A Method and System for Predicting SM-miRNA Association Based on Higher-Order Topological Features

Abstract

This invention relates to the field of SM-miRNA technology, and particularly to a method and system for predicting SM-miRNA associations based on high-order topological features. The method includes: integrating and constructing a graph; obtaining an observation graph by masking some edges; obtaining several K-hop subgraphs using the K-hop algorithm; extracting local topological features of target link nodes using graph convolution; using a graph attention mechanism to aggregate node features to characterize the contours of random walks of target link nodes, and constructing a transition matrix using learned attention coefficients; making assumptions about the existence or non-existence of links for each K-hop subgraph, and constructing a transition matrix; concatenating the node, link, and graph-level features of several batches of subgraphs, and inputting the aggregated feature list into a multilayer perceptron to obtain the final feature list. This invention addresses the problems that complex network methods cannot predict new SMs and miRNAs, and that neural networks do not fully consider the influence of network topology on biomolecular associations.

Description

Technical Field

[0001] This invention relates to the field of SM-miRNA technology, and in particular to a method and system for predicting SM-miRNA associations based on higher-order topological features. Background Technology

[0002] Small RNAs (miRNAs) are a class of non-coding single-stranded RNA molecules, approximately 22 nucleotides in length, encoded by endogenous genes. They participate in the post-transcriptional regulation of gene expression in plants and animals. In molecular biology and pharmacology, small molecule drugs (SMs) are typically organic compounds with a molecular weight not exceeding 500 Daltons. Studies have shown that SM drugs can promote miRNA processing and regulate miRNA expression and function. For example, in preadipocytes, enoxacin leads to a decrease in miR-34a-5p expression and upregulation of its target genes. Furthermore, vorinostat triggers miR-769-5p / 3p-mediated proliferation inhibition in human gastric cancer and induces apoptosis through the STAT3-IGF1R-HDAC3 complex. Morphine regulates dopaminergic neuron differentiation through miR-133b. Therefore, studying the association between SMs and miRNAs can not only deepen our understanding of the characteristics of small molecule drugs and the functions of miRNAs but also promote the development of modern drugs and the treatment of clinical diseases.

[0003] Traditional biological experimental methods for exploring the association between SMs and miRNAs are time-consuming and labor-intensive. With the generation of a large number of miRNA and SM-related biological datasets, numerous computational models have emerged to explore accurate and reliable SM-miRNA associations. Wang et al. proposed a novel computational model for predicting SM-miRNA associations based on random forests (RFSMMA). Wang et al. proposed an ensemble method for predicting SM-miRNA associations based on kernel ridge regression (EKRRSMMA). This method reveals potential SM-miRNA associations by combining feature dimensionality reduction and ensemble learning. Guan et al. proposed a model for predicting and reasoning SM-miRNA associations based on graphlet interaction (GISMMA). GSIMMA uses multiple graphlet interactions to describe the complex relationships between two SMs or two miRNAs. The above are methods based on complex networks. Although complex network methods can capture nonlinear relationships between nodes, they can only predict associated SMs and miRNAs and cannot predict new SMs and miRNAs.

[0004] With the rapid development of neural networks, many deep learning-based methods have also emerged. Shen et al. adopted a new framework called SMAJL, which utilizes a joint learning model of Restricted Boltzmann Machines (RBMs) to predict SM-miRNA association scores. Ni et al. proposed a graph convolutional network model with layer attention mechanism for SM-miRNA association prediction (GCNLASMMA). Li et al. combined heterogeneous network representation learning algorithms using HeGAN and HIN2Vec to utilize the topological information of heterogeneous information networks. Compared with complex network methods, neural network-based methods have strong nonlinear modeling capabilities, but these methods do not fully consider the influence of network topology on biomolecular associations. Summary of the Invention

[0005] To address the shortcomings of existing methods, this invention solves the problems that complex network methods cannot predict new SMs and miRNAs, and that neural networks do not fully consider the impact of network topology on biomolecular associations.

[0006] The technical solution adopted in this invention is: a drug-miRNA association prediction method based on high-order topological features, comprising the following steps:

[0007] Step 1: Calculate the Gaussian interaction spectrum kernel similarity of SM and miRNA using the SM-miRNA association matrix, and integrate the Gaussian interaction spectrum kernel similarity, comprehensive similarity data, and SM-miRNA association matrix to construct a graph; based on the constructed graph, obtain the observation graph using masking, and use the K-hop algorithm to obtain several K-hop subgraphs from the observation graph;

[0008] In a preferred embodiment of the present invention, a graph is constructed as G(V,E,X), where V and E are the sets of nodes and edges in graph G, respectively, and X is the node feature, with the formula for X being:

[0009]

[0010] in, and Both represent all-zero matrices, where m and n represent the number of SMs and miRNAs, respectively; IS is the comprehensive SM similarity matrix and IM is the comprehensive miRNA similarity matrix.

[0011] As a preferred embodiment of the present invention, obtaining several K-hop subgraphs from the observation graph using the K-hop algorithm includes:

[0012] Suppose the observation graph is a graph G with N nodes. O (V,E O ,X O ); where V represents the set of nodes in graph G, E OX is the edge set in the observation graph. O This indicates the characteristics of the nodes observed in the diagram;

[0013] Let the set containing the known links in the observation graph and the unknown links in the original graph be denoted as the target link set E. C ;

[0014] For each target link ξ∈E C Constructing a K-jump subgraph G ξ .

[0015] Step 2: Input multiple K-hop subgraphs into GCN in batches to extract the local topological features of the target linked nodes, and obtain the aggregated node features, that is, the node features in the aggregated K-hop subgraph.

[0016] In a preferred embodiment of the present invention, step two specifically includes:

[0017] A two-layer GCN is used to transfer and fuse graph information to form node representations in the subgraph. The formula is:

[0018] Z = f(A', f(A', X) O )+X O (5)

[0019] Where f represents the GCN operation, F′ is the dimension of the node feature Z after GCN, and A' is a matrix composed of several K-hop subgraphs. A ξ Let be the adjacency matrix of subgraph ξ, n be the batch size, and D be the adjacency matrix of subgraph ξ. ξ For A ξ The degree matrix; X O To observe the features of the graph nodes.

[0020] Step 3: Use graph attention mechanism to characterize the outline of the random walk of the target link nodes by aggregating node features, and construct a transition matrix using the learned attention coefficients; make an assumption about the existence or non-existence of links for each K-hop subgraph, and construct a transition matrix for each; obtain the node-level, link-level, and graph-level features of the target links from the two transition matrices.

[0021] In a preferred embodiment of the present invention, step three specifically includes:

[0022] Step 31: Encode the relevance of the target linked nodes x and y into effective edge weights, using the following formula:

[0023]

[0024] Where {x,y}∈E C , There are two MLPs, F″ is the output dimension of the MLP, and zx For the features of node x, z y Features of node y;

[0025] Step 32: Calculate the random walk transition matrix based on the correlation weights of nodes x and y, using the following formula;

[0026]

[0027] Where {x,y}∈E C ;like Then p x,y =0, N(x) is the neighborhood set of x in the closed subgraph, w x,y Let x be the effective edge weights of nodes x and y;

[0028] Step 33: Denote the subgraph assuming the existence of the target link as G. + The non-existent is denoted as G. - ;

[0029] Step 34, for G + Perform at the node level Link Level Map level Feature extraction;

[0030] In a preferred embodiment of the present invention, the characteristic formulas for node level, link level, and graph level are as follows:

[0031]

[0032] in, G for τ jump + The transition matrix, where x and y are the target nodes; and tr(·) represents the transition probabilities from x to y and from y to x; tr(·) represents the trace of the chosen matrix.

[0033] Step 35, for G - Perform at the node level Link Level Map level Feature extraction.

[0034] Step 4: G of several batches of subgraphs + and G - Features at the node level, link level, and graph level are concatenated to obtain an aggregated feature list, which is then input into the MLP to obtain the final feature list.

[0035] As a preferred embodiment of the present invention, the graph-level feature is that G + and G -The difference between the graph-level return probabilities is used as the final graph-level feature. τ .

[0036] In a preferred embodiment of the present invention, the final feature list is used to obtain a score matrix through the cross-entropy loss function, and the score matrix is ​​used to obtain the predicted score.

[0037] As a preferred embodiment of the present invention, a small molecule drug-miRNA association prediction system based on higher-order topological features includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement the SM-miRNA association prediction method based on higher-order topological features.

[0038] As a preferred embodiment of the present invention, a computer-readable medium storing computer program code implements a high-order topological feature-based SM-miRNA association prediction method when executed by a processor.

[0039] The beneficial effects of this invention are:

[0040] 1. The SM-miRNA association prediction task is transformed into a K-hops subgraph classification task. Multiple K-hops subgraphs are constructed based on multiple target link nodes. 0-1 labels are assigned according to the existence of target links. The purpose of constructing K-hops subgraphs is to amplify the local topology of the corresponding target link nodes.

[0041] 2. This invention divides multiple subgraphs into batches and puts them into GCN according to the batches to extract node features, which are more detailed and distinguishable features compared to the target node.

[0042] 3. In HTFSMMA, complex network algorithms and neural network algorithms are integrated. GCN is used as the node feature of the aggregate graph, and the transition matrix constructed by the graph attention mechanism is used to simulate the state of random walk, thereby obtaining detailed high-order topological features. Attached Figure Description

[0043] Figure 1 This is a logic diagram of the High-Order Topological Feature-Based SM-miRNA Association Prediction Method (HTFSMMA);

[0044] Figure 2 (a) and (b) show the ROC curves of HTFSMMA and other methods in the five-fold cross-validation experiment on datasets 1 and 2, respectively.

[0045] Figure 3 (a) and (b) are comparison graphs of ROC curves of HTFSMMA and other methods under global leave-one cross-validation experiments on dataset 1 and dataset 2, respectively.

[0046] Figure 4 (a) and (b) are comparison plots of ROC curves of HTFSMMA and other methods on datasets 1 and 2, respectively, under local leave-one cross-validation experiments with fixed small molecule drugs.

[0047] Figure 5 (a) and (b) are comparison plots of ROC curves of HTFSMMA and other methods in local leave-one cross-validation experiments with fixed miRNAs on datasets 1 and 2, respectively. Detailed Implementation

[0048] The present invention will be further described below with reference to the accompanying drawings and embodiments. The drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0049] Starting from the topological characteristics between biomolecules, such as Figure 1 As shown, the SM-miRNA association prediction method based on high-order topological features includes the following steps:

[0050] Step 1: Calculate the Gaussian interaction spectrum kernel similarity of SM and miRNA using the SM-miRNA association matrix, and integrate the Gaussian interaction spectrum kernel similarity, comprehensive similarity data and SM-miRNA association matrix to construct a graph; based on the constructed graph, construct multiple K-jump subgraphs using masking.

[0051] The SM-miRNA association data were downloaded from the SM2miR V1.0 database and divided into dataset 1 and dataset 2 for the study. SMs data were obtained from three databases: SM2miRv1.0, DrugBank, and PubChem. miRNA data came from four databases: SM2miR, HMDD, miR2Disease, and PhenomiR. Dataset 1 contains 831 SMs, 541 miRNAs, and 664 association pairs between them. By removing SMs and miRNAs with no associations from dataset 1, dataset 2 was obtained, which contains 39 SMs, 286 miRNAs, and 664 known SM-miRNA association pairs. Here, the SM-miRNA association matrix A∈R is defined. ns×nm (ns represents the number of SMs, nm represents the number of miRNAs) to represent the association information between SMs and miRNAs; if SMs i With miRNA m j There is a correlation, A(s) i ,m j If the value is 1, then the value is 0; otherwise, the value is 0.

[0052] HTFSMMA integrates four types of SM similarity, namely chemical structure-based similarity (SS). C Similarity based on side effects SS S SS based on indication phenotype P Similarity based on functional consistency (SS) F A weighted average strategy is used to form the SM similarity, assigning equal weights to the four types of similarity; the SM similarity matrix SS is defined as follows:

[0053]

[0054] miRNA-miRNA similarity information includes a similarity matrix (MS) based on miRNA-related diseases. D and miRNA target genome functional consistency matrix MS F A weighted average strategy is used to form miRNA similarity, assigning equal weights to both types of similarity. The miRNA similarity matrix MS is defined as follows:

[0055]

[0056] Gaussian interaction spectral similarity (GIP) is calculated using the SM-miRNA association matrix to determine the Gaussian interaction spectral similarity between SMs or between miRNAs, denoted as matrices GS and GM, respectively. The combined SM similarity matrix IS is calculated as follows: Similarly, the comprehensive miRNA similarity matrix IM is calculated as follows:

[0057]

[0058] Subsequently, multiple K-hop subgraphs are constructed based on the predicted target link nodes.

[0059] Let a graph be defined as G(V,E,X), where V and E are obtained from the incidence matrix A, and are the sets of nodes and edges in graph G, respectively, and the node characteristics are... Represented as:

[0060]

[0061] in, and All represent all-zero matrices, m and n represent the number of SMs and miRNAs, respectively, IS is the comprehensive SM similarity matrix and IM is the comprehensive miRNA similarity matrix.

[0062] The observation graph is formed by covering part of the edges of graph G; assuming the observation graph is a graph G with N nodes. O (V,E O ,X O); where V represents the set of nodes in graph G, E O X is the edge set in the observation graph. O The characteristics of nodes in the observation graph are represented; the set containing known links in the observation graph and unknown links in the original graph is denoted as the target link set E. C And use this to generate training data.

[0063] According to each target link ξ∈E C They all construct a K-jump subgraph G for it. ξ .

[0064] The K-hop algorithm refers to finding the set of vertices that have a shortest path of K hops from a given vertex; 1≤K≤τ, where τ represents the maximum number of hops in K.

[0065] Constructing a K-hop subgraph utilizes the features between node pairs as a measure of link features. The method of using multi-hop neighborhood overlap is not limited to capturing only 2-hop neighbor relationships in local neighborhood overlap, nor does it extract features through the global network structure. It is a trade-off operation between local and global neighborhoods. The theoretical basis of multi-hop neighborhood overlap is to obtain higher-order topological features in the graph, which does not necessarily require a large number of hops. Higher-order topological features of the graph can be approximated by constructing a K-hop subgraph. Therefore, choosing an appropriate K is important. Too large a K will not bring much benefit but will increase memory consumption, while too small a K will lead to incomplete topological features.

[0066] For each target link, a K-hop subgraph is constructed. First, the subgraph node set is initialized with the target node's number. Target nodes are labeled using a mask operation, and their first-order neighbors are obtained by observing the edges of the graph. If the maximum hop count for a node is not set, the first-order neighbors of the target node are selected by default. If the maximum hop count is specified, after selecting the first-order neighbors of the target node, neighbor nodes with longer paths (limited by the maximum hop count) are iteratively added to the subgraph node set. After these operations, the adjacent neighbors and multi-hop neighbors of the target node are selected, and corresponding edge masks are set in the observation graph. It is important to note that when constructing the K-hop subgraph, the target node must be distinguished from other nodes, and the target node's index in the subgraph is preserved. The node features of the subgraph are extracted from the corresponding nodes in the observation graph.

[0067] When constructing the K-hop subgraph, the K-hop subgraph is labeled to generate training and test sets. Thus, the link prediction problem of the target node is transformed into a subgraph classification problem.

[0068] Step 2: Aggregate node features in the K-hop subgraph. After constructing the K-hop subgraph, input multiple K-hop subgraphs into the GCN in batches to extract the local topological features of the target linked nodes.

[0069] GCN can fuse features of subgraph nodes and features of neighboring nodes, pass messages through edges, and aggregate information from a wider range of neighboring nodes. As a neural network method, GCN propagates from layer l to layer (l+1) as follows:

[0070]

[0071] Among them, H (l+1) and H (l) Let represent the features of the (l+1)th layer and the lth layer; W(l) represents the weights of the lth layer; σ(·) represents the nonlinear activation function; A' is a matrix composed of multiple K-hop subgraphs. A ξ Let be the adjacency matrix of subgraph ξ, n be the batch size, and D be the adjacency matrix of subgraph ξ. ξ For A ξ The degree matrix; for the input layer, H 0 It is the observation graph node feature X O .

[0072] A two-layer GCN is used to transfer and fuse graph information to form node representations in the subgraph. The specific calculations are as follows:

[0073] Z = f(A', f(A', X) O )+X O (5)

[0074] Where f represents the GCN operation, and F′ is the dimension of the node feature Z after GCN.

[0075] The graph data is processed using the torch_geometric library. During data loading, appropriate indexes are generated for each batch of subgraph data samples so that the batch data can be used correctly later. During batch processing, these subgraphs share a single GCN for forward propagation and parameter updates. Each subgraph can pass through the network in parallel, which helps improve training efficiency and memory utilization.

[0076] Step 3: Construct the transition matrix. First, using the aggregated node features, the graph attention mechanism is used to characterize the outline of the random walk of the target link nodes, and the learned attention coefficients are used to construct the transition matrix. Second, make an assumption about whether the link exists or not for each subgraph, and construct transition matrices for these two types of assumptions respectively. Then, obtain the node-level, link-level, and graph-level features of the target link from the two transition matrices, and concatenate them to form an aggregated feature list.

[0077] High-order topological features are extracted. In the SM-miRNA association information, there is a topological imbalance problem that can cause bias in association prediction, namely, the link height between nodes is unbalanced. Graph attention mechanism is used to alleviate this problem.

[0078] For K-jump graph G ξ =(V ξ E ξ Z ξ The node Z is used to encode edge weights, encoding the correlation between the target linked nodes x and y as effective edge weights, as follows:

[0079]

[0080] Where {x,y}∈E C , It consists of two multilayer perceptrons (MLPs), where F″ is the output dimension of the MLP, and z... x For the features of node x, z y Let y be the feature of node y; then, calculate the random walk transition matrix P = (p x,y ):

[0081]

[0082] Where {x,y}∈E C ;like Then p x,y =0, N(x) is the set of neighbors of x in the closed subgraph; a two-headed graph attention mechanism is used, and the learned attention coefficients are used to construct the transition matrix.

[0083] The transition matrix of a graph can be used to model random walks within a graph. Random walks simulate the characteristics of a graph by randomly moving between nodes, and can help explore the topological properties of a graph to some extent; therefore, using [P τ ] x,y This represents the probability of moving from x to y in τ hops during a random walk. This model can generate random walk sequences of different lengths, thereby helping to reveal the contours of focal links, which can infer higher-order topological information of the graph and information about the relationships between nodes.

[0084] Let the subgraph assuming the existence of the target link be denoted as G. + The non-existent is denoted as G. - ; with G + For example, based on the set τ jumps, the target nodes are x and y. G for τ jump + The transition matrix for G + Perform at the node level Link Level Map level The features are extracted as follows:

[0085]

[0086] in, G for τ jump + The transition matrix, where x and y are the target nodes; and tr(·) represents the transition probabilities from x to y and from y to x; tr(·) represents the trace of the chosen matrix.

[0087] Correspondingly, G - On It's calculated in a similar way.

[0088] Step 4: Multi-level feature extraction. Obtain node-level, link-level, and graph-level features of the target link from the two transition matrices, concatenate them to form an aggregated feature list, and input it into the MLP to obtain the final feature list.

[0089] Through G + and G - The node-level, link-level, and graph-level features provide a more comprehensive understanding of the network's properties and structure; subsequently, G... + and G - The difference between the return probabilities at the graph level is used as the final graph-level feature of the subgraph. τ ; however, node-level and link-level features do not need to utilize the difference between return probabilities as the final node- and link-level features; when processing multi-batch K-hop subgraphs, the graph is processed according to batches. τ The features are stored in tensor form, which contains the differences of each graph in the batch. Finally, the node-level, link-level, and graph-level features are concatenated to obtain an aggregated feature list, which is then input into the MLP to obtain the final feature list.

[0090] Since the subgraphs constructed from SM-miRNA association data contain far fewer positive samples than those constructed from unknown links, a sigmoid-based cross-entropy loss function is used to address the imbalance between positive and negative samples in SM-miRNA association prediction, thereby improving the model's training effectiveness and classification performance. The formula for calculating the sigmoid-based cross-entropy loss function is as follows:

[0091]

[0092] Where N represents the number of samples, z i Y is a sample feature. i σ represents the actual label of the sample, and σ is the activation function.

[0093] The final feature list is used to obtain a score matrix through the cross-entropy loss function. The score matrix is ​​then used to obtain the predicted score, which determines whether the target link exists, thus achieving the purpose of SM-miRNA link prediction.

[0094] The association prediction results are input into the MLP classifier to obtain the target node association prediction score. The choice of the number of hidden layers in the MLP classifier affects the effect of different numbers of hidden layers. Taking dataset 2 as an example, the MLP uses batch normalization and ReLU as activation functions. Table 1 shows the comparison of AUC values ​​of different hidden linear layers in the five-fold cross-validation task.

[0095] Table 1 shows the AUC of different MLP hidden layers on dataset 2.

[0096]

[0097] This shows that as the number of hidden linear layers in the MLP increases, regardless of whether Gaussian interaction spectral kernel similarity (GIP) normalization similarity information is added, the AUC is highest when the number of hidden layers is 2 on dataset 2.

[0098] To measure the impact of K on the model during the construction of K-hops subgraphs, K was trained with values ​​of 1, 3, 5, 7, and 9 to observe the time consumption, memory usage, and AUC performance comparison. Here, dataset 2 is used as an example, and five-fold cross-validation was performed. Other parameters were fixed, the number of MLP layers was set to 1, and GIP features were not used. The specific comparison is shown in Table 2 below. It can be clearly seen that when the number of hops K is 7, the time consumption is only 1 minute longer than when K=5, the increase in memory consumption is small, the AUC is the best, and the overall performance is the best.

[0099]

[0100] Based on multiple experiments, the other hyperparameters of HTFSMMA are set as follows: the maximum number of nodes per hop is set to 5, the number of hops τ in the random walk is 4, the number of hidden layers in GCN is 16, the learning rate lr = 0.003, the batch size = 32, and the epoch = 50.

[0101] Model performance evaluation:

[0102] The model was evaluated using five-fold cross-validation, global leave-one-out cross-validation, and local leave-one-out cross-validation, and compared with three other advanced SM-miRNA association prediction computational models: EKRRSMMA, GCNLASMMA, and RFSMMA.

[0103] Leave-one-out cross-validation uses each known sample in the dataset as a test sample in turn, and all other known samples as training samples. In global leave-one-out cross-validation, all unknown association samples are used as candidate samples. Local leave-one-out cross-validation with a fixed miRNA uses known associations of the fixed miRNA as test samples and unknown associations with the fixed miRNA as candidate samples. Leave-one-out cross-validation with a fixed miRNA operates similarly. Five-fold cross-validation randomly divides all known associations into five equal groups, selects one group as the test sample, and uses the remaining four groups as training samples. All unknown association pairs are considered as candidate samples. Five-fold cross-validation will be run 100 times to eliminate random errors.

[0104] like Figure 2-5 Tables 3 and 4 show the AUC comparisons of HTFSMMA, EKRRSMMA, GCNSMMA, and RFSMMA on different tasks in datasets 1 and 2, respectively.

[0105] Table 3 compares the AUC of HTFSMMA, EKRRSMMA, GCNSMMA, and RFSMMA on dataset 1.

[0106]

[0107] Table 4 compares the AUC of HTFSMMA, EKRRSMMA, GCNSMMA, and RFSMMA on dataset 2.

[0108]

[0109] In addition, case studies were conducted on several specific miRNAs. When studying a specific miRNA, the information of that specific miRNA was removed from the known associations, and the miRNA was treated as a new miRNA. The model was used to predict the top 50 miRNAs most relevant to it, and the number of miRNAs that were validated by published experimental literature was observed.

[0110] The first case study uses SM5-FU as an example. SM5-FU is a synthetic fluoropyrimidine analog and a major chemotherapy drug used in advanced colorectal cancer or when the risk of recurrence is high. A total ranking list of the top 50 5-FU-related miRNAs predicted and validated using HTFSMMA was obtained, as shown in Table 5. Among the top 10 (50) predicted 5-FU-related miRNAs, 6 (27) miRNAs were confirmed, of which 4 (19) miRNAs were confirmed by the database SM2miR v1.0 and 2 (8) miRNAs were confirmed by literature. Zhang et al. found that miR-22 inhibits autophagy and promotes apoptosis to increase the sensitivity of colorectal cancer (CRC) cells to 5-fluorouracil (5-FU) treatment in vitro and in vivo. Yu et al. found that miR-125b conferred 5-FU resistance by increasing autophagy, showing enhancement of beclin-1 and cleaved LC3-II and increased formation of autophagosomes (GFP-LC3 dots).

[0111] Table 5 shows the top 50 miRNAs associated with 5-FU (CID:3385) prediction based on dataset 1.

[0112]

[0113]

[0114] The second case study used is 5-Aza-2'-deoxycytidine (5-Aza-CdR), a nucleoside analog. Incorporating 5-AZA-CdR into DNA blocks DNA methylation and may lead to the activation of specific genes (such as tumor suppressor genes). After performing HTFSMMA, the top 50 miRNAs associated with 5-Aza-dC were obtained and validated (Table 6). Among the top 10 (50) predicted 5-Aza-dC-associated miRNAs, 6 (36) were confirmed miRNAs, of which 4 (22) were confirmed by the SM2miRv1.0 database and 2 (14) were confirmed by published experimental literature. The association between 5-Aza-CdR and miRNAs is supported by a large body of literature. Campos-Viguri et al. pointed out that decreased miR-23b expression levels are associated with cervical cancer, and that miR-23b expression increased after DNA demethylation using 5-Aza-CdR. In addition, Watanabe et al. used 5-aza-CdR as a DNA methylation inhibitor to treat six NSCLC cell lines and found that mir-126 was silenced by DNA methylation of its host gene EGFL7.

[0115] Table 6 lists the top 50 miRNAs associated with 5-Aza-CdR (CID:451668) prediction, validated based on dataset 1.

[0116]

[0117]

[0118] The third type of study uses 17β-estradiol (E2) as an example. It is the main circulating estrogen in women. 17β-estradiol can not only improve insulin sensitivity, but also has cardioprotective and vasodilatory effects. The results of the first 50 miRNAs related to 5-AzadC predicted and verified using HTFSMMA are shown in Table 7. Among the first 10 (50) predicted E2-related miRNAs, 6 (26) were confirmed miRNAs. Among them, 6 (18) miRNAs were confirmed by the database SM2miRv1.0, and 1 (8) miRNAs were confirmed by published experimental literature. There is a large amount of literature to support the association between E2 and miRNAs. Reed et al. showed that BPA and E2 can downregulate miR-27b, thereby leading to the upregulation of genes important for endometrial vascularization and angiogenesis during the menstrual cycle and decidualization. Huang et al. found that E2 can activate miR-23a and p53 expression through ERα transcription, thereby enhancing the activation of miR-23a expression by p53.

[0119] Table 7 shows the top 50 miRNAs associated with E2 (CID:5757) prediction based on dataset 1.

[0120]

[0121]

[0122] Note: The first column of Table 5-8 records the first 1-25 predicted miRNAs, and the third column records the 26-50 predicted miRNAs; the above three case studies confirm the reliable predictive performance of HTFSMMA.

[0123] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for predicting SM-miRNA association based on high-order topological features, characterized in that, Includes the following steps: Step 1: Calculate the Gaussian interaction spectrum kernel similarity of SM and miRNA using the SM-miRNA association matrix, and integrate the Gaussian interaction spectrum kernel similarity, comprehensive similarity data, and SM-miRNA association matrix to construct a graph; based on the constructed graph, cover some edges to obtain the observation graph, and use the K-hop algorithm to obtain several K-hop subgraphs from the observation graph; The construction diagram is as follows , V and E The figures are respectively G The set of nodes and edges in the middle. X For node features, X The formula is: （3） in, and Both represent matrices containing all zeros. m and n These represent the number of SMs and the number of miRNAs, respectively. IS To integrate the SM similarity matrix and IM For comprehensive miRNA similarity matrix; Step 2: Extract the local topological features of the target linked nodes by batch convolving multiple K-hop subgraphs into the input graph to obtain aggregate node features; Step two includes: A two-layer GCN is used to transfer and fuse graph information to form node representations in the subgraph. The formula is: in, Indicates GCN operation, Node features after GCN Dimensions It is a matrix composed of several K-jump subgraphs. , For subgraph The adjacency matrix, For batch size, for The degree matrix; To observe the features of graph nodes; Step 3: Use graph attention mechanism to characterize the outline of the random walk of the target link nodes by aggregating node features, and construct a transition matrix using the learned attention coefficients; make an assumption about the existence or non-existence of links for each K-hop subgraph, and construct a transition matrix for each; obtain the node-level, link-level, and graph-level features of the target links from the two transition matrices. Step 4: Concatenate the node-level, link-level, and graph-level features of several batches of subgraphs to obtain an aggregated feature list. Input the aggregated feature list into the MLP to obtain the final feature list.

2. The SM-miRNA association prediction method based on high-order topological features according to claim 1, characterized in that, The K-hop algorithm was used to obtain several K-hop subgraphs from the observed graph, including: Assume the observation diagram has N A graph with 1 node ;in, V Representation diagram G The set of nodes in It is about observing the edge set in the graph. This indicates the characteristics of the nodes observed in the diagram; The set containing both known links in the observed graph and unknown links in the original graph is denoted as the target link set. ; For each target link Constructing a K-jump subgraph .

3. The SM-miRNA association prediction method based on high-order topological features according to claim 1, characterized in that, Step three specifically includes: Step 31: Link the target node and The relevance encoding is the effective edge weight, and the formula is: in, , There are two MLPs. It is the output dimension of the MLP, z x For nodes x Features, z y For nodes y Features; Step 32, based on nodes and The correlation weights are used to calculate the random walk transition matrix, and the formula is as follows; in, ;like but , It is a closed subgraph Neighborhood set, For nodes and nodes Effective edge weights between; Step 33: Denote the subgraph assuming the existence of the target link as... Non-existent is recorded as ; Step 34, for Perform at the node level Link level , diagram level Feature extraction; Step 35, for Perform at the node level Link level , diagram level Feature extraction.

4. The SM-miRNA association prediction method based on high-order topological features according to claim 3, characterized in that, The feature formulas for node level, link level, and graph level are: in, for Jumping The transition matrix, and For the target node; and for x arrive y and y arrive x The transition probability; To select the trace of the matrix.

5. The SM-miRNA association prediction method based on high-order topological features according to claim 1, characterized in that, The characteristic at the graph level is that... and The difference between graph-level return probabilities is used as the final graph-level feature. .

6. The SM-miRNA association prediction method based on high-order topological features according to claim 1, characterized in that, The final feature list is used to obtain a score matrix through the cross-entropy loss function, and the predicted score is obtained using the score matrix.

7. A small molecule drug-miRNA association prediction system based on high-order topological features, characterized in that, include: Memory is used to store instructions that can be executed by the processor; A processor for executing instructions to implement the SM-miRNA association prediction method based on high-order topological features as described in any one of claims 1-6.

8. A computer-readable medium storing computer program code, characterized in that, The computer program code, when executed by a processor, implements the SM-miRNA association prediction method based on any one of claims 1-6.

Images