An Anti-CRISPR protein recognition method based on graph neural network

Abstract

The present invention relates to the field of gene editing and the field of artificial intelligence, and discloses an Anti-CRISPR protein recognition method based on a graph neural network. The method models the similarity relationship between proteins by constructing a heterogeneous network, and uses a graph neural network to learn protein features, thereby realizing the recognition of Anti-CRISPR proteins. This method improves the recognition performance by integrating multiple protein features, including composition features, evolutionary features, repetition and distribution features, sequence correlation features, structure-related features, and embedded representations generated by the ESM-2 protein language model, and combining a graph attention mechanism (GAT) and a spatial domain-based graph neural network (GraphSAGE). The present invention shows excellent accuracy and robustness on multiple benchmark tests, especially higher recognition accuracy on independent test data sets, and can effectively distinguish between positive and negative samples. This model has important application value in the study of Anti-CRISPR proteins.

Description

Technical Field

[0001] The present invention relates to the fields of gene editing and artificial intelligence, and relates to but is not limited to an Anti-CRISPR protein recognition method based on a graph neural network. Background Art

[0002] As a powerful gene editing tool, the CRISPR-Cas system has shown great application potential in many fields such as biomedicine, agriculture, and synthetic biology. However, the system faces a serious problem in practical applications: off-target effects. Off-target effects refer to the fact that when the CRISPR-Cas system performs gene editing, it may mistakenly cut non-target DNA sequences, leading to unexpected gene mutations or cell damage, thereby causing safety issues. This problem greatly limits the further development and widespread application of the CRISPR-Cas system. In order to solve this problem, researchers have discovered Anti-CRISPR protein, which, as a natural inhibitor of the CRISPR system, can effectively inhibit the activity of the Cas system, thereby improving the accuracy of gene editing. Therefore, Anti-CRISPR protein has extremely high application value in the field of gene editing.

[0003] At present, there are two main methods for identifying Anti-CRISPR proteins: traditional bioinformatics methods and machine learning-based methods. Traditional bioinformatics methods, such as "guilt by association" and "self-targeting analysis", rely on prior biological knowledge to identify Anti-CRISPR proteins. Although this method has a certain degree of accuracy, it is limited by the acquisition and update speed of prior knowledge, and it is difficult to adapt to the needs of large-scale, high-throughput Anti-CRISPR protein identification. Another type of method is based on machine learning. This type of method realizes the automatic identification of Anti-CRISPR proteins by feature extraction and modeling of protein sequences or structures. However, current machine learning methods mainly focus on the modeling of individual protein features and ignore the association between proteins. This method of considering individual proteins in isolation may reduce the accuracy of identifying Anti-CRISPR proteins, because the interaction and association between proteins are crucial in organisms.

[0004] However, most of the existing Anti-CRISPR protein recognition methods still rely on protein sequence information, ignoring the multidimensional characteristics of structure and evolution. Due to the extremely low similarity of Anti-CRISPR protein sequences, traditional sequence alignment methods often have difficulty in effectively extracting features during the recognition process, resulting in low recognition accuracy. Therefore, there is an urgent need for a new recognition method that can integrate the multidimensional characteristics of sequence, structure, and evolution, and has the ability to dynamically adapt to data changes, in order to build a more accurate and reliable Anti-CRISPR protein recognition model. Summary of the invention

[0005] In view of this, an embodiment of the present invention provides an Anti-CRISPR protein recognition method based on a graph neural network, aiming to make full use of the association between proteins and improve the recognition accuracy of Anti-CRISPR proteins.

[0006] The technical solutions of the embodiments of the present invention are as follows:

[0007] An embodiment of the present invention provides an Anti-CRISPR protein recognition method based on a graph neural network, the method comprising:

[0008] Construct a sample data set; wherein the sample data set includes Anti-CRISPR protein sequences and non-Anti-CRISPR protein sequences; extract the multidimensional protein features corresponding to the sample data set, including composition features, evolutionary features, repetition and distribution features, sequence correlation features, structure-related features, and embedded representations generated by the ESM-2 protein language model; construct a protein comprehensive heterogeneous network by combining the multidimensional protein features and the similarity network construction method; wherein the similarity network construction method is used to capture local similarities between protein sequences and structural similarities between proteins; input the protein comprehensive heterogeneous network into the constructed graph neural network model for training; wherein the graph neural network model applies the graph attention mechanism GAT and the spatial domain-based graph neural network GraphSAGE for node feature aggregation and learning; and perform node-by-node recognition of the protein sequence to be identified through the trained graph neural network model.

[0009] In some embodiments, the constructing of the sample data set includes: collecting experimentally verified and literature-mentioned Anti-CRISPR protein sequences from multiple protein databases as a positive sample set; selecting samples from the PaCRISPR negative class data set as a negative sample set; using the CD-HIT tool to remove samples in the positive sample set whose sequence similarity is greater than a first threshold; and combining the positive sample set after de-redundancy processing with the negative sample set to obtain the sample data set.

[0010] In some embodiments, the method of constructing a protein comprehensive heterogeneous network by combining the multidimensional protein features and the similarity network construction method includes: using the Smith-Waterman algorithm to align the protein sequences in the sample data set to obtain a sequence similarity matrix; using the TM-align algorithm to align the protein structures of all sample sequences in the sample data set, calculating the protein structure similarity, and obtaining a structural similarity matrix; fusing the sequence similarity matrix and the structural similarity matrix through a similarity network fusion method to generate a comprehensive similarity matrix; processing the comprehensive similarity matrix based on the K nearest neighbor algorithm, selecting the most similar neighbors of each protein node, and obtaining the protein comprehensive heterogeneous network.

[0011] In some embodiments, during the protein structure alignment process, the method further comprises: retrieving a PDB structure file of each sample sequence in the sample data set from the RCSB database; and performing structure recognition using a protein structure recognition model ESMFold for sample sequences with unknown structures to generate the PDB structure file.

[0012] In some embodiments, the graph neural network model includes a GAT layer, a GraphSAGE layer and a linear layer, wherein the GAT layer is used to calculate the attention weights between each protein node, and adaptively adjust the connection strength between neighboring nodes through a learnable weight matrix; the GraphSAGE layer has four layers in total, and aggregates information from the neighborhood of each protein node through multiple iterations to achieve comprehensive learning of global and local information between nodes; the linear layer is used to convert the aggregated features into the probability of Anti-CRISPR protein recognition.

[0013] In some embodiments, the number of GAT layers is 1, the number of attention heads is set to 16, and the dimension of the hidden layer is 1024; the number of GraphSAGE layers is 4, and the dimensions of the hidden layers are set to 512, 1024, 512, and 256, respectively.

[0014] In some embodiments, the protein sequence to be identified is performed node by node through the trained graph neural network model, including: adding the protein sequence to be identified as the target protein node one by one to the protein comprehensive heterogeneous network currently constructed, and then gradually identifying it based on the similarity relationship between the target protein node and the existing protein nodes.

[0015] In some embodiments, the method further includes: using a binary cross entropy loss function in combination with an L2 regularization term for training, and using a cosine annealing learning rate scheduling strategy to optimize the learning process of the model, and using an Adam optimizer for parameter updating.

[0016] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0017] In an embodiment of the present invention, by constructing all protein sequences in the sample data set into a protein comprehensive heterogeneous network, and utilizing the graph attention mechanism (GAT) and the spatial domain-based graph neural network (GraphSAGE), effective modeling of protein association relationships can be achieved, and multiple protein features can be effectively integrated and the similarities between proteins can be dynamically learned, thereby improving the accuracy of Anti-CRISPR protein recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work, among which:

[0019] Figure 1 A schematic diagram of the process of the Anti-CRISPR protein recognition method based on graph neural network provided in an embodiment of the present invention;

[0020] Figure 2 A schematic diagram of the construction process of a protein comprehensive heterogeneous network provided in an embodiment of the present invention;

[0021] Figure 3 A structural framework diagram of the graph neural network model provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0023] In the following description, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0024] It should be pointed out that the terms "first\second\third" involved in the embodiments of the present invention are only used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present invention described here can be implemented in an order other than that illustrated or described here.

[0025] Those skilled in the art will appreciate that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art in the field to which the embodiments of the present invention belong. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with those in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless specifically defined as here.

[0026] Figure 1 A schematic diagram of a flow chart of an Anti-CRISPR protein recognition method based on a graph neural network provided in an embodiment of the present invention, such as Figure 1 As shown, the method comprises at least the following steps:

[0027] Step S110, constructing a sample data set.

[0028] Here, the sample data set includes Anti-CRISPR protein sequences and non-Anti-CRISPR protein sequences. Among them, Anti-CRISPR protein sequences are positive samples, which can be collected from multiple protein databases; non-Anti-CRISPR protein sequences are negative samples, which can be collected from PaCRISPR's negative data set. In the implementation, the sample data set is randomly divided into a training set (80%) and a validation set (20%), which are used for the training and validation of the subsequent graph neural network model, respectively.

[0029] Step S120, extracting multidimensional protein features corresponding to the sample data set, including composition features, evolutionary features, repetition and distribution features, sequence correlation features, structure correlation features, and embedded representations generated by the ESM-2 protein language model.

[0030] Here, composition features include AAC (amino acid composition), DPC (dipeptide composition), CTDC (composition, transition, distribution comprehensive descriptor (composition)), PCP (physicochemical property component); evolutionary features include derived features calculated by PSSM (position-specific scoring matrix): AADP-PSSM, PSSM-AC, MEDP, PSSM-composition, RPSSM; repetition and distribution features include RRI (repeat residue information), SER (Shannon entropy at the residue level), CTDD (composition, transition, distribution comprehensive descriptor (distribution)); sequence-related features include CTDT (composition, transition, distribution comprehensive descriptor (transition)), SOCN (sequence coupling number), PAAC (pseudo-amino acid composition); structure-related features include B-factor (temperature factor) and SSC (secondary structure composition).

[0031] It is worth noting that the present invention uses the embedding representation generated by the ESM-2 protein language model as the initial feature vector. The parameters of the ESM-2 model are: 33 transformer layers, 20 attention heads, 1280 hidden layers, and a total of 650 million parameters. As part of the input features of the present invention, it can enhance the feature learning ability of the subsequent graph neural network model and reduce the problem of data scarcity.

[0032] Step S130, constructing a protein comprehensive heterogeneous network by combining the multi-dimensional protein features and the similarity network construction method.

[0033] Here, the similarity network construction method is used to capture the local similarity between protein sequences and the structural similarity between proteins. Thus, protein nodes are constructed based on multidimensional protein features, and the edge relationship between protein nodes is established in combination with the similarity network construction method to form a protein comprehensive heterogeneous network. This protein comprehensive heterogeneous network ensures the density and effectiveness of the data, providing strong support for subsequent functional identification.

[0034] Step S140, inputting the protein comprehensive heterogeneous network into the constructed graph neural network model for training.

[0035] Here, the graph neural network model uses the graph attention mechanism GAT and the spatial domain-based graph neural network GraphSAGE to aggregate and learn node features. The GAT layer adaptively adjusts the connection strength between neighbor nodes through a learnable weight matrix, allowing the model to flexibly capture the similarities between proteins. The GraphSAGE layer achieves comprehensive learning of global and local information between nodes by sampling neighbors and aggregating their features.

[0036] Step S150, performing node-by-node identification on the protein sequence to be identified using the trained graph neural network model.

[0037] Here, in the recognition process, in order to be closer to the actual application environment, the protein sequences to be identified are added one by one as target protein nodes to the existing protein comprehensive heterogeneous network, and the network structure is dynamically updated to gradually perform the recognition. This method improves the flexibility and accuracy of the model in processing protein data.

[0038] In an embodiment of the present invention, by constructing all protein sequences in the sample data set into a protein comprehensive heterogeneous network, and utilizing the graph attention mechanism (GAT) and the spatial domain-based graph neural network (GraphSAGE), effective modeling of protein association relationships can be achieved, and multiple protein features can be effectively integrated and the similarities between proteins can be dynamically learned, thereby improving the accuracy of Anti-CRISPR protein recognition.

[0039] In some embodiments, constructing a sample data set includes: collecting experimentally verified and literature-mentioned Anti-CRISPR protein sequences from multiple protein databases as a positive sample set; selecting samples from a negative class data set of PaCRISPR as a negative sample set; using a CD-HIT tool to remove samples in the positive sample set whose sequence similarity is greater than a first threshold; and combining the positive sample set after de-redundancy processing with the negative sample set to obtain the sample data set.

[0040] Here, the multiple protein databases are multiple authoritative data sets (Anti-CRISPRdb, PaCRISPR, AcrHub, and the unified Anti-CRISPR protein naming resource (Bondy-Denomy et al., 2018)).

[0041] The collected Anti-CRISPR protein sequences were de-redundanted using the CD-HIT tool. Preferably, the first threshold was 90% to ensure the uniqueness and representativeness of the data set. After processing, 1004 Anti-CRISPR protein sequences were obtained as positive samples.

[0042] 1004 non-Anti-CRISPR protein sequences were selected from the PaCRISPR negative dataset as negative samples. These sequences were from bacteriophage or bacterial MGE and from bacterial genera known to contain Anti-CRISPR proteins. The sequence lengths ranged from 50 to 350 residues, and the sequence similarity between them was less than 40%.

[0043] The collected 2008 sequences were randomly divided into a training set (80%) and a validation set (20%), with a total of 1606 sequences in the training set and 402 sequences in the validation set for subsequent model training and validation.

[0044] In some embodiments, the method of constructing a protein comprehensive heterogeneous network by combining the multidimensional protein features and the similarity network construction method comprises the following steps: using the Smith-Waterman algorithm to align the protein sequences in the sample data set to obtain a sequence similarity matrix; using the TM-align algorithm to align the protein structures of all sample sequences in the sample data set, calculating the protein structure similarity, and obtaining a structural similarity matrix; fusing the sequence similarity matrix and the structural similarity matrix through a similarity network fusion method to generate a comprehensive similarity matrix; processing the comprehensive similarity matrix based on a nearest neighbor algorithm, selecting the most similar neighbor of each protein node, and obtaining the protein comprehensive heterogeneous network.

[0045] Here, the construction process of protein comprehensive heterogeneous network is as follows Figure 2 As shown, on the one hand, the local similarity of protein sequences is calculated based on the Smith-Waterman algorithm, the proteins are aligned, the BLOSUM62 scoring matrix is ​​used as the scoring standard for amino acid substitution, and a gap penalty is introduced to penalize insertions or deletions in the sequence alignment. Preferably, the scores are standardized to obtain normalized sequence similarity.

[0046] On the other hand, TM-align is used to calculate the protein structure similarity TM-score, which is obtained by comparing the distance between each pair of aligned residues and adjusting the score according to the protein size. , the calculation formula is as follows:

[0047] ;

[0048] In the formula, is the length of the original protein structure, i.e. the number of residues; is the residue length aligned to the template; It is Align the distances between residue pairs; is a constant used for normalization to eliminate the effect of protein size. The structural similarity matrix can be obtained through TM-score.

[0049] The two similarity matrices were fused by the Similarity Network Fusion (SNF) technique, with 20 iterations and 8 nearest neighbors. Finally, the K nearest neighbor algorithm was used to select the K most similar neighbors of each protein node to construct the network edges. Preferably, K = 400 is the optimal value to ensure the balance between network connectivity and computational complexity.

[0050] In some embodiments, during the protein structure alignment process, the method further comprises: retrieving a PDB structure file of each sample sequence in the sample data set from the RCSB database; and performing structure recognition using a protein structure recognition model ESMFold for sample sequences with unknown structures to generate the PDB structure file.

[0051] Here, the PDB structure files of all protein sequences can be used for protein structure alignment, and structural related features such as B-Factor features and SSC features can be extracted from them. For sample sequences with unknown structures, PDB structure files are generated through the ESMFold model to ensure the integrity of the data set.

[0052] In some embodiments, the graph neural network model includes a GAT layer, a GraphSAGE layer and a linear layer, wherein the GAT layer is used to calculate the attention weights between each protein node, and adaptively adjust the connection strength between neighboring nodes through a learnable weight matrix; the GraphSAGE layer has four layers in total, and aggregates information from the neighborhood of each protein node through multiple iterations to achieve comprehensive learning of global and local information between nodes; the linear layer is used to convert the aggregated features into the probability of Anti-CRISPR protein recognition.

[0053] Here, the structural framework of the graph neural network model is as follows Figure 3 As shown in the figure, the graph attention layer is used to dynamically adjust the feature weights to achieve flexible capture of node relationships, and the GraphSAGE layer is combined to extract local and global information, thereby improving the recognition accuracy of the model. The final linear layer maps the feature dimension from 256 to 1 to obtain the output probability of the model. The weight decay parameter of the L2 regularization term is set to 1e -5 .

[0054] GAT layer output node The eigenvector of for:

[0055] ;

[0056] In the formula, Representation Node With neighbor nodes The attention coefficient between is the learnable weight matrix, represents the set of neighbor nodes of node i, represents the input features of node j, Represents the activation function.

[0057] The node feature aggregation formula of the GraphSAGE layer is:

[0058] ;

[0059] In the formula, , Respectively represent The node features of the layer and the l+1th layer, W is the learnable weight matrix, represents the set of neighbor nodes of node i, denote the degrees of nodes i and j respectively.

[0060] The GAT layer adaptively adjusts the connection strength between neighbor nodes through a learnable weight matrix, allowing the model to flexibly capture the similarities between proteins. The GraphSAGE layer achieves comprehensive learning of global and local information between nodes by sampling neighbors and aggregating their features.

[0061] In some embodiments, the number of GAT layers is 1, the number of attention heads is set to 16, and the dimension of the hidden layer is 1024; the number of GraphSAGE layers is 4, and the dimensions of the hidden layers are set to 512, 1024, 512, and 256, respectively.

[0062] Here, the setting of the above parameters ensures the complexity of the model while maintaining the efficiency of the model. Experimental results show that the architecture of 4-layer GraphSAGE combined with GAT layer can capture multi-scale network features to the greatest extent and provide the best feature expression capability.

[0063] In some embodiments, the protein sequence to be identified is performed node by node through the trained graph neural network model, including: adding the protein sequence to be identified as the target protein node one by one to the protein comprehensive heterogeneous network currently constructed, and then gradually identifying it based on the similarity relationship between the target protein node and the existing protein nodes.

[0064] In this way, new Anti-CRISPR protein sequences can be flexibly identified without destroying the existing network structure. This method effectively improves the model's ability to identify unknown nodes and can cope with the challenges of data scarcity and changes in network structure.

[0065] In some embodiments, the method further includes: using a binary cross entropy loss function in combination with an L2 regularization term for training, and using a cosine annealing learning rate scheduling strategy to optimize the learning process of the model, and using an Adam optimizer for parameter updating.

[0066] Here, the binary cross entropy loss function formula is as follows:

[0067] ;

[0068] in, is the loss function, is the sample size, is the true label, is the model recognition result, is the regularization coefficient; Represents the learnable weights of the last layer, the regularization term By penalizing excessively large weight values, the model is forced to maintain a simpler structure during training, thereby improving its performance on unknown data.

[0069] The binary cross entropy loss function is used for model optimization, combined with The regularization term is used to prevent overfitting and make the model more generalizable on limited data. The training process is adjusted by the cosine annealing learning rate scheduling strategy, so that the graph neural network model constructed by the present invention can stably converge to the optimal state.

[0070] The present invention randomly divides the collected 2008 sequences into a training set (80%) and a validation set (20%), with a total of 1606 sequences in the training set and 402 sequences in the validation set, which are used for model training and validation, respectively. At the same time, newly annotated Anti-CRISPR protein sequences published in NCBI from January to October 2024 are collected as test positive classes, and unused samples in the PaCRISPR negative class dataset are combined as test negative classes, resulting in a total of 43 positive samples and 43 negative samples as independent test sets for subsequent model node-by-node recognition testing.

[0071] The indicators on the validation set are shown in Table 1, where TP stands for the number of true positives, which refers to the number of samples that the model correctly predicts to be positive and are actually positive; TN stands for the number of true negatives, which refers to the number of samples that the model correctly predicts to be negative and are actually negative; FP stands for the number of false positives, which refers to the number of samples that the model incorrectly predicts to be positive but are actually negative; and FN stands for the number of false negatives, which refers to the number of samples that the model incorrectly predicts to be negative but are actually positive. MCC stands for the Matthews Correlation Coefficient. ROC-AUC and PR-AUC are both important indicators for evaluating the performance of binary classification models, and they represent different evaluation angles. The ROC (Receiver Operating Characteristic) curve is a curve that reflects the relationship between the true positive rate (TPR) and the false positive rate (FPR) of the classification model. AUC (Area Under Curve) is the area under the ROC curve. This area value can be used as a numerical indicator to measure the quality of the classification model. Among them, the true positive rate (TPR) = TP / (TP + FN), which indicates the proportion of all positive classes predicted as positive; the false positive rate (FPR) = FP / (TN + FP), which indicates the proportion of all negative classes incorrectly predicted as positive. By selecting different thresholds in the range of 0 to 1, calculating the corresponding TPR and FPR, and then connecting these points, the ROC curve is formed. AUC is the area under the curve. The PR (Precision-Recall) curve is a curve that reflects the relationship between the precision (Precision) and recall (Recall) of the classification model. PR-AUC is the area under the PR curve. It can be seen that the accuracy (Accuracy) of the present invention reaches 0.9577, the F1 value reaches 0.9572, and the geometric evaluation number (G-mean) reaches 0.9576.

[0072] The indicators on the independent test set are shown in Table 2. The accuracy of the present invention reaches 0.8256, the F1 value is 0.7887, and the PR-AUC reaches 0.8806. On the independent test set, the recognition accuracy of the present invention remains at a high level, indicating its generalization ability and robustness.

[0073] Table 1: Performance comparison of the present invention and different methods on the validation set

[0074] ;

[0075] Table 2: Performance comparison of the present invention and different methods on independent test sets

[0076] ;

[0077] As can be seen from Tables 1 and 2, the present invention has significant advantages over existing methods (PaCRISPR, AcrPred, DeepAcr, AcRanker, HMM) in terms of accuracy, F1 value, and PR-AUC. Through the node-by-node network update method, the present invention can effectively improve the accuracy of recognition, especially when processing protein data with high similarity, showing better adaptability and stability.

[0078] The present invention innovatively integrates compositional features, evolutionary features, repetitive and distribution features, sequence correlation features, structural correlation features, and embedded representations generated by the ESM-2 protein language model, thereby improving the accuracy of feature representation. At the same time, the present invention combines the graph attention mechanism (GAT) and the spatial domain-based graph neural network (GraphSAGE), which can adaptively meet the feature requirements of different protein nodes, achieve efficient feature aggregation, and greatly improve recognition accuracy. In addition, the present invention allows new nodes to be added to the network one by one and recognized in real time, so that the model has stronger flexibility and generalization ability in practical applications, and effectively responds to the challenges brought by data scarcity and network changes. The present invention is significantly superior to existing methods in various evaluation indicators, and performs well in accuracy, recall rate and generalization ability. This result not only verifies the effectiveness of graph neural networks in protein recognition, but also demonstrates its application potential in Anti-CRISPR protein recognition.

[0079] It should be understood that "one embodiment" or "an embodiment" mentioned throughout the specification means that specific features, structures or characteristics related to the embodiment are included in at least one embodiment of the present invention. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner. It should be understood that in various embodiments of the present invention, the size of the serial number of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention. The serial numbers of the above-mentioned embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments.

[0080] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0081] In the several embodiments provided by the present invention, it should be understood that the disclosed methods can be implemented in other ways. The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments. The features disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0082] The above is only an embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. An Anti-CRISPR protein recognition method based on graph neural network, characterized in that: include: Constructing a sample data set; wherein the sample data set includes Anti-CRISPR protein sequences and non-Anti-CRISPR protein sequences; Extracting multidimensional protein features corresponding to the sample data set, including composition features, evolutionary features, repetition and distribution features, sequence correlation features, structure correlation features, and embedded representations generated by the ESM-2 protein language model; Combining the multidimensional protein features and the similarity network construction method to construct a protein comprehensive heterogeneous network; wherein the similarity network construction method is used to capture the local similarity between protein sequences and the protein structure similarity; Inputting the protein comprehensive heterogeneous network into the constructed graph neural network model for training; wherein the graph neural network model uses the graph attention mechanism GAT and the spatial domain-based graph neural network GraphSAGE to aggregate and learn node features; The protein sequence to be identified is identified node by node through the trained graph neural network model; The method of constructing a protein comprehensive heterogeneous network by combining the multidimensional protein features and the similarity network construction method comprises: The Smith-Waterman algorithm is used to compare the protein sequences in the sample data set to obtain a sequence similarity matrix; Using the TM-align algorithm to perform protein structure alignment on all sample sequences in the sample data set, calculate the protein structure similarity, and obtain a structure similarity matrix; The sequence similarity matrix and the structural similarity matrix are fused by a similarity network fusion method to generate a comprehensive similarity matrix; Processing the comprehensive similarity matrix based on the K-nearest neighbor algorithm, selecting the most similar neighbors of each protein node, and obtaining the protein comprehensive heterogeneous network; The graph neural network model includes a GAT layer, a GraphSAGE layer and a linear layer, wherein the GAT layer is used to calculate the attention weights between each protein node, and adaptively adjust the connection strength between neighboring nodes through a learnable weight matrix; the GraphSAGE layer has four layers in total, and aggregates information from the neighborhood of each protein node through multiple iterations to achieve comprehensive learning of global and local information between nodes; the linear layer is used to convert the aggregated features into the probability of Anti-CRISPR protein recognition.

2. The method according to claim 1, characterized in that The constructing of the sample data set includes: Anti-CRISPR protein sequences that have been experimentally verified and mentioned in the literature were collected from multiple protein databases as positive sample sets; Select samples from the negative class dataset of PaCRISPR as the negative sample set; Using the CD-HIT tool to remove samples in the positive sample set whose sequence similarity is greater than a first threshold; The sample data set is obtained by combining the positive sample set after redundancy removal processing and the negative sample set.

3. The method according to claim 1, characterized in that During the protein structure alignment process, the method further comprises: Retrieve the PDB structure file of each sample sequence in the sample data set from the RCSB database; For sample sequences with unknown structures, the protein structure recognition model ESMFold is used to perform structure recognition to generate the PDB structure file.

4. The method according to claim 1, characterized in that: The number of GAT layers is 1, the number of attention heads is set to 16, and the dimension of the hidden layer is 1024; the number of GraphSAGE layers is 4, and the dimensions of the hidden layers are set to 512, 1024, 512, and 256, respectively.

5. The method according to any one of claims 1 to 4, characterized in that: The method of performing node-by-node identification of the protein sequence to be identified by using the trained graph neural network model includes: The protein sequences to be identified are added one by one as target protein nodes to the currently constructed protein comprehensive heterogeneous network, and then gradually identified based on the similarity relationship between the target protein node and the existing protein nodes.

6. The method according to any one of claims 1 to 4, characterized in that: The method further comprises: The binary cross entropy loss function combined with the L2 regularization term is used for training, and the cosine annealing learning rate scheduling strategy is used to optimize the learning process of the model, and the Adam optimizer is used for parameter update.

Images