Anticancer peptide recognition method and system based on attention mechanism and multi-granularity hierarchical features

Abstract

The present invention constructs a method and system for identifying anticancer peptides based on an attention mechanism and multi-granularity hierarchical features. This method is of great significance in the field of anticancer peptide identification technology. First, atomic-level features are learned through transfer learning, revealing potential features that were difficult to discover in previous work. Secondly, the ChouFasman algorithm is used to represent the secondary structure in the amino acid sequence layer extraction process, which increases the richness of information. Next, the problem that the existing technology cannot capture the high-order structural similarity of anticancer peptide sequences is solved by constructing a hypergraph, and the importance of subsequences to the overall sequence is learned through the attention mechanism. Finally, the multi-granularity hierarchical features are fused, so that the model can fully understand and describe the characteristics of the anticancer peptide sequence. This invention provides a more accurate and comprehensive analysis tool for the discovery and design of anticancer peptides, and can contribute to the development of anticancer drugs.

Description

Technical Field

[0001] The present invention belongs to the technical field of anticancer peptide identification, and in particular relates to a method and system for identifying anticancer peptides based on attention mechanism and multi-granularity hierarchical features. Background Art

[0002] Cancer is a disease caused by the uncontrolled proliferation of abnormal cells. These cells, which proliferate indefinitely, invade and destroy normal tissues and organs, threatening people's lives and health. Currently, chemotherapy and radiotherapy are commonly used clinically to treat cancer. These drugs primarily kill proliferating cancer cells. However, because these drugs cannot completely distinguish between cancer cells and healthy human cells, they can cause irreversible damage to normal cells and tissues while killing cancer cells. Patients then experience a series of side effects from radiotherapy or chemotherapy. Furthermore, chemotherapy can also lead to drug resistance. Therefore, there is an urgent need to develop new, effective anticancer drugs with minimal side effects to reduce patients' suffering and prolong their lives.

[0003] Anticancer peptides (ACPs) are bioactive peptides that naturally exist in a variety of animals, plants, and microorganisms and have anti-cancer effects. In recent years, anticancer peptides have become a new technology for cancer treatment due to their unique mechanism of action, attracting widespread attention from researchers. Anticancer peptides inactivate cancer cells through their unique membrane lytic mechanisms (membranolytic mechanismsms) or non-membranolytic mechanisms (non-membranolytic mechanisms), accurately identify cancer cells, and are non-toxic to healthy cells. The mechanism of action and anti-tumor activity exhibited by anticancer peptides provide a new treatment option for cancer. Using anticancer peptides as anticancer drugs has lower treatment costs and fewer side effects.

[0004] However, the traditional wet lab method can only identify a small number of anticancer peptides in a large protein library, which is very time-consuming and expensive. Currently, computer-based recognition methods have become an important method for identifying anticancer peptides. They are more cost-effective and efficient, and various recognition and classification algorithms have emerged one after another, showing good prospects. However, the identification of anticancer peptides based on computational technology still needs to overcome various technical challenges, such as (1) how to design a reasonable data representation method to carry more possible features; (2) how to design a model to fully extract the potential features of peptide sequences.

[0005] Existing methods for detecting anticancer peptides include:

[0006] 1. Machine Learning (ML)-based methods, such as support vector machines, random forests, and logistic regression. In ML-based methods, researchers use handcrafted features as input into traditional machine learning models to identify anticancer peptides. For example, they use amino acid composition and binary profiles to construct features, which are then passed into a support vector machine model for feature extraction. This type of method can extract features of anticancer peptides, but because it requires manual collection and production of features such as the physicochemical properties of the anticancer peptides, followed by training and testing, the computational complexity is extremely high, making it difficult to apply to the increasingly large protein library.

[0007] 2. A multi-classifier ensemble approach. This method analyzes different feature representation methods and experimentally demonstrates the performance of different combinations. It then combines the results of multiple classifiers using different optimization algorithms. This model combines the strengths of most single machine learning algorithms to fully exploit the potential features of anticancer peptide sequences. Experiments have shown that this model performs outstandingly in anticancer peptide classification tests, outperforming existing single classifiers. However, this approach requires simultaneous training of multiple classifiers, and the model construction process is cumbersome and difficult to implement.

[0008] 3. Methods based on neural networks and deep learning, such as MLP (multilayer perception), LSTM (long short-term memory), graph convolutional networks (GCNs), and self-attention mechanisms. These methods utilize neural networks for feature extraction and representation. Their encoding is simple and effective, eliminating the need for feature engineering while increasing feature extraction capabilities. The self-attention mechanism, among others, effectively mines potential associations between non-adjacent sites by learning contextual features. In a graph convolutional network, each anticancer peptide sequence is represented as a graph node, transforming the anticancer peptide sequence classification task into a graph classification task. The GCN automatically learns node features and predicts anticancer peptides. While these efforts have significantly advanced the detection of anticancer peptides, they still face challenges such as complex encoding, long training times, and the potential for information redundancy. Furthermore, these graph learning tasks typically employ simple regular graphs, considering only binary relationships between nodes. This fails to capture potentially more complex relationships and fails to assess the importance of each subsequence within the sequence.

[0009] Through the above analysis, the problems and defects of the existing technology are as follows:

[0010] (1) Most methods require manual feature representation, which is computationally intensive and prone to information redundancy.

[0011] (2) Using traditional machine learning methods (support vector machine, random forest, etc.), the recognition accuracy is not high;

[0012] (3) Existing methods have never explored features at the atomic level smaller than the sequence, which may contain more potential features;

[0013] (4) Most existing graph-based methods operate on a simple regular graph, only considering binary relationships between nodes and failing to capture more complex relationships that may exist. Summary of the Invention

[0014] In response to the problems existing in the prior art, the present invention provides a method and system for identifying anticancer peptides based on an attention mechanism and multi-granularity hierarchical features.

[0015] The present invention is implemented as follows: first, atomic-level features are learned through transfer learning; then, amino acid sequences and protein secondary structure features are learned through convolutional neural networks and self-attention mechanisms; finally, the relationship features between multiple sequences are captured through a hypergraph attention network model.

[0016] Furthermore, the specific steps of the anticancer peptide identification method based on the attention mechanism and multi-granularity hierarchical features include:

[0017] Step 1: Obtain anticancer peptide dataset;

[0018] Step 2: Atomic-level feature extraction: The obtained atomic sequence features are passed into the molecular pre-trained neural network model to obtain atomic-level features;

[0019] Step 3: Feature extraction at the amino acid sequence level: The obtained anticancer peptide dataset is sequence data. The original sequence features and secondary structure features are extracted using different methods. The obtained sequence features and secondary structure features are then concatenated to obtain the final amino acid sequence features.

[0020] Step 4: Extract multi-sequence similarity relationship features: Divide the subsequences, construct a hypergraph, and use the hypergraph attention feature extraction model to learn multi-sequence similarity relationship features;

[0021] Step 5: Multi-level feature fusion extraction: The atomic layer features, amino acid sequence layer features, and multi-sequence relationship layer features obtained separately are spliced ​​together, and then passed into the extractor composed of a multi-layer linear network for final fusion extraction, and then the final category probability is output.

[0022] Furthermore, the atomic sequence features in step 2 are obtained as follows:

[0023] After obtaining the anticancer peptide dataset, the amino acid sequence was parsed using the online server Phyre2, and each amino acid sequence was decomposed into an atomic sequence with atomic groups as the smallest granularity.

[0024] Furthermore, the molecular pre-training neural network model in the step 2 adopts a variant of the encoder, VQ-VAE, as a context-aware tagger, which encodes atoms into meaningful discrete values ​​to expand the atomic vocabulary and alleviate the quantitative differences between atoms. By expanding the "vocabulary" of atoms, a new node-level pre-training task called masked atomic modeling (MAM) is proposed, and masked discrete values ​​are predicted by pre-training a graph neural network.

[0025] Furthermore, the specific steps of extracting sequence features in step three include:

[0026] The amino acid sequence is encoded using different sequence encoding methods, and the sequence feature vector is obtained by combining different encoding methods. The sequence features are then extracted through multi-layer convolution.

[0027] Furthermore, the specific steps of extracting the secondary structure features in step three include:

[0028] The secondary structure sequence of amino acids is obtained through the ChouFasman algorithm, and then passed into the multi-layer self-attention mechanism for secondary structure feature extraction;

[0029] Furthermore, the specific steps of constructing the hypergraph in step 4 include:

[0030] First, the sequence is decomposed into a set of subsequences as the main feature representation of the sequence. Then, this set of subsequences is represented as nodes of a hypergraph, and each sequence containing a set of subsequences is a hyperedge; each hyperedge can be connected to other hyperedges through some shared nodes as subsequences.

[0031] Furthermore, the hypergraph attention feature extraction model in step 4 includes a two-level attention network optimized using class labels, and the two-level attention networks are an attention aggregation layer for hyperedge features and an attention aggregation layer for node features.

[0032] Another object of the present invention is to provide an anticancer peptide recognition system based on an attention mechanism and multi-granularity hierarchical features, the anticancer peptide recognition system based on an attention mechanism and multi-granularity hierarchical features comprising:

[0033] Data acquisition module, used to obtain anticancer peptide datasets;

[0034] Atomic-level feature extraction module, used to pass the obtained atomic sequence features into the pre-trained model to obtain atomic-level features;

[0035] The amino acid sequence feature extraction module is used to extract sequence features and secondary structure features, and to combine the obtained sequence features and secondary structure features to obtain the final amino acid sequence features;

[0036] The multi-sequence similarity relationship feature extraction module is used to divide subsequences, construct hypergraphs, and use the hypergraph attention feature extraction model to learn multi-sequence similarity relationship features;

[0037] The multi-level feature fusion extraction module is used to splice the atomic layer features, amino acid sequence layer features and multi-sequence relationship layer features obtained separately, and then perform the final fusion extraction to output the final category probability.

[0038] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0039] First, in view of the technical problems existing in the above-mentioned prior art and the difficulty of solving these problems, this paper closely combines the technical solutions to be protected by the present invention and the results and data during the research and development process, and analyzes in detail and in depth how the technical solutions of the present invention solve the technical problems and some creative technical effects brought about by solving the problems. The specific description is as follows:

[0040] This paper constructs a new anticancer peptide identification method based on attention mechanism and multi-granularity hierarchical features, which can extract features at a smaller atomic level and help discover potential features that were difficult to find in previous work;

[0041] The present invention uses the ChouFasman algorithm to represent the secondary structure of the sequence during the amino acid sequence layer extraction process, making the carried information richer;

[0042] The present invention can solve the problem that the existing technology cannot capture the high-order structural similarity between anticancer peptide sequences by constructing a hypergraph, and solve the problem that the existing technology cannot learn the importance of subsequences to sequences by using an attention mechanism;

[0043] The present invention integrates hierarchical features of three granularities to make the model more comprehensive in feature extraction of peptide sequences. For the trained network, it can provide good technical support for the identification and detection of anticancer peptides and is simple and convenient to use.

[0044] The present invention further decomposes the amino acid sequence into an atomic sequence with atomic groups as the smallest granularity, uses transfer learning and adopts the latest molecular pre-training model Mole-BERT to extract the atomic-level features of the amino acid sequence. In the process of extracting amino acid sequence features, the present invention uses the ChouFasman algorithm to predict the secondary structure and further extract it, making the information carried richer; the present invention proposes a new anti-cancer peptide identification method based on attention mechanism and multi-granularity hierarchical features, which extracts three layers of features, namely atomic layer features, amino acid sequence layer features and multi-sequence relationship layer features, and then fuses and extracts the three hierarchical features to finally obtain the category probability for sequence classification. This hierarchical feature fusion method is more conducive to extracting the potential features of peptide sequences.

[0045] Second, considering the technical solution as a whole or from the perspective of the product, the technical effects and advantages of the technical solution to be protected by the present invention are described in detail as follows:

[0046] (1) The present invention can extract features of the amino acid atomic composition of anticancer peptide sequences based on transfer learning, and obtain features of peptide sequences with atoms as the smallest granularity;

[0047] (2) The present invention can use convolutional neural networks and self-attention mechanisms to extract features of the amino acid sequence itself and its secondary structure, thereby obtaining the features of the amino acid sequence itself;

[0048] (3) The present invention can capture the relationship characteristics between anticancer peptide sequences by constructing a hypergraph;

[0049] (4) The present invention can construct a hierarchical peptide sequence feature extraction algorithm by extracting features from three layers of peptide sequences (atomic layer, amino acid sequence layer and multi-sequence relationship layer).

[0050] Third, as auxiliary evidence for the inventiveness of the claims of the present invention, it is also reflected in the following important aspects:

[0051] The technical solution of the present invention solves the technical problems that people have been eager to solve but have never been able to solve successfully:

[0052] This invention introduces an innovative method for extracting smaller, atomic-scale features that were previously difficult to detect. By analyzing the atomic composition of proteins, this method provides a deeper understanding of their structure and function, revealing hidden underlying features. This fine-scale feature extraction helps deepen our understanding of proteins and opens new possibilities for further research and applications.

[0053] ② The present invention utilizes the ChouFasman algorithm during amino acid sequence extraction. Based on the secondary structure of amino acid sequences, this algorithm identifies features such as helices, folds, and disordered regions within proteins, providing a more refined representation of the sequence. This approach makes the extracted features more informative and more accurately reflects the structure and function of proteins, providing a more reliable foundation for subsequent analysis and prediction.

[0054] ③ The present invention also addresses the limitations of existing technologies in capturing high-order structural similarities between anticancer peptide sequences. By constructing a hypergraph, the present invention can more comprehensively consider the interactions and connections between peptide chains, thereby better capturing the high-order structural information of peptide sequences. At the same time, the present invention also introduces an attention mechanism that can effectively learn and focus on important subsequences in the sequence, further improving the ability to capture sequence features. This combined approach enables the present invention to more accurately assess the similarities between anticancer peptide sequences and provide more valuable clues for drug design and discovery.

[0055] Finally, the present invention further enhances the model's ability to extract peptide sequence features by integrating hierarchical features at three granularities. This comprehensive approach, which considers features at different granularities, enables the model to comprehensively understand and describe the properties and characteristics of peptide sequences from multiple perspectives. By fully utilizing information at different levels, the present invention can more accurately characterize the complexity of peptide sequences, providing a more in-depth analysis and application tool for related research fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Schematic diagram of the anticancer peptide identification method based on attention mechanism and multi-granularity hierarchical features provided by an embodiment of the present invention;

[0057] Figure 2 This is a flow chart of the anticancer peptide identification method based on the attention mechanism and multi-granularity hierarchical features provided by an embodiment of the present invention;

[0058] Figure 3 is a schematic diagram of the k-mers technology provided by an embodiment of the present invention;

[0059] Figure 4 This is a diagram showing the principle of constructing a hypergraph provided by an embodiment of the present invention. Implementation Method

[0060] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the 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.

[0061] like Figure 1As shown, the anticancer peptide identification method based on an attention mechanism and multi-granularity hierarchical features provided by an embodiment of the present invention incorporates three layers of features: atomic-level features, amino acid sequence-level features, and multi-sequence relationship features. Atomic-level features are learned through transfer learning, followed by amino acid sequence and protein secondary structure features using a convolutional neural network and self-attention mechanism. Finally, multi-sequence relationship features are captured using a hypergraph attention network model. This hierarchical anticancer peptide feature extraction algorithm, constructed in this way, offers enhanced feature extraction capabilities and can provide technical support and useful information to relevant personnel.

[0062] like Figure 2 As shown, the specific steps of the anticancer peptide identification method based on the attention mechanism and multi-granularity hierarchical features provided by the embodiment of the present invention include:

[0063] S101, acquisition of anticancer peptide dataset;

[0064] S102, atomic-level feature extraction: The obtained atomic sequence features are passed into the molecular pre-trained neural network model to obtain atomic-level features;

[0065] S103, feature extraction at the amino acid sequence level: The obtained anticancer peptide dataset is sequence data, and feature extraction is performed through sequence feature extraction and secondary structure feature extraction. The obtained sequence features and secondary structure features are spliced ​​to obtain the final amino acid sequence features;

[0066] S104, multi-sequence similarity relationship feature extraction: divide the subsequences, construct a hypergraph, and use the hypergraph attention feature extraction model to learn multi-sequence similarity relationship features;

[0067] S105, multi-level feature fusion extraction: The atomic layer features, amino acid sequence layer features and multi-sequence relationship layer features obtained respectively are spliced, and then passed into the extractor composed of a multi-layer linear network for final fusion extraction, and then the final category probability is output.

[0068] The technical solution of the present invention is further described below in conjunction with specific embodiments.

[0069] 1. Obtaining anticancer peptide dataset

[0070] In order to fully analyze the anticancer peptide data, the present invention obtains a large amount of data from different literature, including LEE, Independent, ACPFL-500, ACPFL-164, ACPred-Fuse 500, ACPred-Fuse 2710, AntiCP2.0 main dataset (AntiCP 2.0 Main) and AntiCP 2.0 alternative dataset (AntiCP 2.0 Alternate). Among them, ACP500 is used for training and consists of 250 positive samples and 250 negative samples. ACP164 is used as an independent test set and consists of 82 positive samples and 82 negative samples. The AntiCP 2.0 main dataset and alternative dataset have 861 and 970 samples, respectively.

[0071] 2. Atomic-level feature extraction

[0072] After obtaining the anticancer peptide dataset, the amino acid sequence was parsed using the online server Phyre2, which can decompose each amino acid sequence into an atomic sequence with atomic groups as the smallest granularity.

[0073] Mole-BERT is a newly proposed molecule pre-training neural network. It uses a variant of the encoder, VQ-VAE, as a context-aware tokenizer to encode atoms as meaningful discrete values. This expands the atomic vocabulary and mitigates quantitative differences between atoms. By expanding the atomic vocabulary, it proposes a new node-level pre-training task called Masked Atom Modeling (MAM), and uses a pre-trained graph neural network to predict masked discrete values. In this paper, the obtained atomic sequence features are passed to this pre-trained model to obtain atomic-level features.

[0074] 3. Feature extraction at the amino acid sequence level

[0075] The obtained anticancer peptide dataset is sequence data. The feature extraction of this part is mainly divided into two parts: sequence feature extraction and secondary structure feature extraction.

[0076] (1) Sequence feature extraction

[0077] Amino acid sequences are encoded using various sequence encoding methods, such as AAINDEX, BLOSUM62, and BPF. AAINDEX is a database that records numerical indicators representing various physicochemical and biochemical properties of amino acids or amino acid pairs. BLOSUM62 is a set of feature matrices that indicates the mutation probability between amino acids throughout evolution. BPF, or hot-spot encoding, is the most direct way to encode peptide sequence residues and has been widely used in machine learning-based peptide prediction models. Combining different encoding methods can generate sequence feature vectors, which are then extracted through multi-layer convolution.

[0078] (2) Secondary structure feature extraction

[0079] The ChouFasman algorithm is a probabilistic protein structure prediction method. It analyzes the relative frequencies of each amino acid in α-helices, β-sheets, and turns in known protein structures. From these frequencies, it derives a set of probability parameters for the presence of each amino acid in each secondary structure type. These parameters are used to predict the probability of a given amino acid sequence forming α-helices, β-sheets, β-turns, and random coils. This method accurately reflects the formation process of protein secondary structure. The ChouFasman algorithm obtains the secondary structure sequence of amino acids, which is then fed into a multi-layer self-attention mechanism for secondary structure feature extraction.

[0080] Finally, the obtained sequence features and secondary structure features are spliced ​​together to obtain the final amino acid sequence features.

[0081] 4. Multiple sequence similarity relationship feature extraction

[0082] (1) Divide subsequences

[0083] K-mers technology is a popular and effective tool widely used in biological sequence data analysis (such as sequence matching). It divides the sequential input sequence data into a list of overlapping subsequence strings of length k. It is very similar to the n-gram model used in natural language processing. The specific method of generating all K-mers from the input string sequence is to start from the first K characters, then move only one character to get the next subsequence, and so on. Figure 3 Taking "M" in the sequence "PAMNH" as an example, the 1-mer is divided into: "M", the 2-mers are divided into: "AM, MN", and the 3-mers are divided into: "PAM, AMN, MNH". In the present invention, the 5-mer segmentation method is mainly used.

[0084] (2) Constructing a hypergraph

[0085] In order to learn the similarity of sequences for sequence classification, it is necessary to define the relationship between sequences and the relationship between subsequences within each sequence. To this end, a hypergraph can be constructed to capture the high-order similarity of sequences. First, the sequence is decomposed into a group of subsequences as the main feature expression of the sequence. Then, the present invention represents this group of subsequences as nodes of the hypergraph, and each sequence containing a group of subsequences is a hyperedge. Each hyperedge can be connected to other hyperedges as subsequences through some shared nodes. Through this hypergraph construction method, the hypergraph defines a higher level of connection between sequences and subsequences, and helps to capture the complex similarities between sequences. The specific construction process is as follows Figure 4 As shown, the graph contains three sequences, namely: "SKIVGAFD", "FDGSLKA" and "VGALLDKA". These three sequences share six subsequence nodes, and the sharing relationship defines the connection relationship between sequences.

[0086] (3) Hypergraph Attention Feature Extraction Model

[0087] The present invention uses a hypergraph attention feature extraction model to learn multi-sequence similarity relationship features. The model includes a two-level attention network optimized using class labels, namely an attention aggregation layer for hyperedge features and an attention aggregation layer for node features. This attention network can not only explore the complex relationships between sequences but also help discover important subsequences of a sequence.

[0088] 5. Multi-level feature fusion extraction

[0089] The atomic layer features, amino acid sequence layer features and multi-sequence relationship layer features obtained respectively are spliced, and then passed into the extractor composed of a multi-layer linear network for final fusion extraction, and then the final category probability is output.

[0090] 2. Application Examples: In order to demonstrate the creativity and technical value of the technical solution of the present invention, this section provides application examples of the claimed technical solution on specific products or related technologies.

[0091] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the anti-cancer peptide identification method based on the attention mechanism and multi-granularity hierarchical features.

[0092] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the processor executes the steps of the anti-cancer peptide identification method based on the attention mechanism and multi-granularity hierarchical features.

[0093] An information data processing terminal is used to implement the steps of the anti-cancer peptide identification method based on the attention mechanism and multi-granularity hierarchical features.

[0094] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0095] 3. Evidence of the effects of the embodiments: The embodiments of the present invention have achieved some positive effects during the development or use process, and indeed have great advantages over the existing technology. The following content describes them with reference to the data, charts, etc. of the experimental process.

[0096] The present invention first extracts three layers of features: atomic-level features, amino acid sequence-level features, and multi-sequence relationship-level features. These features are then concatenated and fed into an extractor consisting of a multi-layer linear network for fusion extraction, ultimately outputting class probabilities. The final fusion extraction model is still under construction, so the existing experimental results for the amino acid sequence-level feature extraction model and the multi-sequence relationship-level feature extraction model are described separately.

[0097] Datasets, compared models, and evaluation metrics

[0098] ① Dataset

[0099] The datasets used in the experiment are the AntiCP 2.0 main dataset (AntiCP 2.0 Main) and the AntiCP 2.0 alternative dataset (AntiCP 2.0 Alternate). The AntiCP 2.0 main dataset and the alternative dataset have 861 and 970 samples respectively. The data can be divided into training set, validation set and test set according to the proportion.

[0100] ② Comparison model

[0101] Comparisons were made with the following models: ACPred-LAF, iACP-FSCM, ACPred, ACPred-FL, AntiCP, AntiCp_2.0, iACP, PEPred-Suite, and ACPred-Fuse.

[0102] ③Evaluation indicators

[0103] Five common indicators were used in the experiment to evaluate the performance of ACP-ST, namely accuracy (Acc), specificity (Sp), sensitivity (Sn), Matthews correlation coefficient (MCC) and AUC (area under the ROC curve) value.

[0104] Experimental results and analysis

[0105] Table 1 Comparison with multiple models

[0106] Dataset Model ACC SE SP MCC AntiCP 2.0 Alternate seqFeature_model 0.869 0.840 0.897 0.74 AntiCP 2.0 Alternate Multi_seqFeature_model 0.870 0.946 0.432 0.43 AntiCP 2.0 Alternate iACP-FSCM 0.889 0.876 0.902 0.78 AntiCP 2.0 Alternate iACP 0.776 0.784 0.768 0.55 AntiCP 2.0 Alternate PEPred-Suite 0.575 0.402 0.747 0.16 AntiCP 2.0 Alternate ACPpred-Fuse 0.789 0.644 0.933 0.60 AntiCP 2.0 Alternate ACPred-FL 0.438 0.602 0.256 -0.15 AntiCP 2.0 Alternate ACPred 0.853 0.871 0.835 0.71 AntiCP 2.0 Alternate AntiCP 0.900 0.897 0.902 0.80 AntiCP 2.0 Main seqFeature_model 0.741 0.762 0.721 0.48 AntiCP 2.0 Main Multi_seqFeature_model 0.852 0.968 0.220 0.28 AntiCP 2.0 Main AntiCP_2.0 0.754 0.775 0.734 0.51 AntiCP 2.0 Main AntiCP 0.506 1.000 0.012 0.07 AntiCP 2.0 Main ACPred 0.5347 0.8555 0.2139 0.09 AntiCP 2.0 Main iACP 0.551 0.779 0.322 0.11 AntiCP 2.0 Main PEPred-Suite 0.535 0.331 0.738 0.08 AntiCP 2.0 Main ACPpred-Fuse 0.689 0.692 0.686 0.38 AntiCP 2.0 Main ACPred-FL 0.448 0.671 0.225 -0.12

[0107] As shown in Table 1, seqFeature_model and Multi_seqFeature_model represent the amino acid sequence layer feature extraction model and the multi-sequence relationship layer feature extraction model of the present invention, respectively. In the AntiCP 2.0 Alternate dataset, Multi_seqFeature_model achieved the best specificity score result of 0.946. For other indicators in this dataset, the accuracy of the two models of the present invention was 0.869 and 0.946, respectively, which were close to the highest results. In the AntiCP 2.0Main dataset, Multi_seqFeature_model achieved the best specificity score result of 0.852 and obtained a result that was only 0.032 different from the highest sensitivity. The accuracy index of seqFeature_model was lower than that of Multi_seqFeature_model, but its other evaluation indicators were higher than those of most compared models.

[0108] The above experimental results are preliminary results of the present invention, and the model will be further improved and optimized in future work.

[0109] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for identifying anticancer peptides based on attention mechanism and multi-granularity hierarchical features, characterized in that: The anticancer peptide identification method based on attention mechanism and multi-granularity hierarchical features includes: First, atomic-level features are learned through transfer learning. Then, amino acid sequences and protein secondary structure features are learned through convolutional neural networks and self-attention mechanisms. Finally, the relationship features between multiple sequences are captured through the hypergraph attention network model. The specific steps of the anticancer peptide identification method based on the attention mechanism and multi-granularity hierarchical features include: Step 1: Obtain anticancer peptide dataset; Step 2: Atomic-level feature extraction: The obtained atomic sequence features are passed into the molecular pre-trained neural network model to obtain atomic-level features; Step 3: Feature extraction at the amino acid sequence level: The obtained anticancer peptide dataset is sequence data. Feature extraction is performed through sequence feature extraction and secondary structure feature extraction. The obtained sequence features and secondary structure features are spliced ​​to obtain the final amino acid sequence features. Step 4: Extract multi-sequence similarity relationship features: Divide the subsequences, construct a hypergraph, and use the hypergraph attention feature extraction model to learn multi-sequence similarity relationship features; Step 5: Multi-level feature fusion extraction: The atomic layer features, amino acid sequence layer features, and multi-sequence relationship layer features obtained separately are spliced ​​together, and then passed into the extractor composed of a multi-layer linear network for final fusion extraction, and then the final category probability is output.

2. The anticancer peptide identification method based on attention mechanism and multi-granularity hierarchical features according to claim 1, characterized in that: The atomic sequence features in step 2 are obtained as follows: After obtaining the anticancer peptide dataset, the amino acid sequence was parsed using the online server Phyre2, and each amino acid sequence was decomposed into an atomic sequence with atomic groups as the smallest granularity.

3. The anticancer peptide identification method based on attention mechanism and multi-granularity hierarchical features according to claim 1, characterized in that: The molecular pre-training neural network model in step 2 uses a variant of the encoder, VQ-VAE, as a context-aware tagger to encode atoms into meaningful discrete values, which is used to expand the atomic vocabulary and reduce the quantitative differences between atoms. By expanding the "vocabulary" of atoms, a new node-level pre-training task is proposed, and masked discrete values ​​are predicted through pre-training graph neural networks.

4. The anticancer peptide identification method based on attention mechanism and multi-granularity hierarchical features according to claim 1, characterized in that: The specific steps of step 3 sequence feature extraction include: The amino acid sequence is encoded using different sequence encoding methods, and the sequence feature vector is obtained by combining different encoding methods. The sequence features are then extracted through multi-layer convolution.

5. The anticancer peptide identification method based on attention mechanism and multi-granularity hierarchical features according to claim 1, characterized in that: The specific steps of extracting the secondary structure features in step 3 include: The secondary structure sequence of amino acids is obtained through the ChouFasman algorithm and then passed into the multi-layer self-attention mechanism for secondary structure feature extraction.

6. The anticancer peptide identification method based on attention mechanism and multi-granularity hierarchical features according to claim 1, characterized in that: The specific steps of constructing the hypergraph in step 4 include: First, the sequence is decomposed into a set of subsequences as the main feature representation of the sequence. Then, this set of subsequences is represented as nodes of a hypergraph, and each sequence containing a set of subsequences is a hyperedge; each hyperedge is connected to other hyperedges as subsequences through some shared nodes.

7. The anticancer peptide identification method based on attention mechanism and multi-granularity hierarchical features according to claim 1, characterized in that: The hypergraph attention feature extraction model in step 4 includes a two-level attention network optimized using class labels, and the two-level attention networks are an attention aggregation layer for hyperedge features and an attention aggregation layer for node features.

8. A system for identifying anticancer peptides based on an attention mechanism and multi-granularity hierarchical features for implementing the method for identifying anticancer peptides based on an attention mechanism and multi-granularity hierarchical features according to any one of claims 1 to 7, characterized in that: The anticancer peptide recognition system based on attention mechanism and multi-granularity hierarchical features includes: Data acquisition module, used to obtain anticancer peptide datasets; Atomic-level feature extraction module, used to pass the obtained atomic sequence features into the pre-trained model to obtain atomic-level features; The amino acid sequence feature extraction module is used to extract sequence features and secondary structure features, and to combine the obtained sequence features and secondary structure features to obtain the final amino acid sequence features; The multi-sequence similarity relationship feature extraction module is used to divide subsequences, construct hypergraphs, and use the hypergraph attention feature extraction model to learn multi-sequence similarity relationship features; The multi-level feature fusion extraction module is used to splice the atomic layer features, amino acid sequence layer features and multi-sequence relationship layer features obtained separately, and then perform the final fusion extraction to output the final category probability.

9. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the anti-cancer peptide identification method based on the attention mechanism and multi-granularity hierarchical features as described in any one of claims 1 to 7.

Images