Correlation prediction method of circular RNA and drug sensitivity, computer device and storage medium

Abstract

The invention provides a circular RNA and drug sensitivity correlation prediction method, a computer device and a storage medium, and the method comprises the steps: constructing a double-type relation heterogeneous graph with intra-class node and inter-class node representations by using multi-source data from circular RNA and drugs, obtaining an intra-class embedding matrix and an inter-class embedding matrix of the circular RNA and the drug; fusing the intra-class embedding matrix and the inter-class embedding matrix by adopting a multi-kernel fusion method to obtain a final circular RNA kernel matrix and a final drug kernel matrix; and performing circular RNA and drug sensitivity correlation prediction on the circular RNA nuclear matrix and the drug nuclear matrix by using a dual Laplacian regularization method to obtain a correlation value of the circular RNA and the drug sensitivity. By applying the correlation prediction method of the circular RNA and the drug sensitivity, the prediction accuracy can be improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular, to a method for predicting the association between circular RNA and drug sensitivity, a computer device applying the method for predicting the association between circular RNA and drug sensitivity, and a computer-readable storage medium applying the method for predicting the association between circular RNA and drug sensitivity. Background Art

[0002] A large number of studies have shown that changes in the expression of circular RNA (circRNA) can often significantly affect the sensitivity of cells to drugs. Therefore, studying the association between circular RNA and drug sensitivity is important for further understanding the function of circular RNA and developing new drugs and new uses of old drugs. However, verifying this association using traditional biological test methods is very expensive and time-consuming. With the continuous accumulation of big data related to circular RNA, there is an urgent need to develop efficient and accurate artificial intelligence inference models to help discover potential associations between circular RNA and drug sensitivity.

[0003] An existing deep learning model is the GATECDA model, which is used to predict the association between circular RNA and drug sensitivity. The GATECDA model uses a graph-structured attention autoencoder method to collect data such as circular RNA sequence information, drug structure, and circular RNA-drug sensitivity association, calculate the similarity between circular RNA and drugs, and then use these data to generate a graph structure in GATE, where the nodes are represented as vectors of circular RNA and drug sensitivity association. Finally, the graph structure generated by GATE is used as input to the fully connected neural network for circular RNA-drug sensitivity association prediction for learning and prediction. There is also a MNGACDA model, which uses a graph attention autoencoder multimodal network to generate a low-dimensional embedding representation of the association between circular RNA and drug sensitivity based on multi-source heterogeneous circular RNA and drug sensitivity association data. This low-dimensional embedding representation is then input into the decoder network, and the association between circular RNA and drug sensitivity is predicted based on the score.

[0004] However, the above methods cannot well infer the potential association between circular RNA and drug sensitivity, and further in-depth research on new association representation patterns and prediction models is needed. Summary of the invention

[0005] The first object of the present invention is to provide a method for predicting the association between circular RNA and drug sensitivity, which can improve the accuracy of predicting the association between circular RNA and drug sensitivity.

[0006] The second object of the present invention is to provide a computer device that can improve the accuracy of predicting the association between circular RNA and drug sensitivity.

[0007] The third object of the present invention is to provide a computer-readable storage medium that can improve the accuracy of predicting the association between circular RNA and drug sensitivity.

[0008] In order to achieve the above-mentioned first purpose, the method for predicting the association between circular RNA and drug sensitivity provided by the present invention includes: using multi-source data from circular RNA and drugs to construct a dual-type relationship heterogeneous graph represented by intra-class nodes and inter-class nodes to obtain the intra-class embedding matrix and the inter-class embedding matrix of circular RNA and drugs; using a multi-core fusion method to fuse the intra-class embedding matrix and the inter-class embedding matrix to obtain the final circular RNA core matrix and the final drug core matrix; applying a dual Laplace regularization method to the circular RNA core matrix and the drug core matrix to predict the association between circular RNA and drug sensitivity, and obtaining the association value between circular RNA and drug sensitivity.

[0009] It can be seen from the above scheme that the method for predicting the association between circular RNA and drug sensitivity of the present invention constructs a dual-type relationship heterogeneous graph with intra-class and inter-class node representations by using circular RNA and drug data, uses an encoder based on an attention mechanism for learning, and adopts a multi-core fusion method to integrate the intra-class embedding matrix and the inter-class embedding matrix association embedding. Finally, the dual Laplace regularization method is used to predict the potential circular RNA-drug sensitivity association, so that the circular RNA and drug data can be more comprehensively integrated, thereby improving the accuracy of the association prediction between circular RNA and drug sensitivity.

[0010] In a further scheme, a dual-type relational heterogeneous graph with intra-class node representation and inter-class node representation is constructed using multi-source data from circular RNA and drugs, and the steps of obtaining the intra-class embedding matrix and the inter-class embedding matrix of circular RNA and drugs include: obtaining circular RNA sequence similarity, drug structure similarity, circular RNA Gaussian interaction kernel similarity and drug Gaussian interaction kernel similarity; using a nonlinear fusion method to calculate the circular RNA integrated similarity based on the circular RNA sequence similarity and the circular RNA Gaussian interaction kernel similarity; using a nonlinear fusion method to calculate the circular RNA integrated similarity based on the drug structure similarity and the drug Gaussian interaction kernel similarity. Calculate the drug integration similarity; obtain the circular RNA initialization representation matrix and the drug initialization representation matrix of the intra-class node attention encoder according to the circular RNA integration similarity and the drug integration similarity; use the intra-class node attention encoder to learn the intra-class nodes according to the circular RNA initialization representation matrix and the drug initialization representation matrix to obtain the circular RNA intra-class embedding matrix and the drug intra-class embedding matrix; use the inter-class node attention encoder to learn the inter-class nodes according to the circular RNA intra-class embedding matrix and the drug intra-class embedding matrix to obtain the circular RNA inter-class embedding matrix and the drug inter-class embedding matrix.

[0011] It can be seen that by calculating the Gaussian interaction kernel similarity between circular RNA and drugs to construct a circular RNA-drug association network, the topological information of circular RNA and drugs can be well measured in the relationship graph. At the same time, the intra-class node encoder can be used to aggregate node information of the same type, while the inter-class node encoder is used to aggregate information of neighboring nodes of different types. Using the intra-class node encoder and the inter-class node encoder for learning and constructing a dual-type relationship heterogeneous graph between circular RNA molecules and drugs, the relationship between different types of nodes can be fully utilized to extract the association information between circular RNA molecules and drug sensitivity, thereby improving the accuracy of prediction.

[0012] In a further embodiment, the step of obtaining the circular RNA initialization representation matrix and the drug initialization representation matrix of the intra-class node attention encoder according to the circular RNA integration similarity and the drug integration similarity includes: the initialization representation matrix of the intra-class node attention encoder corresponding to the circular RNA is obtained by the following formula: Among them, S c is the circular RNA integration similarity, Y train is the sample adjacency matrix related to circular RNA and drug sensitivity, is a learnable parameter; the initialization representation matrix of the intra-class node attention encoder corresponding to the drug is obtained by the following formula: Among them, S d is the drug integration similarity, T train is the sample adjacency matrix related to circular RNA and drug sensitivity, is a learnable parameter.

[0013] It can be seen that by integrating the similarity and the sample adjacency matrix related to circular RNA and drug sensitivity to generate the initialization representation matrix of the intra-class node attention encoder, it is convenient to learn circular RNA and drug data.

[0014] In a further scheme, according to the circular RNA initialization representation matrix and the drug initialization representation matrix, the intra-class node attention encoder is used to learn the intra-class nodes to obtain the circular RNA intra-class embedding matrix and the drug intra-class embedding matrix. The steps include: defining the first importance index of the current circular RNA node pair according to the circular RNA initialization representation matrix; standardizing the first importance index to obtain the second importance index; obtaining the first embedding feature of the current circular RNA node according to the second importance index; standardizing the first embedding feature to obtain the second embedding feature; learning the first embedding feature and the second embedding feature through the intra-class node attention encoder to obtain the circular RNA intra-class embedding matrix of the current circular RNA node.

[0015] In a further scheme, according to the circular RNA initialization representation matrix and the drug initialization representation matrix, the intra-class node attention encoder is used to learn the intra-class nodes, and the steps of obtaining the circular RNA intra-class embedding matrix and the drug intra-class embedding matrix include: defining the third importance index of the current drug node pair according to the drug initialization representation matrix; standardizing the third importance index to obtain the fourth importance index; obtaining the third embedding feature of the current drug node according to the fourth importance index; standardizing the third embedding feature to obtain the fourth embedding feature; learning the third embedding feature and the fourth embedding feature through the intra-class node attention encoder to obtain the drug intra-class embedding matrix of the current drug node.

[0016] It can be seen that by training the intra-class attention encoder, the information of adjacent nodes can be used to aggregate the embedding information of nodes of the same type.

[0017] In a further scheme, according to the intra-class embedding matrix of circular RNA and the intra-class embedding matrix of drugs, the inter-class nodes are learned using the inter-class node attention encoder to obtain the steps of the circular RNA inter-class embedding matrix and the drug inter-class embedding matrix, including: obtaining the importance index of the current inter-class node pair and performing normalization to obtain the inter-class importance index; obtaining the circular RNA adjacent embedding features of the current circular RNA node and the drug adjacent embedding features of the current drug node according to the inter-class importance index; obtaining the first importance embedding association matrix of the drug nodes related to the current circular RNA node according to the circular RNA adjacent embedding features and the intra-class embedding matrix of circular RNA, learning the circular RNA adjacent embedding features and the first importance embedding association matrix through the inter-class node attention encoder to obtain the circular RNA inter-class embedding matrix; obtaining the second importance embedding association matrix of the circular RNA nodes related to the current drug node according to the drug adjacent embedding features and the intra-class embedding matrix of drugs, learning the drug adjacent embedding features and the second importance embedding association matrix through the inter-class node attention encoder to obtain the drug inter-class embedding matrix.

[0018] It can be seen that through the training of the inter-class attention encoder, the interaction nodes between different types can be processed as related information.

[0019] In a further scheme, a multi-core fusion method is used to fuse the intra-class embedding matrix and the inter-class embedding matrix to obtain the final circular RNA core matrix and the final drug core matrix, and the steps include: obtaining a circular RNA coding embedding matrix set and a drug coding embedding matrix set; based on the circular RNA coding embedding matrix set and the drug coding embedding matrix set, using the Gaussian distribution interaction profile kernel similarity function to calculate the circular RNA core matrix and drug core matrix of each layer; performing multi-core integral fusion on all circular RNA core matrices to obtain the final circular RNA core matrix; performing multi-core integral fusion on all drug core matrices to obtain the final drug core matrix.

[0020] It can be seen that the Gaussian distribution interaction profile kernel similarity function is used to calculate the circular RNA kernel matrix and drug kernel matrix of each layer, and multi-kernel integral fusion is performed to obtain the final circular RNA kernel matrix and the final drug kernel matrix, thereby improving the comprehensiveness of information acquisition and improving the performance of predicting circular RNA-drug association.

[0021] In a further embodiment, the correlation value between circular RNA and drug sensitivity is obtained by the following formula: Among them, IC is the final circular RNA core matrix, ID is the final drug core matrix, α c and is the learnable parameter matrix.

[0022] In order to achieve the second object of the present invention, the present invention provides a computer device including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the steps of the above-mentioned method for predicting the association between circular RNA and drug sensitivity are implemented.

[0023] In order to achieve the third object of the present invention, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a controller, the steps of the above-mentioned method for predicting the association between circular RNA and drug sensitivity are implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a flow chart of an embodiment of the method for predicting the association between circular RNA and drug sensitivity of the present invention.

[0025] Figure 2 It is a flow chart of obtaining the intra-class embedding matrix and the inter-class embedding matrix of circular RNA and drugs in the embodiment of the method for predicting the association between circular RNA and drug sensitivity of the present invention.

[0026] Figure 3 It is a coordinate schematic diagram of the degree distribution of circular RNA nodes and drug nodes in the embodiment of the circular RNA and drug sensitivity association prediction method of the present invention.

[0027] Figure 4It is a schematic diagram of the calculation process of the prediction model in the embodiment of the method for predicting the association between circular RNA and drug sensitivity of the present invention.

[0028] The present invention is further described below in conjunction with the accompanying drawings and embodiments. DETAILED DESCRIPTION

[0029] The method for predicting the association between circular RNA and drug sensitivity of the present invention is an application program used in a computer device, and is used to predict the association between circular RNA and drug sensitivity.

[0030] Example of a method for predicting the association between circular RNA and drug sensitivity:

[0031] like Figure 1 As shown, in this embodiment, when the association prediction method of circular RNA and drug sensitivity is working, step S1 is first executed to construct a dual-type relationship heterogeneous graph represented by intra-class nodes and inter-class nodes using multi-source data from circular RNA and drugs, and obtain the intra-class embedding matrix and inter-class embedding matrix of circular RNA and drugs.

[0032] In this embodiment, when constructing a dual-type relationship heterogeneous graph represented by intra-class nodes and inter-class nodes using multi-source data from circular RNA and drugs to obtain the intra-class embedding matrix and the inter-class embedding matrix of circular RNA and drugs, step S11 is first performed to obtain the circular RNA sequence similarity, drug structure similarity, circular RNA Gaussian interaction kernel similarity, and drug Gaussian interaction kernel similarity.

[0033] When obtaining multi-source data on circular RNA and drugs, the circRic database of circular RNA and drug sensitivity and drug sensitivity data from the GDSC database were used. The Wilcoxon test (<0.05) was used to extract significant circular RNA-drug sensitivity association data as the benchmark data set, which included 271 circular RNAs, 218 drugs and 4134 circular RNA and drug sensitivity association data. Circular RNA and drugs are represented as two different types of nodes in the network. The node set of circular RNA is N c , defined as Similarly, the node set of drugs is N d , defined as Adjacency Matrix Used to store circular RNA and drug sensitivity association data. In this matrix, 271 rows represent the number of circular RNAs and 218 columns represent the number of drugs. i (1≤i≤N c ) and drugs j (1≤j≤N d) is related, then Y ij =1, otherwise Y ij = 0. During the training phase, all Y ij = 1 is considered a positive sample, and the others are considered negative samples. Randomly mask some positive samples from the adjacency matrix Y to obtain the matrix Y train The similarity between circular RNA and drugs was calculated by using the circular RNA host gene sequence database from the National Center for Biotechnology (NCBI) and the PubChem database of drug structure information.

[0034] When calculating the similarity of circular RNA, the similarity between circular RNA molecules is defined by calculating the sequence similarity between the host genes associated with the circular RNA molecules. In this way, the calculation of the similarity between circular RNAs is transformed into the calculation of the similarity between their host gene sequences. The sequence similarity between host genes is based on the sequence Levenshtein distance, which can be calculated by the ratio function of the Levenshtein package of Python, and the similarity matrix is ​​defined Storing circular RNA sequence similarity data is a technique known to those skilled in the art and will not be described in detail herein.

[0035] When calculating the similarity of drugs, the structure of the drug has a great influence on its active function. The drug structure similarity calculation adopts a well-known method, and the specific process is as follows: First, the drug structure data is downloaded from the PubChem public database; then the structural similarity between drugs is calculated based on the topological fingerprint data of each drug, and the similarity matrix is ​​defined Storing drug structure similarity data is a technique known to those skilled in the art and will not be described in detail herein.

[0036] After obtaining the circular RNA sequence similarity and the drug structure similarity, in order to be able to measure the topological information of the circular RNA and the drug in the relationship graph, in this embodiment, the circular RNA and drug association network is constructed by calculating the Gaussian interaction kernel similarity between the circular RNA and the drug.

[0037] When calculating the similarity of the circular RNA Gaussian interaction kernel, the binary vector BI (c i ) and BI(c j ) represents circular RNA node c i , circular RNA node c j The association information between all drugs, namely BI (c i ) is the matrix Y train The i-th row of , BI(c j ) is the matrix Y train The jth row of . The circular RNA c is calculated by the following formula iCircRNAc j The Gaussian interaction kernel similarity CGS (c i ,c j ):

[0038] CGS(c i ,c j ) = exp(γ c ||BI(c i )-BI(c j )|| 2 ),

[0039]

[0040] Among them, the parameter α c Generally set to 1.

[0041] Similarly, when calculating the Gaussian interaction kernel similarity of drugs, the binary vector BI (d i ) and BI(d j ) represents drug node d i , drug node d j The association information between all circular RNAs, namely BI(d i ), BI(d j ) is the matrix Y train The i-th and j-th columns of . The drug d is calculated by the following formula i and d j Gaussian interaction kernel similarity DGS(d i ,d j ):

[0042] DGS(d i ,d j ) = exp(γ d ||BI(d i )-BI(d j )|| 2 ),

[0043]

[0044] Among them, the parameter α d Generally set to 1.

[0045] After obtaining the circular RNA sequence similarity, drug structure similarity, circular RNA Gaussian interaction kernel similarity, and drug Gaussian interaction kernel similarity, step S12 is performed to calculate the circular RNA integrated similarity based on the circular RNA sequence similarity and the circular RNA Gaussian interaction kernel similarity, and to calculate the drug integrated similarity based on the drug structure similarity and the drug Gaussian interaction kernel similarity using a nonlinear fusion method.

[0046] When calculating the circular RNA integration similarity, the Levenshtein distance similarity of the host gene sequences of different circular RNAs is normalized, and the formula is as follows:

[0047]

[0048] Then, the K-nearest neighbor (KNN) algorithm was used to calculate the circular RNA node c i The sequence similarity of CKNN1 is as follows:

[0049]

[0050] Among them, N′ i Circular RNA node c i The set of K circRNA neighbor nodes, including the circRNA node c i This operation is based on the following assumption: the higher the local similarity, the higher its reliability. i The similarity between the nearby circular RNA nodes and the distal circular RNA nodes is low and set to 0.

[0051] Similarly, after normalization of the CGS of Gaussian interactions between different circular RNAs, the formula is as follows:

[0052]

[0053] Then, the K-nearest neighbor (KNN) algorithm was used to calculate the circular RNA node c i The Gaussian interaction similarity of CKNN2 is as follows:

[0054]

[0055] Among them, N′ i Circular RNA node c i The set of K circRNA neighbor nodes, including the circRNA node c i This operation is based on the following assumption: the higher the local similarity, the higher its reliability. i The similarity between the nearby circular RNA nodes and the distal circular RNA nodes is low and set to 0.

[0056] Update each circular RNA similarity matrix as follows:

[0057] CSS′ (t+1) =CKNN1×CSS′ (t) ×(CKNN1) T ;

[0058] CGS′(t+1) =CKNN2×CGS′(t)×(CKNN2) T ;

[0059] The above iterations stop when the convergence condition is met. The convergence condition of the algorithm is defined as and Assuming that the above process converges after t iterations, the integrated similarity matrix of circular RNA is:

[0060]

[0061] When calculating the drug integration similarity, the Levenshtein distance similarity of different drug structures is normalized, and the formula is as follows:

[0062]

[0063] Then, the K-nearest neighbor (KNN) algorithm is used to calculate the drug node d i The structure of Levenshtein distance similarity DKNN1 is as follows:

[0064]

[0065] Among them, N′ i is the drug node c i The set of K drug neighbor nodes, including drug node d i This operation is based on the following assumption: the higher the local similarity, the higher its reliability. Therefore, the drug node d i The drug nodes nearby have high similarity with it. The drug nodes far away have low similarity with it and are set to 0.

[0066] Similarly, after normalization of the Gaussian interaction kernel similarity DGS of different drugs, the formula is as follows:

[0067]

[0068] Then, the K-nearest neighbor (KNN) algorithm is used to calculate the drug node d i The Gaussian interaction similarity of DKNN2 is as follows:

[0069]

[0070] Among them, N′ i is the drug node d i The set of K drug neighbor nodes, including drug node d i This operation is based on the following assumption: the higher the local similarity, the higher its reliability. Therefore, the drug node d iThe drug nodes nearby have high similarity with it. The drug nodes far away have low similarity with it and are set to 0.

[0071] Update each drug similarity matrix as follows:

[0072] DSS′ (t+1 ) = DKNN1 × DSS′ (t) ×(DKNN1) T ;

[0073] DGS′ (t+1) =DKNN2×DGS′ (t) ×(DKNN2) T ;

[0074] The above iterations stop when the convergence condition is met. The convergence condition of the algorithm is defined as and Assuming that the above process converges after t iterations, the integrated similarity matrix of the drugs is:

[0075]

[0076] After obtaining the circular RNA integration similarity and the drug integration similarity, step S13 is performed to obtain the circular RNA initialization representation matrix and the drug initialization representation matrix of the intra-class node attention encoder according to the circular RNA integration similarity and the drug integration similarity.

[0077] Since the adjacency networks describing objects in different biomedical fields are often sparse, this means that there are many nodes with small degrees (number of connected edges). Figure 3 As shown in the figure, 100 circular RNAs and 80 drugs are randomly selected from the data set to calculate the node degrees. It can be seen that most of the nodes have low degrees, whether circular RNA nodes or drug nodes, because in the biomedical field, each drug is often only related to a few circular RNA molecules, but not to most circular RNA molecules. Therefore, the nodes are divided into head nodes and tail nodes according to their degree values. V represents the set of nodes in the graph, V h and V t Represent the set of head nodes and tail nodes respectively. For each node v∈V, define N v Represents the set of adjacent (edge-connected) nodes of v. Set a threshold K and define the tail node as the set of nodes whose degree is less than or equal to the threshold K, that is, V t ={v:|N v |≤K}; and the head node set is the complement of the tail node, that is, V h ={v:|N v |>K}, K is a hyperparameter. In this way, the association between circular RNA and drugs is divided into the following four types:

[0078] 1. The head node of the circular RNA is associated with the head node of the drug;

[0079] 2. The circular RNA head node is associated with the drug's tail node;

[0080] 3. The circular RNA tail node is associated with the drug head node;

[0081] 4. The tail node of the circular RNA is associated with the tail node of the drug.

[0082] In the circular RNA similarity network and drug similarity network, the concepts of head node and tail node can also be used to define three types of associations:

[0083] 1. Associations between head nodes;

[0084] 2. The association between the head node and the tail node;

[0085] 3. The association between tail nodes.

[0086] As mentioned above, the three intra-class relationships can be expressed as R intra = {1,2,3}, and the four inter-class relationships are represented by R inter ={1,2,3,4}.

[0087] In this embodiment, when the circular RNA initialization representation matrix and the drug initialization representation matrix of the intra-class node attention encoder are obtained according to the circular RNA integration similarity and the drug integration similarity, for a given intra-class node pair, such as the node pair of circular RNA node i and circular RNA node j The circular RNA initialization representation matrix is in, is a learnable parameter, is the circular RNA node n i In the adjacency matrix Similarly, for a given drug-type node pair, the drug initialization representation matrix is in, is a learnable parameter, h d,j ∈R d′ (1≤j≤N d ) is the drug node m j In the adjacency matrix The j-th column vector in .

[0088] After obtaining the circular RNA initialization representation matrix and the drug initialization representation matrix, execute step S14, and use the intra-class node attention encoder to learn the intra-class nodes according to the circular RNA initialization representation matrix and the drug initialization representation matrix to obtain the circular RNA intra-class embedding matrix and the drug intra-class embedding matrix.

[0089] In this embodiment, according to the circular RNA initialization representation matrix and the drug initialization representation matrix, the intra-class node attention encoder is used to learn the intra-class nodes to obtain the circular RNA intra-class embedding matrix and the drug intra-class embedding matrix. The steps include: defining the first importance index of the current circular RNA node pair according to the circular RNA initialization representation matrix; standardizing the first importance index to obtain the second importance index; obtaining the first embedding feature of the current circular RNA node according to the second importance index; standardizing the first embedding feature to obtain the second embedding feature; learning the first embedding feature and the second embedding feature through the intra-class node attention encoder to obtain the circular RNA intra-class embedding matrix of the current circular RNA node.

[0090] Specifically, for the current circular RNA node pair (n i ,n j ) When training the self-attention encoder, define the first importance index of the current circular RNA node pair The calculation is as follows:

[0091]

[0092] Among them, || represents the link operation, is a learnable weight vector. LeakyRelu is a nonlinear activation function that is widely used in attention-based neural networks.

[0093] Standardize the first importance index to obtain the second importance index

[0094]

[0095] in, Indicates that in a given relationship Φ k When circular RNA node n i The set of adjacent nodes.

[0096] Next, the first embedding feature of the current circular RNA node is obtained according to the second importance index. i Embedding features The calculation is as follows:

[0097]

[0098] in, Represents the normalization operation of different types of layer relationships. This formula is used to normalize the circular RNA node n i In order to obtain the circular RNA node n i For more comprehensive embedding information, the first embedding feature is obtained by the following formula:

[0099]

[0100] Where q∈R 2d′ is a trainable parameter.

[0101] The first embedding feature is normalized to obtain the second embedding feature. In this embodiment, the first embedding feature is normalized using the softmax function. As shown below:

[0102]

[0103] Finally, the first embedding feature and the second embedding feature are learned through the intra-class node attention encoder to obtain the circular RNA node n i The circular RNA class type embedding matrix is ​​as follows:

[0104]

[0105] in, Represents circular RNA node n i The importance parameter of the intra-class type within all circular RNA nodes, λ c is a learnable parameter, adjustable and The proportion of and λ c Can be obtained through training.

[0106] In this embodiment, according to the circular RNA initialization representation matrix and the drug initialization representation matrix, the intra-class node attention encoder is used to learn the intra-class nodes to obtain the circular RNA intra-class embedding matrix and the drug intra-class embedding matrix. The step also includes: defining the third importance index of the current drug node pair according to the drug initialization representation matrix; standardizing the third importance index to obtain the fourth importance index; obtaining the third embedding feature of the current drug node according to the fourth importance index; standardizing the third embedding feature to obtain the fourth embedding feature; learning the third embedding feature and the fourth embedding feature through the intra-class node attention encoder to obtain the drug intra-class embedding matrix of the current drug node.

[0107] Specifically, for the current drug node pair (mi ,m j ) When training the self-attention encoder, define the third importance index of the current drug node pair The calculation is as follows:

[0108]

[0109] Among them, || represents the link operation, is a learnable weight vector. LeakyRelu is a nonlinear activation function that is widely used in attention-based neural networks.

[0110] The fourth importance index is obtained by normalizing the third importance index

[0111]

[0112] in, Indicates that in a given relationship Φ k Time drug node m i The set of adjacent nodes.

[0113] Next, the third embedding feature of the current drug node is obtained according to the fourth importance index, and the drug node m i Embedding features The calculation is as follows:

[0114]

[0115] in, Represents the normalization operation of different types of layer relationships. Using this formula, the drug node m i The aggregate information of drug nodes m is fused to have different specific relationships. i More comprehensive embedding information, the third embedding feature is obtained by the following formula:

[0116]

[0117] Where q∈R 2d′ is a trainable parameter.

[0118] The third embedding feature is normalized to obtain the fourth embedding feature. In this embodiment, the third embedding feature is normalized using the softmax function. As shown below:

[0119]

[0120] Finally, the third and fourth embedding features are learned through the intra-class node attention encoder to obtain the drug node m iThe drug class embedding matrix is ​​as follows:

[0121]

[0122] in, Represents drug node m i The importance parameter of the intra-class type within all drug nodes, λ d is a learnable parameter, adjustable and The proportion of and λ d Can be obtained through training.

[0123] set up and They represent the attention encoding of the input layer of the circular RNA and drug node embedding matrix respectively. Assuming that the intra-class attention encoder has t layers, the output of the previous layer is taken as the input of the next layer. Repeating the above process can obtain the t-layer circular RNA and drug node embedding matrix for circular RNA and drugs

[0124] After the circular RNA intra-class embedding matrix and the drug intra-class embedding matrix are obtained, step S15 is executed to learn the inter-class nodes using the inter-class node attention encoder according to the circular RNA intra-class embedding matrix and the drug intra-class embedding matrix to obtain the circular RNA inter-class embedding matrix and the drug inter-class embedding matrix.

[0125] In this embodiment, according to the intra-class embedding matrix of circular RNA and the intra-class embedding matrix of drugs, the inter-class node attention encoder is used to learn the inter-class nodes, and the steps of obtaining the inter-class embedding matrix of circular RNA and the inter-class embedding matrix of drugs include: obtaining the importance index of the current inter-class node pair and performing normalization processing to obtain the inter-class importance index; obtaining the circular RNA adjacent embedding features of the current circular RNA node and the drug adjacent embedding features of the current drug node according to the inter-class importance index; obtaining the first importance embedding association matrix of the drug nodes related to the current circular RNA node according to the circular RNA adjacent embedding features and the intra-class embedding matrix of circular RNA, learning the circular RNA adjacent embedding features and the first importance embedding association matrix through the inter-class node attention encoder to obtain the circular RNA inter-class embedding matrix; obtaining the second importance embedding association matrix of the circular RNA nodes related to the current drug node according to the drug adjacent embedding features and the intra-class embedding matrix of drugs, learning the drug adjacent embedding features and the second importance embedding association matrix through the inter-class node attention encoder to obtain the drug inter-class embedding matrix.

[0126] Specifically, in obtaining the circular RNA inter-type embedding matrix First, let the circular RNA node n i ∈C and drug node m j ∈D. is the circular RNA node n i The circular RNA-like endogenous embedding matrix, is the drug node m j The intra-drug type embedding matrix of the drug class. When the importance index of the current inter-class node pair is obtained and standardized, the inter-class importance index is obtained. i ,m j ) It can be calculated by the following formula:

[0127]

[0128] Importance Index After standardization, we can get the importance index between classes:

[0129]

[0130] in, Represents circular RNA node n i The mutual relationship between adjacent nodes between classes φ m . are two sets of learnable parameters, is a learnable weight vector.

[0131] Next, the circular RNA adjacency embedding features of the current circular RNA node are obtained based on the inter-class importance index. i The circular RNA neighbor embedding features aggregated from different types of nodes are calculated as follows:

[0132]

[0133] in, A layer normalization operation that represents an inter-class type association.

[0134] With circular RNA node n i Related drug nodes m j The first importance embedding correlation matrix of the inter-class type Calculated by the following formula:

[0135]

[0136]

[0137] Finally, the inter-class node attention encoder is used to learn the circular RNA adjacency embedding features and the first importance embedding association matrix, which is related to the circular RNA node ni CircRNA intertype embedding matrix of all drug-related nodes The fusion representation is as follows:

[0138]

[0139] Similarly, the drug adjacency embedding feature of the current drug node is obtained according to the inter-class importance index; the second importance embedding association matrix of the circular RNA node related to the current drug node is obtained according to the drug adjacency embedding feature and the drug intra-class embedding matrix, and the drug adjacency embedding feature and the second importance embedding association matrix are learned through the inter-class node attention encoder to obtain the drug inter-class embedding matrix, thereby obtaining the drug node m j The drug class embedding matrix

[0140] Specifically, in the drug class embedding matrix When, first, let drug node m j ∈D and circular RNA node n i ∈C. is the drug node m j The drug class embedding matrix, is the circular RNA node n i When obtaining the importance index of the current inter-class node pair and performing normalization to obtain the inter-class importance index, the inter-class node pair (m j ,n i ) It can be calculated by the following formula:

[0141]

[0142] Importance Index After standardization, we can get the importance index between classes:

[0143]

[0144] in, Represents drug node m j The mutual relationship between adjacent nodes between classes Φ n . are two sets of learnable parameters, is a learnable weight vector.

[0145] Next, the drug adjacency embedding features of the current drug node are obtained based on the inter-class importance index. j The drug adjacency embedding features aggregated from different types of nodes are calculated as follows:

[0146]

[0147] in, A layer normalization operation that represents an inter-class type association.

[0148] With drug node m j Related circular RNA nodes n i The second importance of the inter-class embedding correlation matrix Calculated by the following formula:

[0149]

[0150]

[0151] Finally, the inter-class node attention encoder is used to learn the circular RNA adjacent embedding features and the second important embedding association matrix, and the drug node m j Drug class inter-type embedding matrix of all related circular RNA nodes The fusion representation is as follows:

[0152]

[0153] set up and They represent the attention encoding of the input layer, that is, the inter-type embedding matrix set of circular RNA and drug nodes. Assuming that the inter-type attention encoder has k layers, the output of the previous layer is taken as the input of the next layer, and the above process is repeated, the k node embedding matrices of circular RNA and drugs can be obtained as follows:

[0154] After obtaining the intra-class embedding matrix and the inter-class embedding matrix, step S2 is performed to fuse the intra-class embedding matrix and the inter-class embedding matrix using a multi-core fusion method to obtain a final circular RNA core matrix and a final drug core matrix.

[0155] In this embodiment, a multi-core fusion method is used to fuse the intra-class embedding matrix and the inter-class embedding matrix to obtain the final circular RNA core matrix and the final drug core matrix, including the following steps: obtaining a circular RNA coding embedding matrix set and a drug coding embedding matrix set; based on the circular RNA coding embedding matrix set and the drug coding embedding matrix set, using a Gaussian distribution interaction profile kernel similarity function to calculate the circular RNA core matrix and the drug core matrix of each layer; performing multi-core integral fusion on all circular RNA core matrices to obtain the final circular RNA core matrix; performing multi-core integral fusion on all drug core matrices to obtain the final drug core matrix.

[0156] Specifically, after obtaining the intra-class embedding matrix and inter-class embedding matrix of circular RNA and drug, the circular RNA encoding embedding matrix set is obtained. and drug encoding embedding matrix set

[0157] Next, the Gaussian distribution interaction profile kernel similarity function was used to calculate the circular RNA kernel matrix and drug kernel matrix at each layer.

[0158] The circular RNA core matrix is:

[0159] The drug core matrix is:

[0160] in, and Respectively represent H c and H d The lth element of the i-th row, γ l Denotes the corresponding bandwidth, let γ l =γ, γ is a hyperparameter, l = 1,···,K+1, K+1 is the circular RNA encoding embedding moment set H c and drug encoding embedding matrix set H d The sum of the number of columns.

[0161] In order to make full use of the information to improve the performance of predicting the association between circular RNA and drugs, all circular RNA core matrices were fused by multi-core integration to obtain the final circular RNA core matrix as follows: All drug kernel matrices are integrated by multi-kernel integration to obtain the final drug kernel matrix as follows: in, and Represents the kernel matrix weights corresponding to circular RNA and drugs respectively.

[0162] After obtaining the final circular RNA core matrix and the final drug core matrix, step S4 is performed to apply the dual Laplace regularization method to the circular RNA core matrix and the drug core matrix to predict the association between circular RNA and drug sensitivity, and obtain the association value between circular RNA and drug sensitivity.

[0163] Specifically, when the dual Laplace regularized least squares model (DLapRLS) is used to predict the association between circular RNA and drugs, the loss function of DLapRLS is defined as follows:

[0164]

[0165] in‖·‖ F is the Frobenius norm, α c and is the learnable parameter matrix, Φ c and Φ dis the regularization parameter and tr(·) is the trace number. is the normalized Laplacian matrix:

[0166] L c =V c -1 / 2 Δ c V c -1 / 2 ,Δ c =V c -IC,

[0167] L d =V d -1 / 2 Δ d V d -1 / 2 ,Δ d =V d -ID,

[0168] in, and is a diagonal matrix.

[0169] Finally, the association value of circular RNA with drug sensitivity was predicted It is obtained from the IC and ID matrices as shown in the following formula: Among them, IC is the final circular RNA core matrix, ID is the final drug core matrix, α c and is the learnable parameter matrix.

[0170] When training the prediction model of this application, except for the parameter α c and α d In addition, the prediction model is trained and updated using the Adam algorithm. c Update by calculating the partial derivative of the loss function: set up Then α c The calculation is as follows:

[0171]

[0172] Similarly, the loss function is related to α d The partial derivatives can be calculated as follows:

[0173]

[0174] set up Then α d The calculation is as follows:

[0175]

[0176] αc and α d Initialized randomly at the beginning of model training, then

[0177] and Tune according to the calculation formula, and use the standard deep learning Adam algorithm to train and optimize other parameters. The calculation flow diagram of the prediction model of this application is as follows Figure 4 shown.

[0178] It can be seen from the above scheme that the circular RNA and drug sensitivity association prediction method of the present invention constructs a dual-type relationship heterogeneous graph with intra-class and inter-class node representations by using circular RNA and drug data, uses an encoder based on an attention mechanism for learning, and adopts a multi-core fusion method to integrate the intra-class embedding matrix and the inter-class embedding matrix association embedding. Finally, the dual Laplace regularization method is used to predict the potential circular RNA-drug sensitivity association, so that the circular RNA and drug data can be more comprehensively integrated to improve the prediction accuracy.

[0179] Computer device embodiment:

[0180] The computer device of this embodiment includes a controller, and when the controller executes a computer program, the steps in the above-mentioned embodiment of the method for predicting the association between circular RNA and drug sensitivity are implemented.

[0181] For example, a computer program may be divided into one or more modules, one or more modules are stored in a memory and executed by a controller to implement the present invention. One or more modules may be a series of computer program instruction segments capable of implementing specific functions, and the instruction segments are used to describe the execution process of the computer program in a computer device.

[0182] The computer device may include, but is not limited to, a controller and a memory. Those skilled in the art will appreciate that the computer device may include more or fewer components, or a combination of certain components, or different components, for example, the computer device may also include input and output devices, network access devices, buses, etc.

[0183] For example, the controller may be a central processing unit (CPU), other general controllers, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general controller may be a microcontroller or any conventional controller, etc. The controller is the control center of the computer device, and uses various interfaces and lines to connect the various parts of the entire computer device.

[0184] The memory can be used to store computer programs and / or modules. The controller realizes various functions of the computer device by running or executing the computer programs and / or modules stored in the memory, and calling the data stored in the memory. For example, the memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as a sound receiving function, a sound conversion to text function, etc.), etc.; the data storage area can store data created according to the use of the mobile phone (such as audio data, text data, etc.), etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (FlashCard), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0185] Computer readable storage medium embodiment:

[0186] If the module integrated in the computer device of the above embodiment is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiment of the method for predicting the association between circular RNA and drug sensitivity can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the controller, the steps of the embodiment of the method for predicting the association between circular RNA and drug sensitivity can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The storage medium may include: any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media does not include electrical carrier signals and telecommunications signals.

[0187] It should be noted that the above are only preferred embodiments of the present invention, but the design concept of the invention is not limited thereto, and any non-substantial modifications made to the present invention using this concept also fall within the protection scope of the present invention.

Claims

1. A method for predicting the association between circular RNA and drug sensitivity, characterized in that: Using multi-source data from circular RNA and drugs, a dual-type relational heterogeneous graph with intra-class node representation and inter-class node representation is constructed to obtain the intra-class embedding matrix and inter-class embedding matrix of circular RNA and drugs. The intra-class embedding matrix and the inter-class embedding matrix are fused by a multi-core fusion method to obtain a final circular RNA core matrix and a final drug core matrix; A dual Laplace regularization method is applied to the circular RNA core matrix and the drug core matrix to predict the association between the circular RNA and the drug sensitivity, so as to obtain the association value between the circular RNA and the drug sensitivity.

2. The method for predicting the association between circular RNA and drug sensitivity according to claim 1, characterized in that: The steps of constructing a dual-type relational heterogeneous graph with intra-class node representation and inter-class node representation by using multi-source data from circular RNA and drugs, and obtaining the intra-class embedding matrix and the inter-class embedding matrix of circular RNA and drugs include: Obtain circular RNA sequence similarity, drug structure similarity, circular RNA Gaussian interaction kernel similarity, and drug Gaussian interaction kernel similarity; A nonlinear fusion method is used to calculate the circular RNA integration similarity according to the circular RNA sequence similarity and the circular RNA Gaussian interaction kernel similarity, and a nonlinear fusion method is used to calculate the drug integration similarity according to the drug structure similarity and the drug Gaussian interaction kernel similarity; Obtaining a circular RNA initialization representation matrix and a drug initialization representation matrix of an intra-class node attention encoder according to the circular RNA integration similarity and the drug integration similarity; According to the circular RNA initialization representation matrix and the drug initialization representation matrix, using an intra-class node attention encoder to learn the intra-class nodes to obtain a circular RNA intra-class embedding matrix and a drug intra-class embedding matrix; According to the intra-class embedding matrix of the circular RNA and the intra-class embedding matrix of the drug, the inter-class node attention encoder is used to learn the inter-class nodes to obtain the inter-class embedding matrix of the circular RNA and the inter-class embedding matrix of the drug.

3. The method for predicting the association between circular RNA and drug sensitivity according to claim 2, characterized in that: The step of obtaining a circular RNA initialization representation matrix and a drug initialization representation matrix of an intra-class node attention encoder according to the circular RNA integration similarity and the drug integration similarity comprises: The initialization representation matrix of the intra-node attention encoder corresponding to the circular RNA is obtained by the following formula: Among them, S c is the circular RNA integration similarity, Y train is the sample adjacency matrix related to circular RNA and drug sensitivity, is a learnable parameter; The initialization representation matrix of the intra-class node attention encoder corresponding to the drug is obtained by the following formula: Among them, S d is the drug integration similarity, Y train is the sample adjacency matrix related to circular RNA and drug sensitivity, is a learnable parameter.

4. The method for predicting the association between circular RNA and drug sensitivity according to claim 3, characterized in that: The steps of learning the intra-class nodes using an intra-class node attention encoder according to the circular RNA initialization representation matrix and the drug initialization representation matrix to obtain the circular RNA intra-class embedding matrix and the drug intra-class embedding matrix include: Defining a first importance index of the current circular RNA node pair according to the circular RNA initialization representation matrix; Standardizing the first importance index to obtain a second importance index; Obtaining a first embedding feature of the current circular RNA node according to the second importance index; Performing normalization processing on the first embedded feature to obtain a second embedded feature; The first embedding feature and the second embedding feature are learned by the intra-class node attention encoder to obtain the intra-class circular RNA embedding matrix of the current circular RNA node.

5. The method for predicting the association between circular RNA and drug sensitivity according to claim 3, characterized in that: The steps of learning the intra-class nodes using an intra-class node attention encoder according to the circular RNA initialization representation matrix and the drug initialization representation matrix to obtain the circular RNA intra-class embedding matrix and the drug intra-class embedding matrix include: defining a third importance index of the current drug node pair according to the drug initialization representation matrix; Performing standardization on the third importance index to obtain a fourth importance index; Obtaining a third embedding feature of the current drug node according to the fourth importance index; Performing normalization processing on the third embedded feature to obtain a fourth embedded feature; The third embedding feature and the fourth embedding feature are learned by the intra-class node attention encoder to obtain the intra-drug embedding matrix of the current drug node.

6. The method for predicting the association between circular RNA and drug sensitivity according to claim 2, characterized in that: According to the intra-circRNA type embedding matrix and the intra-drug type embedding matrix, the steps of using an inter-class node attention encoder to learn the inter-class nodes to obtain the inter-circRNA type embedding matrix and the inter-drug type embedding matrix include: Obtain the importance index of the current inter-class node pair and perform standardization to obtain the inter-class importance index; Obtaining the circular RNA neighbor embedding feature of the current circular RNA node and the drug neighbor embedding feature of the current drug node according to the inter-class importance index; Obtaining a first importance embedding association matrix of drug nodes related to the current circular RNA node according to the circular RNA adjacency embedding features and the circular RNA inter-class embedding matrix, and learning the circular RNA adjacency embedding features and the first importance embedding association matrix through the inter-class node attention encoder to obtain the circular RNA inter-class embedding matrix; According to the drug adjacency embedding features and the drug inter-class embedding matrix, a second importance embedding association matrix of the circular RNA node related to the current drug node is obtained, and the drug adjacency embedding features and the second importance embedding association matrix are learned through the inter-class node attention encoder to obtain the drug inter-class embedding matrix.

7. The method for predicting the association between circular RNA and drug sensitivity according to claim 2, characterized in that: The steps of fusing the intra-class embedding matrix and the inter-class embedding matrix using a multi-core fusion method to obtain a final circular RNA core matrix and a final drug core matrix include: Obtaining a circular RNA encoding embedding matrix set and a drug encoding embedding matrix set; Based on the circular RNA encoding embedding matrix set and the drug encoding embedding matrix set, the circular RNA kernel matrix and drug kernel matrix of each layer are calculated using the Gaussian distribution interaction profile kernel similarity function; Performing multi-core integration fusion on all the circular RNA core matrices to obtain the final circular RNA core matrix; All the drug core matrices are subjected to multi-core integration fusion to obtain the final drug core matrix.

8. The method for predicting the association between circular RNA and drug sensitivity according to any one of claims 1 to 7, characterized in that: The correlation value between the circular RNA and the drug sensitivity is obtained by the following formula: Wherein, IC is the final circular RNA core matrix, ID is the final drug core matrix, α c and is the learnable parameter matrix.

9. A computer device comprising a processor and a memory, characterized in that: The memory stores a computer program, and when the computer program is executed by the processor, the steps of the method for predicting the association between circular RNA and drug sensitivity according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the controller, the steps of the method for predicting the association between circular RNA and drug sensitivity as described in any one of claims 1 to 8 are implemented.

Images