007] A cancer peptide prediction method based on feature fusion and cross attention mechanism, the method comprises:

[0008] Collecting cancer peptide sequences and non-cancer peptide sequences from a database to construct a dataset containing each protein sequence;

[0009] Inputting the protein sequence into a protein language model ESM-2, and extracting peptide structural features in the protein sequence through a Transformer encoder;

[0010] Inputting the protein sequence into a feature extraction model, and extracting peptide physical and chemical features in the protein sequence through the feature extraction model;

[0011] Performing dimension transformation processing on the peptide structural features to obtain processed peptide structural features; and using BiLSTM to continuous the discrete peptide physical and chemical features to obtain continuous peptide physical and chemical features;

[0012] Using a cross attention mechanism to fuse the processed peptide structural features and the continuous peptide physical and chemical features to obtain target features;

[0013] Inputting the target features into a multi-layer perception MLP to obtain cancer peptide prediction results.

[0014] In one embodiment, collecting cancer peptide sequences and non-cancer peptide sequences from a database to construct a dataset containing each protein sequence, comprising:

[0015] Collecting cancer peptide sequences verified by experiments from a database, using a CD-HIT tool to remove redundancy of the cancer peptide sequences, and filtering to obtain final cancer peptide sequences;

[0016] Randomly collecting non-cancer peptide sequences, using homology bias removal and PSSM extraction standards to extract final non-cancer peptide sequences from the non-cancer peptide sequences;

[0017] Randomly selecting cancer peptide sequences and non-cancer peptide sequences from the final cancer peptide sequences and the final non-cancer peptide sequences respectively to construct a dataset containing each protein sequence.

[0018] In one embodiment, inputting the protein sequence into a protein language model ESM-2, and extracting peptide structural features in the protein sequence through a Transformer encoder, comprising:

[0019] inputting the protein sequence into a protein language model ESM-2, and converting the protein sequence into a numerical vector representation through the protein language model;

[0020] inputting the numerical vector representation into a Transformer encoder in the protein language model ESM-2;

[0021] calculating through a dot-product attention mechanism and a linear layer in the Transformer encoder to obtain a peptide structure feature.

[0022] In one embodiment, the protein sequence is input into a feature extraction model, and a peptide physical and chemical feature in the protein sequence is extracted through the feature extraction model, including:

[0023] inputting the protein sequence into a feature extraction model, and using one-hot encoding in the feature extraction model to represent each amino acid in the protein sequence through a hot encoding to obtain a binary vector corresponding to the protein sequence;

[0024] calculating the sum of the element mass of each amino acid in the protein sequence through the feature extraction model, and taking the sum of the element mass as the molecular weight;

[0025] obtaining the acid-base value of the amino acid in the protein sequence, and calculating the isoelectric point according to the acid-base value;

[0026] calculating the number and properties of the hydrophobic group of the amino acid in the protein sequence, and determining the hydrophobicity of the amino acid based on the number and properties of the hydrophobic group;

[0027] taking the binary vector, the molecular weight, the isoelectric point, and the hydrophobicity as the peptide physical and chemical feature in the protein sequence.

[0028] In one embodiment, the peptide structure feature is subjected to dimension transformation processing to obtain a processed peptide structure feature; and a discrete peptide physical and chemical feature is continuous through BiLSTM to obtain a continuous peptide physical and chemical feature, including:

[0029] inputting the peptide structure feature into a linear layer for dimension transformation processing to obtain a processed peptide structure feature;

[0030] inputting the peptide physical and chemical feature into a bidirectional long short-term memory network BiLSTM, capturing the long-distance dependency relationship of the long sequence in the peptide physical and chemical feature through the BiLSTM, completing the continuous of the peptide physical and chemical feature, and obtaining a continuous peptide physical and chemical feature.

[0031] In one embodiment, the processed peptide structure features and the continuous peptide physicochemical features are fused by using a cross-attention mechanism to obtain target features, including:

[0032] The processed peptide structure features and the continuous peptide physicochemical features are respectively used by the cross-attention mechanism to generate a query matrix, a key matrix, and a value matrix;

[0033] The key vector dimension corresponding to the processed peptide structure features and the continuous peptide physicochemical features is determined;

[0034] The cross-attention is calculated according to the query matrix, the key matrix, the value matrix, and the key vector dimension;

[0035] The feature fusion of the processed peptide structure features and the continuous peptide physicochemical features is completed based on the cross-attention to obtain target features.

[0036] In one embodiment, the method further includes:

[0037] The target features are input into a Transformer architecture, and a multi-head self-attention mechanism in the Transformer architecture is used to calculate a query vector, a key vector, and a value vector for each position;

[0038] The query vector, the key vector, and the value vector are weighted and averaged to obtain a weighted feature;

[0039] The weighted feature is nonlinearly converted position by position using a feedforward neural network.

[0040] In one embodiment, the target features are input into a multi-layer perceptron (MLP) to obtain an anticancer peptide prediction result, including:

[0041] The target features are input into an input layer of the multi-layer perceptron (MLP), and a hidden layer and a linear layer in the multi-layer perceptron (MLP) are used to predict the result;

[0042] The prediction result is output from an output layer of the multi-layer perceptron (MLP) to obtain an anticancer peptide prediction result.

[0043] An anticancer peptide prediction system based on feature fusion and a cross-attention mechanism, including:

[0044] A data collection module is configured to collect anticancer peptide sequences and non-anticancer peptide sequences from a database to construct a dataset containing various protein sequences;

[0045] The structural feature extraction module is configured to input the protein sequence into a protein language model ESM-2, and extract a peptide structural feature in the protein sequence through a Transformer encoder.

[0046] The other feature extraction module is configured to input the protein sequence into a feature extraction model, and extract a peptide physicochemical feature in the protein sequence through the feature extraction model.

[0047] The feature processing module is configured to perform dimension transformation processing on the peptide structural feature to obtain a processed peptide structural feature, and use BiLSTM to continuous the discrete peptide physicochemical feature to obtain a continuous peptide physicochemical feature.

[0048] The feature fusion module is configured to use a cross-attention mechanism to fuse the processed peptide structural feature and the continuous peptide physicochemical feature to obtain a target feature.

[0049] The prediction module is configured to input the target feature into a multi-layer perception MLP to obtain an anticancer peptide prediction result.

[0050] In one embodiment, the data collection module is further configured to collect anticancer peptide sequences verified by experiments from a database, perform de-redundancy processing on the anticancer peptide sequences using a CD-HIT tool, and filter to obtain final anticancer peptide sequences; randomly collect non-anticancer peptide sequences, and extract final non-anticancer peptide sequences from the non-anticancer peptide sequences using homology bias removal and PSSM extraction standards; and randomly select anticancer peptide sequences and non-anticancer peptide sequences from the final anticancer peptide sequences and the final non-anticancer peptide sequences, respectively, to construct a data set containing each protein sequence.

[0051] The above anticancer peptide prediction method and system based on feature fusion and cross-attention mechanism use a protein language model to extract a peptide structural feature, use a traditional feature extraction model to extract a peptide physicochemical feature, use a cross-attention mechanism to fuse features, and finally obtain an anticancer peptide prediction result based on a multi-layer perception MLP. The method and system do not need to provide high cost and do not involve ethical and moral issues, and the extracted features are related to each other, so that the anticancer peptide prediction can be quickly, efficiently and accurately performed. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 An application environment diagram of the anticancer peptide prediction method based on feature fusion and cross-attention mechanism in one embodiment;

[0053] Figure 2 A flowchart of the anticancer peptide prediction method based on feature fusion and cross-attention mechanism in one embodiment;

[0054] Figure 3 Fig. 1 is a schematic diagram of a model framework of ACP-ETBLCA in an embodiment;

[0055] Figure 4 Fig. 2 is a schematic diagram of KDE plot, AUC curve, pre-Recall curve of the model in an embodiment to predict probability distribution on ACP135 and ACP99 dataset;

[0056] Figure 5 Fig. 3 is a schematic diagram of measurement of six indexes of pre, F1, recall, Sp, MCC, ACC respectively for different ESM-2 parameters in an embodiment;

[0057] Figure 6 Fig. 4 is a schematic diagram of exploring whether BiLSTM is needed in an embodiment;

[0058] Figure 7 Fig. 5 is a schematic diagram of exploring the source of Q key in cross-attention in an embodiment;

[0059] Figure 8 Fig. 6 is a schematic diagram of exploring the number of transformer encoder layer in an embodiment;

[0060] Figure 9 Fig. 7 is a schematic diagram of sample classification after data in ACP135 is extracted by traditional feature extraction model, BiLSTM, ESM-2, linear layer, cross-attention in an embodiment;

[0061] Figure 10 Fig. 8 is a schematic diagram of sample classification after data in ACP135 is extracted by 6-layer transformer encoder layer and average pooling layer respectively in an embodiment;

[0062] Figure 11 Fig. 9 is a structural block diagram of an anti-cancer peptide prediction system based on feature fusion and cross-attention mechanism in an embodiment;

[0063] Figure 12 Fig. 10 is an internal structure diagram of a computer device in an embodiment. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0065] The application embodiment provides an anti-cancer peptide prediction method based on feature fusion and cross attention mechanism, which can be applied to an application environment as shown in Figure 1 As shown in Figure 1 The application environment includes a computer device 110. The computer device 110 can collect anti-cancer peptide sequences and non-anti-cancer peptide sequences from a database to construct a data set containing various protein sequences. The computer device 110 can input the protein sequences into a protein language model ESM-2, extract peptide structural features in the protein sequences through a Transformer encoder. The computer device 110 can input the protein sequences into a feature extraction model, and extract peptide physical and chemical features in the protein sequences through the feature extraction model. The computer device 110 can perform dimension transformation processing on the peptide structural features to obtain processed peptide structural features. The discrete peptide physical and chemical features are continuous using BiLSTM to obtain continuous peptide physical and chemical features. The computer device 110 can use a cross attention mechanism to fuse the processed peptide structural features and the continuous peptide physical and chemical features to obtain target features. The computer device 110 can input the target features into a multi-layer perception MLP to obtain an anti-cancer peptide prediction result. The computer device 110 can be, but is not limited to, various personal computers, notebook computers, smartphones, robots and other devices.

[0066] In one embodiment, as shown in Figure 2 An anti-cancer peptide prediction method based on feature fusion and cross attention mechanism is provided, including the following steps:

[0067] Step 202, collecting anti-cancer peptide sequences and non-anti-cancer peptide sequences from a database to construct a data set containing various protein sequences.

[0068] The database can be a public database such as CancerPPD, APD3, SATPd, etc. The computer device can obtain 1350 experimentally verified anti-cancer peptide sequences from the public database, and find 1839 non-anti-cancer peptide data in the literature to construct a data set containing various protein sequences.

[0069] In one embodiment, the provided anti-cancer peptide prediction method based on feature fusion and cross-attention mechanism can further include a process of constructing a data set, and the specific process includes: collecting experimentally verified anti-cancer peptide sequences from a database, performing de-redundancy processing on the anti-cancer peptide sequences using a CD-HIT tool, and filtering to obtain final anti-cancer peptide sequences; randomly collecting non-anti-cancer peptide sequences, and extracting final non-anti-cancer peptide sequences from the non-anti-cancer peptide sequences using homology bias removal and PSSM extraction standards; and randomly selecting anti-cancer peptide sequences and non-anti-cancer peptide sequences from the final anti-cancer peptide sequences and the final non-anti-cancer peptide sequences, respectively, to construct a data set containing each protein sequence.

[0070] The computer device can fuse the anti-cancer peptides and the non-anti-cancer peptides using CD-HIT to remove anti-cancer peptide samples with more than 90% sequence similarity. Subsequently, sequences with a length of 5 to 50 amino acids are extracted using Seqkit, and only those sequences that can generate a position-specific scoring matrix (PSSM) profile by the PSI-BLAST tool are retained to finally obtain 1839 non-anti-cancer peptide samples. From these samples, 487 positive sequences and 1479 negative sequences are randomly selected to construct a training data set, and the remaining sequences constitute an independent test set ACP135. In this embodiment, the widely used ACP99 is also used.

[0071] Specifically, the training set and ACP135 were initially developed by Jilong et al. They collected 1350 experimentally verified anti-cancer peptide sequences from the CancerPPD, APD3, and SATPdb databases. In order to alleviate the homology bias and prevent artificially exaggerating the recognition accuracy, the filtering process in this embodiment finally obtained 622 positive anti-cancer peptide samples. For negative samples, Jilong et al. randomly selected non-anti-cancer peptide sequences and applied the same homology bias removal and PSSM extraction standards to finally obtain 1839 non-anti-cancer peptide samples. From these samples, 487 positive sequences and 1479 negative sequences were randomly selected to construct a training data set.

[0072] ACP135 contains 135 anti-cancer peptide sequences and 360 non-anti-cancer peptide sequences, while the training set includes 487 anti-cancer peptide sequences and 1479 non-anti-cancer peptide sequences. ACP99 was developed by Agrawal et al. and contains 256 sequences, of which 99 are positive samples and 157 are negative samples. The data set has also undergone homology bias removal and PSSM matrix extraction. Training and testing using these strictly screened databases can ensure that the classification performance of the model is accurately evaluated and avoids interference from homology-related artifacts.

[0073] Step 204, input the protein sequence into the protein language model ESM-2, and extract the peptide structure features in the protein sequence through the Transformer encoder.

[0074] The computer device can input the obtained protein sequence into the ESM-2 model for feature extraction. The ESM-2 model is composed of 30 encoding layers, each of which includes a multi-head self-attention layer, a feedforward network layer, a residual connection layer, etc., to fully exploit the peptide structure information.

[0075] ESM-2 is a deep learning pre-training model for protein sequence feature extraction. Similar to BERT, ESM-2 uses self-attention mechanisms to learn representations of amino acid sequences to capture complex features and long-term dependencies in the sequence.

[0076] In one embodiment, the anti-cancer peptide prediction method based on feature fusion and cross-attention mechanism can further include a process of extracting peptide structure features, and the specific process includes: inputting the protein sequence into the protein language model ESM-2, and converting the protein sequence into a numerical vector representation through the protein language model; inputting the numerical vector representation into the Transformer encoder in the protein language model ESM-2; and calculating through the dot-product attention mechanism and linear layer in the Transformer encoder to obtain the peptide structure features.

[0077] In this embodiment, in order to adapt to the background of the protein sequence, ESM-2 converts the input protein sequence into a numerical vector representation, and then extracts sequence features through a Transformer encoder with 6 layers, 12 layers or more layers. Specifically, each layer of the Transformer encoder includes multiple self-attention heads, which are calculated through the dot-product attention mechanism and the linear layer to generate high-quality sequence features, which are then used for downstream tasks such as protein function prediction. The encoding process of ESM-2 can be represented as: Where Q, K, and V represent query vectors, key vectors, and value vectors, respectively, d k is the dimension of the key vector.

[0078] Step 206, input the protein sequence into the feature extraction model, and extract the peptide physical and chemical features in the protein sequence through the feature extraction model.

[0079] The feature extraction model can use traditional feature extraction methods for feature extraction. Specifically, the obtained protein sequence can be based on hydrophobicity, isoelectricity, one-hot encoding, molecular weight, etc. Traditional feature extraction methods based on peptide biochemical information features are obtained.

[0080] In one embodiment, the provided anti-cancer peptide prediction method based on feature fusion and cross-attention mechanism can further include a process of extracting physical and chemical characteristics of peptides, and the specific process includes: inputting a protein sequence into a feature extraction model, using one-hot encoding in the feature extraction model to represent each amino acid in the protein sequence through a hot encoding, to obtain a binary vector corresponding to the protein sequence; calculating the sum of the element mass of each amino acid in the protein sequence through the feature extraction model, and taking the sum of the element mass as the molecular weight; obtaining the acid-base value of the amino acid in the protein sequence, and calculating the isoelectric point according to the acid-base value; calculating the number and nature of the hydrophobic group of the amino acid in the protein sequence, and determining the hydrophobicity of the amino acid based on the number and nature of the hydrophobic group; and taking the binary vector, the molecular weight, the isoelectric point and the hydrophobicity as the physical and chemical characteristics of the peptides in the protein sequence.

[0081] In the embodiment, each amino acid is represented by a hot encoding to generate a binary vector, in which a single position is set to 1 to represent the existence of a specific category, and all other positions are 0. Given C categories, one-hot encoding converts each category C i into a vector v i of length C. The definition formula can be represented as:

[0082] The molecular weight (MW) refers to the total mass of all atoms in the molecule, usually in units of daltons (Da). For an amino acid AAi, its molecular weight MWi is the sum of the mass of all constituent elements.

[0083] The isoelectric point (Isoelectric Point, pI) refers to the pH value at which the amino acid has no net charge in the solution, and is calculated according to the acidity and alkalinity of the amino acid side chain, involving the dissociation constants of carboxyl and amino groups. The calculation formula can be represented as: where pK a1 and pK a2 represent the dissociation constants of carboxyl and amino groups, respectively.

[0084] Hydrophobicity is a measure of the number and nature of the hydrophobic group in the amino acid, which is usually evaluated using the Kyte-Doolittle scale. The hydrophobicity score H_iH_iH_i is defined as the value derived from the Kyte-Doolittle scale. The Kyte-Doolittle scale is a method of measuring the hydrophobicity of amino acids to water, which divides the hydrophobicity of amino acids into positive values (indicating strong hydrophobicity) and negative values (indicating strong hydrophilicity). Each amino acid is assigned a score based on its hydrophobicity, and the higher the score, the more hydrophobic the amino acid.

[0085] Step 208 : Perform dimensionality transformation on the peptide structural features to obtain processed peptide structural features; and use BiLSTM to continuousize the discrete peptide physicochemical features to obtain continuous peptide physicochemical features.

[0086] The computer device can use the linear layer to transform the dimension of the features extracted by ESM-2, and use BiLSTM to make the discrete data continuous.

[0087] Specifically, in one embodiment, a method for predicting anticancer peptides based on feature fusion and cross-attention mechanism is provided, which may also include a feature processing process, the specific process including: inputting the peptide structural features into the linear layer for dimensionality transformation processing to obtain the processed peptide structural features; inputting the peptide physicochemical features into the bidirectional long short-term memory network BiLSTM, capturing the long-distance dependencies of long sequences in the peptide physicochemical features through BiLSTM, completing the continuity of the peptide physicochemical features, and obtaining continuous peptide physicochemical features.

[0088] Among them, the bidirectional long short-term memory (BiLSTM) network is an enhanced version of the traditional long short-term memory (LSTM) network. Its main advantage is its ability to simultaneously capture the global contextual information of the sequence, significantly improving the modeling of long-distance dependencies. Unlike traditional LSTM models, which generally process information in a unidirectional manner, BiLSTM overcomes this limitation by enabling information to flow in both directions. Protein sequences can be viewed as a biological language, where peptide segments correspond to sentences and individual amino acid residues correspond to words. The contextual relationships between these residues are crucial for accurate prediction.

[0089] In this embodiment, although LSTM effectively alleviates the problems of gradient vanishing and gradient exploding, its inherent defect is that it cannot utilize future context information. Therefore, BiLSTM combines two independent LSTM layers: one for forward processing and the other for backward processing. Mathematically, the forward and backward transfer can be expressed as: h t → =LSTM forward (x t ,h t-1 → );h t ← =LSTM forward (x t ,h t+1 ← ); where h t → 、h t ←respectively, denote the hidden states produced by the forward and backward LSTM layers at time step t. The final output of a BiLSTM is obtained by concatenating or combining the hidden states from the forward and backward directions:

[0090] This concatenated hidden state h t provides a comprehensive context representation that encompasses both past and future state information. The bidirectional mechanism enables BiLSTMs to capture long-range dependencies more accurately in long sequences, making them particularly suitable for tasks that require extensive context understanding.

[0091] In this embodiment, the features extracted by the traditional feature extraction method contain physical and chemical properties. However, due to the relatively weak correlation between features from different methods, it may affect the quality of the synthesized features after fusion with ESM-2 features. To preserve the physical and chemical information extracted by the traditional feature extraction model and enhance the relationship between different features, BiLSTM is introduced to ensure that the physical and chemical properties are effectively preserved and the relationship between different feature dimensions is enhanced, thereby improving the overall feature representation and the prediction performance of the model.

[0092] In step 210, the processed peptide structure features and the continuous peptide physical and chemical features are fused using a cross-attention mechanism to obtain target features.

[0093] The computer device can use the cross-attention mechanism to fuse the processed features to obtain a structure information feature containing physical and chemical information. The cross-attention mechanism is a variant of the attention mechanism, which is particularly suitable for multi-modal tasks, multi-sequence tasks, and scenarios involving upstream and downstream feature interaction. The core idea is to extract context information from one input sequence and enhance or guide the representation of another input sequence through the attention mechanism.

[0094] In one embodiment, the anti-cancer peptide prediction method based on feature fusion and cross-attention mechanism provided can further include a feature fusion process, and the specific process includes: generating a query matrix, a key matrix, and a value matrix using the processed peptide structure features and the continuous peptide physical and chemical features through the cross-attention mechanism; determining the key vector dimension corresponding to the processed peptide structure features and the continuous peptide physical and chemical features; calculating the cross-attention according to the query matrix, the key matrix, the value matrix, and the key vector dimension; and completing the feature fusion of the processed peptide structure features and the continuous peptide physical and chemical features based on the cross-attention to obtain target features.

[0095] The cross attention mechanism generates query, key, and value vectors from the input sequence itself. Cross attention uses two different sequences to generate query and key-value vectors respectively. The calculation formula of cross attention is: Where Q represents the query matrix obtained from one input sequence, K and V represent the key and value matrices obtained from another input sequence, respectively. k is the dimension of the key vector. Combining features extracted from ESM-2 and fusing them through the cross-attention mechanism facilitates the effective integration of structural features obtained through ESM-2 with traditional features (physicochemical features processed by BiLSTM). This configuration enables the cross-attention mechanism to effectively combine physicochemical information with structural features, generating a feature set that captures the comprehensive properties of anticancer peptides, thereby improving the overall performance of the model.

[0096] In one embodiment, a method for predicting anticancer peptides based on feature fusion and cross-attention mechanism is provided, which may also include a feature processing process, specifically including: inputting the target feature into the Transformer architecture, calculating the query vector, key vector, and value vector for each position through the multi-head self-attention mechanism in the Transformer architecture; performing weighted average calculation on the query vector, key vector, and value vector to obtain weighted features; and using a feedforward neural network to perform nonlinear transformation on the weighted features position by position.

[0097] The computer device can use the 6-layer transformer encoder layer to further process the fused features to obtain processed features and pass them through the pooling layer.

[0098] Specifically, the Transformer encoder layer is a basic component in the Transformer architecture, which is used to encode the input sequence into a high-dimensional feature representation that can effectively capture local and global dependencies in the data. Each encoder layer contains two main submodules: a multi-head self-attention mechanism and a feed-forward neural network (FFN). Among them, the multi-head self-attention mechanism enables each position in the sequence to pay attention to all other positions, thereby capturing long-range dependencies and contextual information. The multi-head self-attention mechanism involves calculating the query vector (Q), key vector (K), and value vector (V) for each position, and then applying weighted averaging to calculate the weighted representation, where multiple attention heads work in parallel to capture multiple features of the input data.

[0099] By using multiple attention heads, different aspects of input features can be focused on, enhancing the representation ability. After the self-attention mechanism, a feed-forward neural network (FFN) performs a position-wise nonlinear transformation on the features, which usually consists of two linear transformations and an activation function such as ReLU or GELU. This structure enables the model to capture more complex feature interactions.

[0100] Step 212, input the target feature into the multi-layer perception MLP to obtain the anticancer peptide prediction result.

[0101] The computer device can input the processed target feature into the MLP to obtain the final prediction result, and divide whether it is an anticancer peptide according to whether the prediction probability is greater than 0.5.

[0102] Multi-layer perception (MLP, Multi-Layer Perceptron) is a kind of feedforward artificial neural network, which is composed of multiple layers, each layer contains several neurons, and can perform complex nonlinear mapping. MLP is usually used to process supervised learning problems, especially classification and regression tasks. In MLP, data is transmitted through the input layer, passes through several hidden layers (each layer is composed of multiple neurons), and finally outputs the result. The hidden layer and the output layer introduce nonlinear characteristics through the activation function, which improves the expression ability of the model.

[0103] Specifically, in one embodiment, the provided anticancer peptide prediction method based on feature fusion and cross attention mechanism can further include the process of prediction, and the specific process includes: inputting the target feature into the input layer of the multi-layer perception MLP, and predicting the result through the hidden layer and the linear layer in the multi-layer perception MLP; outputting the prediction result from the output layer of the multi-layer perception MLP to obtain the anticancer peptide prediction result.

[0104] The application model ACP-ETBLCA framework provided in the present application is as follows: Figure 3As shown, including A, database collection data; B, extract features cross attention mechanism for fusion and at most layer perception machine MLP; C, practical application. Specifically, the ACP-ETBLCA prediction model contains two stages, the first stage is used to extract features from the input protein sequence, and the second stage uses the constructed model for prediction. In the first stage, the input protein sequence is respectively put into ESM-2 and the traditional feature extraction method, i.e., the feature extraction model, which focuses on extracting the physical and chemical properties of peptides, to extract features; the features extracted by ESM-2 are dimensionally transformed using a linear layer; the features extracted by the traditional feature extraction method are further processed using BiLSTM, and the two processed features are fused using Cross-attention, and then the fused features are input into a 6-layer Transformer Encoder Layer to obtain new features; the obtained features are input into an MLP in the second step for prediction and to obtain results. Through the diversity of feature extraction, the method of deep learning improves the accuracy and generalization of the model.

[0105] In one embodiment, in order to verify the effect of the anti-cancer peptide prediction method application model based on feature fusion and cross attention mechanism provided in the present application, five commonly used statistical indicators are used: accuracy (ACC), Matthews correlation coefficient (MCC), sensitivity (Sn) and specificity (Sp). In addition, considering the imbalance problem of the training and test data sets of the model in the present application, the area under the curve (AUC) and the F1 score are also included in the evaluation. The calculation formulas of these indicators are as follows:

[0106]

[0107] Wherein, TP, TN, FP and FN represent the counts of true positives, true negatives, false positives and false negatives, respectively. In addition, in the present embodiment, the precision-recall curve is also calculated to further evaluate the performance of the model, and these indicators provide a comprehensive evaluation of the classification ability of the model, especially in the case of unbalanced data set, the traditional accuracy may not fully capture the subtle differences in model performance.

[0108] In one embodiment, in order to demonstrate the superiority of the anti-cancer peptide prediction method application model ACP-ETBLCA based on feature fusion and cross attention mechanism provided in the present application, the ACP-ETBLCA model is compared with the existing model on the ACP135 and ACP99 data sets. Among them, the comparison of the constructed ACP-ETBLCA model and the existing model on the ACP135 data set is as follows:

[0109] Comparison of the constructed model with existing models on the ACP135 dataset

[0110]

[0111] Among them, the comparison between the constructed ACP-ETBLCA model and the existing model on the ACP99 dataset is as follows:

[0112] Comparison of the constructed model with existing models on the ACP99 dataset

[0113]

[0114]

[0115] It can be seen that ACP-ETBLCA is basically superior to all previously constructed models in terms of indicators such as ACC (accuracy), Sp (specificity) and MCC (Matthews correlation coefficient), which shows that iBitter-GRE effectively mines the local and global information of peptide sequences and demonstrates excellent ability in predicting bitter peptides.

[0116] Specifically, when analyzing the prediction results of ACP135 and ACP99 samples, we can observe the distribution characteristics of the two. Figure 4 As shown in Figure A, for the ACP135 sample, the predicted probabilities of negative samples are mostly concentrated in the area close to 0, and the probability of a small number of misclassified negative samples is slightly higher than 0.5. Although the positive samples are slightly scattered, they are mainly concentrated in the position close to 1, and the probability of misclassification is low. In contrast, Figure 4 As shown in Figure B, the predicted probability of negative samples for ACP99 samples is also concentrated near 0, but it misclassifies fewer negative samples and has a significantly higher predicted probability of positive samples, concentrated in an area close to 1. This explains the higher accuracy of ACP99 compared to ACP135.

[0117] In further performance evaluation, as Figure 4 The ROC curves for the models shown in Figures C and D show that ACP135 achieves an AUC of 0.9589, while ACP99 achieves an AUC of 0.9893, demonstrating that ACP99 has slightly better discriminatory capabilities. The PR curves further reveal that ACP99 outperforms ACP135, with an average precision (AP) of 0.9905, compared to ACP135's 0.9380. While both significantly outperform a random classifier, ACP99 demonstrates a more consistent advantage in balancing precision and recall.

[0118] Specifically, the measurements of different ESM-2 parameters under the six indicators of pre, F1, recall, Sp, MCC, and ACC are as follows: Figure 5As shown in the table, using ESM-2 with 650M parameters can make the model have higher average pre, F1, recall, Sp, MCC, ACC and more concentrated data under 10-fold cross-validation, better generalization ability, and better classification effect of the model.

[0119] As shown in the table, using BiLSTM to continuous the features extracted by the traditional feature extraction method can make the model have higher average pre, F1, recall, Sp, MCC, ACC and more concentrated data under 10-fold cross-validation, better generalization ability, and better classification effect of the model. Figure 6

[0120] As shown in the table, using the features obtained by BiLSTM as Q in cross-attention, i.e., obtaining structure features containing physical and chemical information, can make the model have higher average F1, recall, MCC, ACC and more concentrated data under 10-fold cross-validation, better generalization ability, which shows that using the features obtained by BiLSTM as cross-attention can improve the classification ability of the model. Figure 7

[0121] It can be found that the number of layers of the transformer encoder layer has a significant impact on the generalization ability of the model. In the case of using 6 layers (the same as the transformer encoder layer used in the 28M parameter version of ESM-2), the result distribution of 10-fold cross-validation is more concentrated, which shows that the model has better generalization ability. In this embodiment, T-SNE graph can also be used for visual analysis of each block of the model. Figure 8

[0122] As shown in the table, after the data in ACP135 is extracted by the traditional feature extraction model, BiLSTM, ESM-2, linear layer, and cross-attention, the classification of the samples is as shown in the table. Figure 9 As shown in the table, by comparing (A) and (B), it can be found that the classification of the samples has been significantly improved after using BiLSTM. Figure 9 As shown in the table, by comparing (C) and (D), it can be found that using linear layer only changes the dimension of the features extracted by ESM-2, and does not change the quality of the features. Figure 9 As shown in the table, by comparing (E), it can be found that after using cross-attention, the samples have been significantly classified. Figure 9 Figure 9 Figure 9

[0123] ​​​​​​After the data in ACP135 passes through 6 layers of transformer encoder layer and average pooling layer, the classification of samples is as follows Figure 10 As shown by Figure 10 Comparing (A), (B), (C), (D), (E), and (F), we can see that the classification effect without a transformer encoder layer is better. Figure 10 (G) It can be found that the use of average pooling further improves the classification effect.

[0124] It should be understood that, although the various steps in the above flow chart are shown in sequence as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be performed in other orders. Moreover, at least a portion of the steps in the above flow chart may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily to be performed in sequence, but can be performed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0125] In one embodiment, Figure 11 As shown, an anticancer peptide prediction system based on feature fusion and cross-attention mechanism is provided, including: a data collection module 1110, a structural feature extraction module 1120, other feature extraction module 1130, a feature processing module 1140, a feature fusion module 1150 and a prediction module 1160, wherein:

[0126] A data collection module 1110 is configured to collect anticancer peptide sequences and non-anticancer peptide sequences from a database to construct a data set containing various protein sequences;

[0127] Structural feature extraction module 1120, for inputting protein sequences into protein language model ESM-2 and extracting peptide structural features in the protein sequences through Transformer encoder;

[0128] Other feature extraction modules 1130 are used to input the protein sequence into the feature extraction model and extract the physical and chemical features of peptides in the protein sequence through the feature extraction model;

[0129] The feature processing module 1140 is used to perform dimensionality transformation processing on the peptide structural features to obtain processed peptide structural features; use BiLSTM to continuousize the discrete peptide physicochemical features to obtain continuous peptide physicochemical features;

[0130] The feature fusion module 1150 is configured to fuse the processed peptide structure features and the continuous peptide physicochemical features by using a cross-attention mechanism to obtain target features.

[0131] The prediction module 1160 is configured to input the target features into a multi-layer perception (MLP) to obtain an anticancer peptide prediction result.

[0132] In an embodiment, the data collection module 1110 is further configured to collect anticancer peptide sequences verified by experiments from a database, perform de-redundancy processing on the anticancer peptide sequences by using a CD-HIT tool, and filter to obtain final anticancer peptide sequences; randomly collect non-anticancer peptide sequences, and extract final non-anticancer peptide sequences from the non-anticancer peptide sequences by using homology bias removal and PSSM extraction standards; and randomly select anticancer peptide sequences and non-anticancer peptide sequences from the final anticancer peptide sequences and the final non-anticancer peptide sequences respectively to construct a dataset containing each protein sequence.

[0133] In an embodiment, the structure feature extraction module 1120 is further configured to input the protein sequence into a protein language model ESM-2, convert the protein sequence into a numerical vector representation by using the protein language model, input the numerical vector representation into a Transformer encoder in the protein language model ESM-2, and calculate by using a dot-product attention mechanism and a linear layer in the Transformer encoder to obtain peptide structure features.

[0134] In an embodiment, the other feature extraction module 1130 is further configured to input the protein sequence into a feature extraction model, represent each amino acid in the protein sequence by using one-hot encoding in the feature extraction model to obtain a binary vector corresponding to the protein sequence, calculate the sum of element masses of each amino acid in the protein sequence by using the feature extraction model, and take the sum of element masses as a molecular weight, obtain the acid-base value of the amino acid in the protein sequence, calculate the isoelectric point according to the acid-base value, calculate the number and properties of the hydrophobic group of the amino acid in the protein sequence, determine the hydrophobicity of the amino acid based on the number and properties of the hydrophobic group, and take the binary vector, the molecular weight, the isoelectric point, and the hydrophobicity as peptide physicochemical features in the protein sequence.

[0135] In an embodiment, the feature processing module 1140 is further configured to input the peptide structure features into a linear layer for dimension transformation processing to obtain processed peptide structure features, and input the peptide physicochemical features into a bidirectional long short-term memory network (BiLSTM) to capture long-distance dependency relationships of long sequences in the peptide physicochemical features by using the BiLSTM, complete the continuous of the peptide physicochemical features, and obtain continuous peptide physicochemical features.

[0136] In an embodiment, the feature fusion module 1150 is further configured to generate a query matrix, a key matrix, and a value matrix by using the processed peptide structure features and the continuous peptide physicochemical features respectively through a cross-attention mechanism; determine a key vector dimension corresponding to the processed peptide structure features and the continuous peptide physicochemical features; calculate cross-attention according to the query matrix, the key matrix, the value matrix, and the key vector dimension; and complete feature fusion of the processed peptide structure features and the continuous peptide physicochemical features based on the cross-attention to obtain target features.

[0137] In an embodiment, the feature fusion module 1150 is further configured to input the target features into a Transformer architecture, calculate a query vector, a key vector, and a value vector for each position by using a multi-head self-attention mechanism in the Transformer architecture; perform weighted average calculation on the query vector, the key vector, and the value vector to obtain weighted features; and perform position-by-position nonlinear conversion on the weighted features by using a feedforward neural network.

[0138] In an embodiment, the prediction module 1160 is further configured to input the target features into an input layer of a multi-layer perception (MLP), perform result prediction by using a hidden layer and a linear layer in the MLP; and output a prediction result from an output layer of the MLP to obtain an anticancer peptide prediction result.

[0139] In an embodiment, a computer device is provided, which can be a terminal. An internal structure diagram of the computer device can be as shown in FIG. 8. Figure 12 The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is configured to communicate with external terminals through a network connection. The computer program is executed by the processor to implement an anticancer peptide prediction method based on feature fusion and a cross-attention mechanism. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or can be a key, trackball, or touchpad arranged on the shell of the computer device, or can be an external keyboard, touchpad, or mouse, etc.

[0140] Those skilled in the art can understand that, Figure 12The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0141] In one embodiment, a computer device is provided, comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the anti-cancer peptide prediction method based on feature fusion and cross-attention mechanism when executing the computer program.

[0142] In one embodiment, a computer readable storage medium is provided, storing a computer program, and the computer program implements the steps of the anti-cancer peptide prediction method based on feature fusion and cross-attention mechanism when executed by a processor.

[0143] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium and can include the processes of the above-mentioned embodiments when executed. Any reference to memory, storage, database or other medium used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0144] The technical features of the above embodiments can be combined in any way. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

[0145] The above-described embodiments are merely illustrative of several embodiments of the present application, which are described in more detail and in a specific and detailed manner, but should not be construed as limiting the scope of the patent. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these are all within the scope of the present application. Therefore, the scope of protection of the patent of the present application should be subject to the appended claims.

Claims

1. A method for predicting anticancer peptides based on feature fusion and cross-attention mechanism, characterized in that: The method comprises: Anti-cancer peptide sequences and non-anti-cancer peptide sequences are collected from the database to construct a data set containing various protein sequences; Inputting the protein sequence into the protein language model ESM-2, and extracting peptide structural features in the protein sequence through the Transformer encoder; Inputting the protein sequence into a feature extraction model, and extracting the physicochemical characteristics of peptides in the protein sequence through the feature extraction model; Performing dimensionality transformation processing on the peptide structural features to obtain processed peptide structural features; using BiLSTM to continuousize the discrete peptide physicochemical features to obtain continuous peptide physicochemical features; Using a cross-attention mechanism to fuse the processed peptide structural features and the continuous peptide physicochemical features to obtain target features; The target features are input into a multi-layer perceptron (MLP) to obtain the anticancer peptide prediction results.

2. The anticancer peptide prediction method based on feature fusion and cross-attention mechanism according to claim 1, characterized in that: Anti-cancer peptide sequences and non-anti-cancer peptide sequences were collected from the database to construct a dataset containing various protein sequences, including: Collecting experimentally verified anticancer peptide sequences from the database, using the CD-HIT tool to perform redundancy processing on the anticancer peptide sequences, and filtering to obtain the final anticancer peptide sequences; randomly collecting non-anticancer peptide sequences, and extracting final non-anticancer peptide sequences from the non-anticancer peptide sequences using homology deviation removal and PSSM extraction criteria; Anticancer peptide sequences and non-anticancer peptide sequences are randomly selected from the final anticancer peptide sequences and the final non-anticancer peptide sequences respectively to construct a data set containing various protein sequences.

3. The anticancer peptide prediction method based on feature fusion and cross-attention mechanism according to claim 1, characterized in that: The protein sequence is input into the protein language model ESM-2, and the peptide structural features in the protein sequence are extracted through the Transformer encoder, including: Inputting the protein sequence into a protein language model ESM-2, and converting the protein sequence into a numerical vector representation by the protein language model; Inputting the numerical vector representation into the Transformer encoder in the protein language model ESM-2; The peptide structural features are obtained by performing calculations through the point-wise attention mechanism and linear layer in the Transformer encoder.

4. The anticancer peptide prediction method based on feature fusion and cross-attention mechanism according to claim 1, characterized in that: The protein sequence is input into a feature extraction model, and the physicochemical characteristics of peptides in the protein sequence are extracted by the feature extraction model, including: Inputting the protein sequence into a feature extraction model, using one-hot encoding in the feature extraction model to represent each amino acid in the protein sequence by one-hot encoding, to obtain a binary vector corresponding to the protein sequence; Calculating the sum of all element masses of each amino acid in the protein sequence using the feature extraction model, and taking the sum of the element masses as the molecular weight; Obtaining the pH values ​​of amino acids in the protein sequence, and calculating the isoelectric point based on the pH values; calculating the number and properties of amino acid hydrophobic groups in the protein sequence, and determining the hydrophobicity of the amino acid based on the number and properties of the hydrophobic groups; The binary vector, molecular weight, isoelectric point, and hydrophobicity are used as physicochemical characteristics of peptides in protein sequences.

5. The anticancer peptide prediction method based on feature fusion and cross-attention mechanism according to claim 1, characterized in that: Performing dimensionality transformation on the peptide structural features to obtain processed peptide structural features; BiLSTM is used to continuousize the discrete physicochemical features of peptides to obtain continuous physicochemical features of peptides, including: Inputting the peptide structural features into a linear layer for dimensionality transformation processing to obtain processed peptide structural features; The peptide physicochemical characteristics are input into a bidirectional long short-term memory network BiLSTM, and the long-range dependencies of long sequences in the peptide physicochemical characteristics are captured by BiLSTM, thereby completing the continuity of the peptide physicochemical characteristics and obtaining continuous peptide physicochemical characteristics.

6. The anticancer peptide prediction method based on feature fusion and cross-attention mechanism according to claim 1, characterized in that: The processed peptide structural features and the continuous peptide physicochemical features are fused using a cross-attention mechanism to obtain target features, including: Using the processed peptide structural features and the continuous peptide physicochemical features through a cross-attention mechanism, a query matrix, a bond matrix, and a value matrix are generated respectively; Determining bond vector dimensions corresponding to the structural features of the processed peptides and the continuous physicochemical features of the peptides; Calculating cross attention based on the query matrix, key matrix, value matrix, and key vector dimensions; The processed peptide structural features and the continuous peptide physicochemical features are fused based on the cross attention to obtain target features.

7. The anticancer peptide prediction method based on feature fusion and cross-attention mechanism according to claim 6, characterized in that: The method further comprises: Input the target features into the Transformer architecture, and calculate the query vector, key vector, and value vector for each position through the multi-head self-attention mechanism in the Transformer architecture; Performing weighted average calculation on the query vector, key vector, and value vector to obtain weighted features; A feedforward neural network is used to perform position-by-position nonlinear transformation on the weighted features.

8. The anticancer peptide prediction method based on feature fusion and cross-attention mechanism according to claim 1, characterized in that: The target features are input into a multi-layer perceptron (MLP) to obtain the anticancer peptide prediction results, including: Input the target feature into the input layer of a multi-layer perceptron (MLP), and perform result prediction through the hidden layer and linear layer in the multi-layer perceptron (MLP); The prediction result is output from the output layer of the multi-layer perceptron MLP to obtain the anticancer peptide prediction result.

9. An anticancer peptide prediction system based on feature fusion and cross-attention mechanism, characterized in that: The system comprises: A data collection module is used to collect anticancer peptide sequences and non-anticancer peptide sequences from a database to construct a data set containing various protein sequences; A structural feature extraction module is used to input the protein sequence into the protein language model ESM-2 and extract the peptide structural features in the protein sequence through the Transformer encoder; Another feature extraction module is used to input the protein sequence into a feature extraction model, and extract the physicochemical features of peptides in the protein sequence through the feature extraction model; A feature processing module is used to perform dimensionality transformation processing on the peptide structural features to obtain processed peptide structural features; and use BiLSTM to continuousize the discrete peptide physicochemical features to obtain continuous peptide physicochemical features; A feature fusion module is used to fuse the processed peptide structural features and the continuous peptide physicochemical features using a cross-attention mechanism to obtain target features; The prediction module is used to input the target features into a multi-layer perceptron (MLP) to obtain an anticancer peptide prediction result.

10. The anticancer peptide prediction system based on feature fusion and cross-attention mechanism according to claim 9, characterized in that: The data collection module is further used to collect experimentally verified anticancer peptide sequences from a database, perform redundancy processing on the anticancer peptide sequences using the CD-HIT tool, and filter them to obtain final anticancer peptide sequences; randomly collect non-anticancer peptide sequences, and extract final non-anticancer peptide sequences from the non-anticancer peptide sequences using homology deviation removal and PSSM extraction standards; and randomly select anticancer peptide sequences and non-anticancer peptide sequences from the final anticancer peptide sequences and the final non-anticancer peptide sequences, respectively, to construct a data set containing various protein sequences.

Images