Exosome-related piRNA identification method based on graph neural network

Abstract

The present invention discloses a method for identifying exosome-related piRNA based on graph neural network. The identification method specifically comprises: step S1, constructing an exosome piRNA identification benchmark data set; step S2, using a two-step positive-unlearning method, selecting piRNA sequences consistent with the number of positive sets from an unlabeled data set as reliable negative set samples; step S3, based on sequence information, performing graph construction; step S4, constructing an embedding space from a large amount of piRNA sequence data in a pre-trained model of k-mer fragment sequences so that biologically similar k-mer sequences are closer in the space; step S5, obtaining a word vector representation of the k-mer sequence according to step 4, and using an improved graph convolutional neural network as a classification model to construct a classification model; step S6, determining the sequence data for the classification score obtained in step 5, and taking the classification with a higher probability as the final result of whether the current piRNA is related to exosomes. The present invention can be used to identify exosome-related piRNA.

Description

Technical Field

[0001] The present invention belongs to the technical field of identifying exosome-related piRNA, and in particular relates to an effective method for identifying exosome-related piRNA based on a graph neural network. Background Art

[0002] In recent years, studies have shown that epigenetics, especially non-coding RNA, plays an important role in the occurrence and development of diseases, and has gradually become a hot topic in international life science research. With the increasing maturity of research on microRNA (microRNA, miRNA) and long non-coding RNA (long non-coding RNA, lncRNA) in complex disease-related non-coding RNA, PIWI-interacting RNA (piRNA) has become the forefront of the field.

[0003] piRNA was discovered in 2006. It is a type of non-coding RNA that interacts with PIWI family proteins. This discovery was named one of the top ten scientific advances of the year by Science magazine. Studies have found that piRNA has important biological functions, such as maintaining genome integrity, epigenetic regulation, and heterochromatin formation. Early piRNA research focused on gonadal tissues. Recent studies have found that abnormally expressed piRNAs also exist in a variety of non-gonadal tissues such as the central nervous system (brain), colon, heart, kidney, liver, lung, small intestine, large intestine, spleen, and stomach. They are also associated with a variety of cancers and are expected to become new diagnostic or prognostic biomarkers and drug targets.

[0004] Recently, a number of breakthrough new discoveries have shown that exosome-related piRNAs are potential disease markers and therapeutic targets with good clinical application prospects. For example, the Shanghai Institute of Immunochemistry analyzed the exosomes of patients with cholangiocarcinoma and gallbladder cancer and found that piRNAs in plasma had significant changes.

[0005] Identifying exosome-associated piRNAs is a fundamental and challenging research problem. There are two types of existing identification methods, namely biological experimental methods and computational methods. For example, the biological experimental method of Valencia-Burton et al. using fluorescent protein complementation to monitor RNA localization in prokaryotic cells can obtain the subcellular location of piRNAs more accurately and reliably, but it requires professional researchers, corresponding experimental reagents and equipment, etc., which is very expensive and time-consuming. With the rapid development of genome sequencing technology, a large amount of piRNA data has been accumulated. Therefore, it is possible to use computational methods to predict the study of exosome-associated piRNAs. Summary of the invention

[0006] In order to overcome the shortcomings of the above-mentioned existing identification technologies, the purpose of the present invention is to provide a method for identifying exosome-related piRNAs based on graph neural networks, which aims to adopt strict filtering conditions and rigorous learning strategies to construct a benchmark data set, so that the results output by subsequent classification models are more reliable and credible, and the system error is reduced; k-mer fragments and k-mer fragment sequences containing wildcards are used to construct graphs; a graph attention model is designed to identify exosome-related piRNAs and explore potential disease markers and therapeutic targets.

[0007] This application uses an algorithm suitable for identifying exosome-associated piRNA. piRNA is a class of non-coding RNA that interacts with PIWI family proteins. Exosomes are extracellular vesicles secreted by cells, the main components of which are phospholipid bilayers and related contents. Exosomes can be extracted from a variety of body fluids, such as human blood, urine, semen, and saliva. Studies have shown that exosome-associated piRNA is closely related to the occurrence and development of diseases and is a potential disease biomarker and therapeutic target.

[0008] To achieve the above object, the technical solution adopted by the present invention is: a method for identifying exosome-related piRNA based on a graph neural network, comprising the following steps:

[0009] Step S1, constructing a benchmark dataset for exosomal piRNA identification;

[0010] Step S2, using a two-step positive-unlabeled learning method, select piRNA sequences with the same number as the positive set from the unlabeled dataset as reliable negative set samples;

[0011] Step S3, constructing a graph based on the sequence information;

[0012] Step S4, by learning the biological significance knowledge from a large amount of piRNA sequence data in a pre-trained model of K-mer fragment sequences in an unsupervised manner, thereby constructing an embedding space so that biologically similar k-mer sequences are closer in the space;

[0013] Step S5, obtaining the word vector representation of the k-mer sequence according to step 4, and using the improved graph convolutional neural network as a classification model to construct a classification model;

[0014] Step S6, determining the sequence data for the classification scores obtained in step 5, and taking the classification with a higher probability as the final result of whether the current piRNA is related to exosomes.

[0015] Further, the step S1 specifically includes:

[0016] S11: 20,867 piRNA subcellular localization data were downloaded from the RNALocate 2.0 database, of which 10,438 exosome-related piRNA records were retained;

[0017] S12: According to the piRNA name obtained in step S11, the corresponding sequence data was collected from the piRBase database;

[0018] S13: The human piRNA gold standard dataset, containing 77,242 piRNA sequences, was downloaded from the piRBase database for constructing the unlabeled dataset.

[0019] S14: Based on the piRNA sequence data obtained in step S13, remove the positive set data according to the sequence alignment results;

[0020] S15: Based on the piRNA sequence data obtained in step S14, the CD-HIT tool is used to reduce sequence redundancy, and the sequence identity parameter is still set to 0.8;

[0021] S16: After steps S14 and S15, 21,979 piRNA data with low redundancy are obtained as an unlabeled dataset.

[0022] Furthermore, in step S12, the existing exosome-related piRNA sequence data are clustered using the CD-HIT tool to reduce sequence redundancy, and the sequence identity parameter is set to 0.8.

[0023] Further, the step S2 specifically includes:

[0024] S21: Randomly select samples from the unlabeled dataset, build a training set with the existing positive set samples, and use the training set to train the random forest classifier;

[0025] S22: Use the classifier to predict the remaining unlabeled samples and repeat multiple times;

[0026] S23: To reduce the false negatives of predictions, the average values ​​of the prediction scores were sorted in reverse order, and the middle 8321 piRNA records were selected as reliable negative set samples.

[0027] Further, the step S3 specifically includes:

[0028] S31: Extract k-mer sequences from the piRNA sequences of the benchmark dataset using a sliding window;

[0029] S32: According to all k-mer sequence fragments obtained in step S31, the middle bases are replaced with wildcards;

[0030] S33: A graph representation is constructed using k-mer sequence fragments and k-mer sequence fragments with wildcards, wherein the points of the graph correspond to the k-mer sequence fragments and the k-mer sequence fragments with wildcards, the edges correspond to the links of the k-mer sequence fragments, and the k-mer sequence fragments with wildcards are connected to the k-mer sequence fragments whose intermediate bases may be replaced.

[0031] Further, the step S4 specifically includes:

[0032] S41: In order to obtain a distributed representation of k-mer sequences, k-mer sequences were first extracted from all piRNA sequences collected from the piRBase database;

[0033] S42: Model the relationship between the context sequence and the target k-mer sequence through the Skip-gram model;

[0034] S43: Referring to Word2Vec, we maximize the following function through such unsupervised large-scale corpus learning to obtain the distributed representation of k-mer sequences:

[0035]

[0036]

[0037] Where N is the length of the training sequence, 2c is the window size, and w i is the center of the window, W is the number of k-mers, v w and v' w are the input and output representations of a particular k-mer sequence w. The probability p(w i+j |w i ) uses the softmax function form.

[0038] Further, the step S5 specifically includes:

[0039] S51: Obtain the word vector representation of the k-mer sequence according to step 4, and the word vector of the k-mer sequence containing the wildcard position is represented by the sum of the related k-mers;

[0040] S52: Use the improved graph convolutional neural network as the classification model. The model is divided into a graph convolution module, a convolution module, and a fully connected network module.

[0041] S53: Use two graph convolutional layers, both of which apply the GraphSAGE method to sample neighbor nodes and perform average aggregation based on the node’s own word vectors. The calculation is as follows:

[0042] h' v =σ(Bh v +W·mean({hu |u∈N(v)})) Formula 5

[0043] where h v is the feature of vertex v, N(v) is the set of neighboring points of vertex v in the graph, σ is a nonlinear activation function, and W and B are matrices to be learned;

[0044] S54: The convolution module uses one-dimensional convolution to reduce the dimension of the output of the graph convolution layer;

[0045] S55: A fully connected layer is used at the end of the network to integrate the features after dimensionality reduction, and softmax logistic regression can be used for classification.

[0046] Step S6, judging the sequence data for the classification scores obtained in step 5, and taking the classification with a higher probability as the final result, that is, whether the current piRNA is related to exosomes.

[0047] The beneficial effects of the present invention are:

[0048] Since the present invention uses the latest and most authoritative database as a reliable data source, adopts strict data screening standards to ensure data accuracy, and uses a graph neural network model for the first time to identify exosome-related piRNA, it has the advantages of novel solutions and accurate results.

[0049] In the present invention, piRNA-related data are deeply analyzed to determine exosome-related piRNA. Authoritative databases such as RNALocate and piRBase provide a large amount of high-quality piRNA data, which lays the foundation for studying exosome-related piRNA. PiRNA subcellular localization data can be obtained from the RNALocate database. Based on the gold standard data of piRNA, an identification benchmark data set is constructed, and a graph neural network model is applied to identify exosome-related piRNA.

[0050] In addition to considering k-mer sequence fragments, the present invention also considers k-mer sequence fragments with wildcards as nodes to construct graphs. The exosome-related piRNAs identified by the present invention can be used to reveal the functions of piRNAs in the occurrence and development of diseases, and may also provide new biomarkers and targets for diagnosis and treatment.

[0051] The present invention designs a method for identifying exosome-related piRNA based on a graph neural network model. The method and results are very useful for researchers exploring the function of piRNA through next-generation sequencing applications. The methods and resources provided by the present invention will be helpful for studying the function of piRNA in various complex diseases.

[0052] These and other aspects of the present application will be more concise and understandable in the following description of the embodiments. It should be understood that the above general description and the following detailed description are only exemplary and explanatory and cannot limit the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or the prior art description. Obviously, the drawings described below are only some embodiments of the present application. Among them:

[0054] Figure 1 It is a general flow chart of the method for identifying exosome-related piRNA based on graph neural network of the present invention;

[0055] Figure 2 It is a flowchart for the construction of the exosome-associated piRNA identification benchmark dataset;

[0056] FIG3(A) is a sequence data of piR-hsa-528 in step 3 of the present invention, taking piR-hsa-528 as an example.

[0057] “ACAAAGGGAGGCAGAACGTAAAGCCCAGAAAG”, a graphical representation of the transformation;

[0058] FIG3(A) illustrates the process of selecting k-mer fragments and k-mer sequence fragments with wildcards;

[0059] Figure 3(B) illustrates the result of graph construction;

[0060] Figure 4 The structure of the classification model in step 5 is illustrated. DETAILED DESCRIPTION

[0061] Below, the present application is further described in conjunction with the accompanying drawings and specific implementation methods. It should be noted that, under the premise of no conflict, the various embodiments or technical features described below can be arbitrarily combined to form a new embodiment.

[0062] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0063] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0064] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present application, rather than to limit the present application. It should also be noted that, for ease of description, only the parts related to the present application, rather than all structures, are shown in the accompanying drawings.

[0065] It should be mentioned before discussing the exemplary embodiments in more detail that some exemplary embodiments are described as processes or methods depicted as flow charts. Although the flow charts describe the steps as sequential processes, many of the steps therein can be implemented in parallel, concurrently or simultaneously. In addition, the order of the steps can be rearranged. The process can be terminated when its operation is completed, but there can also be additional steps not included in the accompanying drawings.

[0066] The purpose of the present invention is to solve the shortcomings in the above background and propose a prediction method for piRNA and mRNA target pairs based on an interactive reasoning network. To achieve the above purpose, the present invention provides the following technical solutions, which are further described in detail below in conjunction with the accompanying drawings and examples.

[0067] This application uses an algorithm suitable for identifying exosome-associated piRNA. piRNA is a class of non-coding RNA that interacts with PIWI family proteins. Exosomes are extracellular vesicles secreted by cells, the main components of which are phospholipid bilayers and related contents. Exosomes can be extracted from a variety of body fluids, such as human blood, urine, semen, and saliva. Studies have shown that exosome-associated piRNA is closely related to the occurrence and development of diseases and is a potential disease biomarker and therapeutic target.

[0068] A method for identifying exosome-related piRNA based on a graph neural network model, the method comprising the following steps:

[0069] Step S1, constructing a benchmark dataset for exosomal piRNA identification;

[0070] The steps for constructing the exosome-associated piRNA identification benchmark dataset are as follows: Figure 1 As shown, Figure 1 The processing flow on the left shows how to construct the positive set. Figure 1 The processing on the right shows the method of constructing the unlabeled sample set, which specifically includes the following:

[0071] S11: 20,867 piRNA subcellular localization data were downloaded from the RNALocate 2.0 (http: / / www.rna-society.org / rnalocate / ) database, of which 10,438 exosome-related piRNA records were retained.

[0072] S12: According to the piRNA names obtained in step S11, the corresponding sequence data were collected from the piRBase database; the existing exosome-related piRNA sequence data were clustered using the CD-HIT (Cluster Database at HighIdentity with Tolerance, https: / / github.com / weizhongli / cdhi t) tool to reduce sequence redundancy, and the sequence identity parameter was set to 0.8;

[0073] After steps S11 and S12, 8321 exosome-related piRNA data with low redundancy were finally obtained as the positive data of the benchmark dataset.

[0074] S13: The human piRNA gold standard dataset, containing 77,242 piRNA sequences, was downloaded from the piRBase database for constructing the unlabeled dataset.

[0075] The piRBase 3.0 database is currently the largest piRNA database, and its URL is http: / / bigdata.ibp.ac.cn / piRBase / . Because piRNA is the largest type of non-coding small RNA and its number is huge, the piRBase database specially selects representative piRNA sequences to construct the piRNA gold standard data to facilitate piRNA-related research.

[0076] S14: Based on the piRNA sequence data obtained in step S13, remove the positive set data according to the sequence alignment results;

[0077] The positive set of data refers to all known exosome-related piRNAs downloaded from the RNALocate 2.0 database in step S11.

[0078] S15: Based on the piRNA sequence data obtained in step S14, the CD-HIT tool is used to reduce sequence redundancy, and the sequence identity parameter is still set to 0.8;

[0079] S16: After steps S14 and S15, 21,979 piRNA data with low redundancy are obtained as an unlabeled dataset.

[0080] Step S2, using a two-step Positive-Unlabeled Learning method, selects piRNA sequences with the same number as the positive set from the unlabeled dataset as reliable negative set samples.

[0081] The Positive-Unlabeled Learning method is a semi-supervised learning algorithm that trains a classifier with only positive data and unlabeled data.

[0082] Step S2 specifically includes:

[0083] S21: Randomly select samples from the unlabeled dataset, build a training set with the existing positive set samples, and use the training set to train the random forest classifier;

[0084] S22: Use the classifier to predict the remaining unlabeled samples and repeat multiple times;

[0085] S23: To reduce the false negatives of predictions, the average values ​​of the prediction scores were sorted in reverse order, and the middle 8321 piRNA records were selected as reliable negative set samples.

[0086] Step S3, constructing a graph based on the sequence information;

[0087] FIG3 is a graphical representation of the conversion process in step S3 of the present invention, taking piR-hsa-528 as an example, based on its sequence data "ACAAAGGGAGGCAGAACGTAAAGCCCAGAAAG". FIG3(A) illustrates steps S31 and S32, i.e., the selection process of k-mer fragments and k-mer sequence fragments with wildcards; FIG3(B) illustrates step S32, i.e., the result of graph construction.

[0088] Step S3 includes:

[0089] S31: For the piRNA sequences of the benchmark dataset, k-mer sequences are extracted using a sliding window.

[0090] Since every three adjacent bases on mRNA can constitute the biological background of a codon, it is believed that RNA fragments composed of three bases may have specific biological functions, so 3 is selected as the fixed window size.

[0091] For a given sequence S=(S 1 ,S 2 ,…,S n ), convert S to k-mer sequence fragments, where S i ∈{A,G,C,T,N}, the sequence is in cDNA form:

[0092]

[0093]

[0094] like Figure 2As shown, taking piR-hsa-528 as an example, this piRNA is a piRNA sequence associated with exosomes, and the specific sequence data is "ACAAAGGGAGGCAGAACGTAAAGCCCAGAAAG". Through a sliding window with a window size of 3, as shown in Figure 3(A), the k-mer fragment results are: "ACA", "CAA", "AAA", "AAG", "AGG", "GGG", "GGA", "GAG", "AGG", "GGC", "GCA", "CAG", "AGA", "GAA", "AAC", "ACG", "CGT", "GTA", "TAA", "AAA", "AAG", "AGC", "GCC", "CCC", "CCA", "CAG", "AGA", "GAA", "AAA" and "AAG", a total of 30 sequence fragments.

[0095] S32: According to all k-mer sequence fragments obtained in step S31, the middle bases are replaced with wildcards.

[0096] Taking piR-hsa-528 as an example, as shown in Figure 3(A), the k-mer sequence fragments "ACA", "CAA" and "AAG" were converted into wildcard "A*A", "C*A" and "A*G" fragment representations.

[0097] S33: A graph representation is constructed using k-mer sequence fragments and k-mer sequence fragments with wildcards, wherein the points of the graph correspond to the k-mer sequence fragments and the k-mer sequence fragments with wildcards, the edges correspond to the links of the k-mer sequence fragments, and the k-mer sequence fragments with wildcards are connected to the k-mer sequence fragments whose intermediate bases may be replaced.

[0098] Taking piR-hsa-528 as an example, non-repetitive k-mer sequence fragments were converted into 22 nodes, and k-mer sequence fragments with wildcards were converted into 11 nodes. They were first connected in the order of the k-mer sequences, and then the k-mer sequence fragments with wildcards and the k-mer sequence fragments whose intermediate bases might be replaced were connected (i.e., "A*A" was connected with "ACA", "AAA" and "AGA" were connected). The final graphical representation of the conversion is shown in Figure 3(B).

[0099] Step S4, by learning the biological significance knowledge from a large amount of piRNA sequence data in a pre-trained model of k-mer fragment sequences in an unsupervised manner, an embedding space is constructed so that biologically similar k-mer sequences are closer in the space.

[0100] Step S4 specifically includes:

[0101] S41: In order to obtain a distributed representation of k-mer sequences, k-mer (k=3) sequences were first extracted from all piRNA sequences collected from the piRBase database;

[0102] S42: Model the relationship between the context sequence and the target k-mer sequence through the Skip-gram model;

[0103] S43: Referring to Word2Vec, we maximize the following function through such unsupervised large-scale corpus learning to obtain the distributed representation of k-mer sequences:

[0104]

[0105]

[0106] Where N is the length of the training sequence, 2c is the window size, and w i is the center of the window, W is the number of k-mers, v w and v' w are the input and output representations of a particular k-mer sequence w. The probability p(w i+j |w i ) uses the softmax function form.

[0107] Step S5, obtaining the word vector representation of the k-mer sequence according to step 4, and using the improved graph convolutional neural network as a classification model to construct a classification model;

[0108] Figure 4 The structure of the classification model in step 5 is shown in FIG. Step S5 specifically includes: Figure 4 As shown,

[0109] S51: Obtain the word vector representation of the k-mer sequence according to step 4. The word vector of the k-mer sequence containing the wildcard is represented by the sum of the related k-mers.

[0110] S52: Use the improved graph convolutional neural network as the classification model. The model is divided into a graph convolution module, a convolution module, and a fully connected network module.

[0111] S53: Use two graph convolutional layers, both of which apply the GraphSAGE method to sample neighbor nodes and perform average aggregation based on the node’s own word vectors. The calculation is as follows:

[0112] h' v =σ(Bh v +W·mean({h u |u∈N(v)})) Formula 5

[0113] where h vis the feature of vertex v, N(v) is the set of neighboring points of vertex v in the graph, σ is a nonlinear activation function, and W and B are matrices to be learned;

[0114] S54: The convolution module uses one-dimensional convolution to reduce the dimension of the output of the graph convolution layer;

[0115] S55: A fully connected layer is used at the end of the network to integrate the features after dimensionality reduction, and softmax logistic regression can be used for classification.

[0116] Step S6, determining the sequence data for the classification scores obtained in step 5, and taking the classification with a higher probability as the final result of whether the current piRNA is related to exosomes.

[0117] For example, piR-hsa-528, according to the network, has a probability of belonging to the category of exosome-related piRNA of 0.812, and a probability of not belonging to the category of exosome-related piRNA of 0.198, so it is finally determined to be an exosome-related piRNA.

[0118] The beneficial effects of the present invention are:

[0119] Since the present invention adopts the latest and most authoritative database as a reliable data source, adopts strict filtering conditions to construct a reliable benchmark data set, uses k-mer fragments and k-mer fragments containing wildcards to construct graphs, and applies graph attention networks in the classification process, the subsequent classification results are more reliable and credible, and the prediction error is reduced.

[0120] The graph neural network model was used for the first time to identify exosome-related piRNA, so it has the advantages of novel scheme and accurate results.

[0121] In the present invention, piRNA-related data are deeply analyzed to determine exosome-related piRNA. Authoritative databases such as RNALocate and piRBase provide a large amount of high-quality piRNA data, which lays the foundation for studying exosome-related piRNA. PiRNA subcellular localization data can be obtained from the RNALocate database. Based on the gold standard data of piRNA, an identification benchmark data set is constructed, and a graph neural network model is applied to identify exosome-related piRNA.

[0122] In addition to considering k-mer sequence fragments, the present invention also considers k-mer sequence fragments with wildcards as nodes to construct graphs. The exosome-related piRNAs identified by the present invention can be used to reveal the functions of piRNAs in the occurrence and development of diseases, and may also provide new biomarkers and targets for diagnosis and treatment.

[0123] The present invention designs a method for identifying exosome-related piRNA based on a graph neural network model. The method and results are very useful for researchers exploring the function of piRNA through next-generation sequencing applications. The methods and resources provided by the present invention will be helpful for studying the function of piRNA in various complex diseases.

[0124] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for identifying exosome-related piRNA based on graph neural network, characterized in that: The following steps are involved: Step S1, constructing a benchmark dataset for exosomal piRNA identification; Step S2, using a two-step positive-unlabeled learning method, select piRNA sequences with the same number as the positive set from the unlabeled dataset as reliable negative set samples; Step S3, constructing a graph based on the sequence information; The step S3 specifically includes: S31: extract k-mer sequences from the piRNA sequences of the benchmark dataset by sliding windows, where k = 3; For a given sequence S=(S1, S2, ..., S n ), convert S to k-mer sequence fragments, where S i ∈{A, G, C, T, N}, the sequence is in cDNA form: S32: According to all k-mer sequence fragments obtained in step S31, the middle bases are replaced with wildcards; S33: construct a graph representation using k-mer sequence fragments and k-mer sequence fragments with wildcards, wherein the points of the graph correspond to the k-mer sequence fragments and the k-mer sequence fragments with wildcards, the edges correspond to the links of the k-mer sequence fragments, and the k-mer sequence fragments with wildcards are connected to the k-mer sequence fragments whose intermediate bases may be replaced; Step S4, by learning the biological significance knowledge from a large amount of piRNA sequence data in a pre-trained model of k-mer fragment sequences in an unsupervised manner, thereby constructing an embedding space so that biologically similar k-mer sequences are closer in the space; The step S4 specifically includes: S41: In order to obtain a distributed representation of k-mer sequences, k-mer sequences were first extracted from all piRNA sequences collected from the piRBase database; S42: Model the relationship between the context sequence and the target k-mer sequence through the Skip-gram model; S43: Referring to Word2Vec, we maximize the following function through such unsupervised large-scale corpus learning to obtain the distributed representation of k-mer sequences: Where N is the length of the training sequence, 2c is the window size, and w i is the center of the window, W is the number of k-mers, v w Input for a specific k-mer sequence w, v' w is the output for a specific k-mer sequence w; the probability p(w i+j |w i )Use the softmax function form; Step S5, obtaining the word vector representation of the k-mer sequence according to step 4, and using the improved graph convolutional neural network as a classification model to construct a classification model; The step S5 specifically includes: S51: Obtain the word vector representation of the k-mer sequence according to step 4, and the word vector of the k-mer sequence containing the wildcard position is represented by the sum of the related k-mers; S52: Use the improved graph convolutional neural network as the classification model. The model is divided into a graph convolution module, a convolution module, and a fully connected network module. S53: Use two graph convolutional layers, both of which apply the GraphSAGE method to sample neighbor nodes and perform average aggregation based on the node’s own word vectors. The calculation is as follows: h′ v = σ(Bh v + W·mean({h|u ∈ N(v)})) Equation 5 where h v is the feature of vertex v, N(v) is the set of neighboring points of vertex v in the graph, σ is a nonlinear activation function, and W and B are matrices to be learned; S54: The convolution module uses one-dimensional convolution to reduce the dimension of the output of the graph convolution layer; S55: A fully connected layer is used at the end of the network to integrate the features after dimensionality reduction, and softmax logistic regression can be used for classification; Step S6, determining the sequence data for the classification scores obtained in step 5, and taking the classification with a higher probability as the final result of whether the current piRNA is related to exosomes.

2. The method for identifying exosome-related piRNA based on graph neural network according to claim 1, characterized in that: The step S1 specifically includes: S11: 20,867 piRNA subcellular localization data were downloaded from the RNALocate2.0 database, of which 10,438 exosome-related piRNA records were retained; S12: According to the piRNA names obtained in step S11, the corresponding sequence data were collected from the piRBase database, and the existing exosome-related piRNA sequence data were clustered using the CD-HIT (Cluster Database at High Identity with Tolerance, https: / / github.com / weizhongli / cdhit) tool to reduce sequence redundancy, and the sequence identity parameter was set to 0.8; After steps S11 and S12, 8321 exosome-related piRNA data with low redundancy were finally obtained as the positive data of the benchmark dataset; S13: Download the human piRNA gold standard dataset from the piRBase database, which contains 77,242 piRNA sequences and is used to construct an unlabeled dataset; S14: Based on the piRNA sequence data obtained in step S13, remove the positive set data according to the sequence alignment results; S15: Based on the piRNA sequence data obtained in step S14, the CD-HIT tool is used to reduce sequence redundancy, and the sequence identity parameter is still set to 0.8; S16: After steps S14 and S15, 21,979 piRNA data with low redundancy are obtained as an unlabeled dataset.

3. The method for identifying exosome-related piRNA based on graph neural network according to claim 2, characterized in that: The step S2 specifically includes: S21: Randomly select samples from the unlabeled dataset, build a training set with the existing positive set samples, and use the training set to train the random forest classifier; S22: Use the classifier to predict the remaining unlabeled samples and repeat multiple times; S23: To reduce the false negatives of predictions, the average values ​​of the prediction scores were sorted in reverse order, and the middle 8321 piRNA records were selected as reliable negative set samples.

Images