Drug-lncrna relationship prediction method and system based on embedding constraint

Abstract

The application relates to a drug-lncRNA relationship prediction method and system based on embedding constraints, and the method comprises the following steps: collecting lncRNA-drug association data and preprocessing, and constructing a data set; based on the data set, an lncRNA-drug association matrix is constructed, and an lncRNA-drug bipartite graph is extracted; based on the lncRNA-drug association matrix, an lncRNA similarity matrix and a drug similarity matrix are obtained, and dimension reduction processing is performed; the lncRNA-drug bipartite graph and the dimension reduction processed lncRNA similarity matrix and drug similarity matrix are input into an LDA-GNN model and an LDA-DHG model with an embedding constraint strategy, and the probability score of lncRNA and drug association is obtained; and based on the probability score, the relationship prediction of the drug-lncRNA is realized. Through the embedding constraint strategy, the prediction performance and the distinguishing ability of the model are significantly improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of drug-lncRNA relationship prediction, and particularly relates to a drug-lncRNA relationship prediction method and system based on embedding constraints. BACKGROUND

[0002] The progress of high-throughput experimental techniques and computational methods has greatly promoted the ability of people to understand the mechanism of drug action, especially in revealing the complex interactions between drugs and non-coding RNAs at the molecular level. In recent years, predicting the associations between long non-coding RNAs (LncRNAs) and drugs (LncRNA-drug associations, LDRAs) has become a key research direction in drug repositioning and personalized therapy. LncRNAs are a class of RNA molecules that do not encode proteins, and are widely involved in gene expression regulation, cell function regulation, and disease mechanism formation. Studies have shown that LncRNAs can significantly affect the pharmacological properties of drugs by affecting the expression of drug targets or regulating key signaling pathways.

[0003] Currently, there are several databases that systematically integrate experimentally verified LDRAs. For example, D-lnc and NcRNADrug include data on the effects of drugs on LncRNA expression and LncRNAs related to drug resistance. These data provide valuable resources for in-depth study of the functional relationship between LncRNAs and drugs. However, experimental methods are usually costly and time-consuming, limiting their application in large-scale prediction tasks. Therefore, developing LDRAs prediction models based on computational methods has become a pressing problem.

[0004] Various machine learning methods have been applied to LDRAs prediction. For example, Wang et al. used an ElasticNet (EN) regression model to integrate genomic data from various cancer cell lines with drug response data to identify LDRAs. Ha et al. proposed a matrix factorization method based on lncRNA expression profiles (EMFLDA) to integrate heterogeneous biological data. Chen et al. proposed a LDRAs prediction method based on random walk (RWR), considering the importance of local structure. Although these methods have achieved some success, they often have certain limitations in capturing high-dimensional and nonlinear features and integrating various biological data.

[0005] In recent years, deep learning (DL) techniques have gained prominence due to their ability to learn hierarchical representations and capture complex patterns in data. Tang et al. proposed a GCN framework that integrates multiple similarities between lncRNAs and drugs and constructs a unified graph to perform end-to-end prediction. Liu et al. proposed a method that combines a co-learning framework and graph structure learning to model the co-variance patterns between lncRNAs and drugs. Ha et al. used a matrix factorization method based on lncRNA expression profile data and drug responses to integrate multiple heterogeneous biological features. Although existing deep learning methods have improved prediction performance to some extent, they still cannot fully exploit the potential dependencies between molecular entities due to their reliance on static feature representations and difficulty in fully capturing the complex graph structure relationships between lncRNAs and drugs.

[0006] In deep learning techniques, graph neural networks (GNNs) are increasingly popular in molecular prediction tasks because they can effectively model molecular entities as nodes and capture their relationships as edges in a graph structure. For example, Xu et al. introduced a LDRAs identification model based on graph attention networks (GAT) and graph convolutional networks (GCN) (LDA-GNN). Xu et al. developed a model for identifying DTIs (LDA-DHG) that constructs a dynamic heterogeneous graph based on drug-drug and target-target similarity networks and proposes a new progressive learning strategy to improve prediction accuracy, which is also applied to the LDRA prediction field. Yao et al. proposed a prediction model that integrates GCN and Transformer, which constructs a graph adjacency matrix based on intra-class similarities and inter-class connections between lncRNAs, miRNAs, and diseases.

[0007] However, traditional GNN-based methods still have limitations in fully capturing spatial constraints between related nodes, specifically, they do not always enforce the proximity of biologically related entities, which can hinder the quality of learned representations. How to overcome this limitation is a technical problem that needs to be solved. SUMMARY

[0008] To solve the problems existing in the prior art, the present application provides a drug-lncRNA relationship prediction method and system based on embedding constraints, aiming to solve the shortcomings of existing methods in exploring the geometric relationship between molecular entities and improve the prediction performance and discrimination ability of the model.

[0009] To achieve the above-mentioned purpose, the present application provides the following solutions:

[0010] A drug-lncRNA relationship prediction method based on embedding constraints, the method comprising:

[0011] S1, collect lncRNA-drug association data and preprocess to construct a dataset;

[0012] S2, based on the dataset, construct an lncRNA-drug association matrix, and extract an lncRNA-drug bipartite graph;

[0013] S3, based on the lncRNA-drug association matrix, obtain an lncRNA similarity matrix and a drug similarity matrix, and perform dimensionality reduction processing on the lncRNA similarity matrix and the drug similarity matrix;

[0014] S4, input the lncRNA-drug bipartite graph and the lncRNA similarity matrix and the drug similarity matrix after dimensionality reduction processing into an LDA-GNN model and an LDA-DHG model with embedding constraint strategy, and obtain probability scores of lncRNA and drug association;

[0015] S5, based on the probability scores, realize drug-lncRNA relationship prediction.

[0016] Preferably, the S3 comprises:

[0017] G l (l i ,l j )=exp(-r l ||Y(l i )-Y(l j )|| 2 )

[0018] G d (d i ,d j )=exp(-r d ||Y(d i )-Y(d j )|| 2 )

[0019] Wherein, G l (l i ,l j ) and G d (d i ,d j ) are the Gaussian similarity of lncRNA l i and lncRNA l j and the Gaussian similarity of drug d i and drug d j , r l and r d are the scale functions of the Gaussian kernel function in the control of lncRNA and drug similarity, exp(·) is the natural exponential function, Y(l i), Y(l j ), Y(d i ) and Y(d j () are lncRNA l i lncRNAl j Drug d i and drug d j The eigenvector representation, ||Y(l i )-Y(l j )|| 2 Indicates lncRNAl i With lncRNAl j The squared Euclidean distance between the eigenvectors; ||Y(d i )-Y(d j )|| 2 Indicates drug d i With drug d j The squared Euclidean distance between the eigenvectors.

[0020] Preferably, S4 includes:

[0021]

[0022] in, This represents the output feature of the l-th layer. W represents the output feature of the (l-1)th layer. (l) Let b be the weight matrix of the l-th layer. (l) is the bias term of the l-th layer, and ReLU(·) is the activation function.

[0023] Preferably, step S4 further includes: training the LDA-GNN model and the LDA-DHG model with embedding constraint strategies.

[0024] The lncRNA-drug association data, the dimensionality-reduced lncRNA similarity matrix, and the drug similarity matrix are used as edges and nodes of a graph structure and input into an LDA-GNN model constructed based on GAT and GCN. The node update method is as follows:

[0025]

[0026] in, This represents the feature vector of node j in layer 0 of the LDA-GNN model. Let d(i) represent the feature vector of node i in the first layer of the LDA-GNN model, d(i) represent the degree of node i, d(j) represent the degree of node j, N(i) represent the set containing all neighbors of node i, and w 1 This represents the parameter matrix that GCN needs to learn;

[0027] The lncRNA-drug association data, the lncRNA similarity matrix and the drug similarity matrix after dimension reduction are input into the LDA-DHG model constructed based on GAT and GCN as edges and nodes of a graph structure, and the node update mode is:

[0028]

[0029] wherein, represents the feature of node i in the lth layer of the LDA-DHG model, represents the feature of node i in the (l+1)th layer of the LDA-DHG model, Agg is an aggregation operation, and MLP is a multilayer perceptron for feature updating.

[0030] An attention mechanism is introduced to calculate the attention coefficient between nodes:

[0031]

[0032] wherein, e ij is the attention coefficient of node i to neighbor node j, is the feature representation of node i in the kth layer, is the feature representation of node j in the kth layer, and a is a shared attention mechanism for calculating the attention coefficient; W (k) represents the parameter matrix to be learned in the kth layer, and σ represents an activation function.

[0033] The attention coefficient is normalized using a Softmax function:

[0034]

[0035] wherein, a ij is the normalized value of the attention coefficient, e im is the attention coefficient of node i to all its neighbor nodes m∈N i .

[0036] Weighted summation:

[0037]

[0038] Preferably, the S4 further comprises:

[0039]

[0040] wherein, y i represents the true label of the ith sample, y i ' represents the predicted value of the ith sample, x a represents the embedding vector of the ath association pair, and y bThe embedding vector representing the bth correlation pair, and β are hyperparameters, N is different classification results, and n is the dimension of each vector.

[0041] The application also provides a drug-lncRNA relationship prediction system based on embedding constraints, which is used to realize the drug-lncRNA relationship prediction method based on embedding constraints.

[0042] The acquisition module is used for acquiring lncRNA-drug correlation data and preprocessing, and constructing a data set.

[0043] The bipartite graph extraction module is used for constructing an lncRNA-drug correlation matrix based on the data set, and extracting an lncRNA-drug bipartite graph.

[0044] The similarity matrix acquisition module is used for acquiring an lncRNA similarity matrix and a drug similarity matrix based on the lncRNA-drug correlation matrix, and performing dimension reduction processing on the lncRNA similarity matrix and the drug similarity matrix.

[0045] The probability acquisition module is used for inputting the lncRNA-drug bipartite graph and the lncRNA similarity matrix and the drug similarity matrix after dimension reduction processing into LDA-GNN models and LDA-DHG models with embedding constraint strategies, and acquiring probability scores of lncRNA and drug correlation.

[0046] The prediction module is used for realizing drug-lncRNA relationship prediction based on the probability scores.

[0047] Preferably, the similarity matrix acquisition module comprises:

[0048] G l (l i ,l j )=exp(-r l ||Y(l i )-Y(l j )|| 2 )

[0049] G d (d i ,d j )=exp(-r d ||Y(d i )-Y(d j )|| 2 )

[0050] Wherein, G l (li , j ) and G d (d i , d j ) are the Gaussian similarities of lncRNA l i and lncRNA l j , and drug d i and drug d j , respectively, r l and r d are the scale functions controlling the Gaussian kernel functions in lncRNA and drug similarities, respectively, exp(·) is the natural exponential function, Y(l i ), Y(l j ), Y(d i ), and Y(d j ) are the feature vector representations of lncRNA l i , lncRNA l j , drug d i , and drug d j , respectively, ||Y(l i )-Y(l j )||2 2 represents the squared Euclidean distance between the feature vectors of lncRNA l i and lncRNA l j ; ||Y(d i )-Y(d j )||2 2 represents the squared Euclidean distance between the feature vectors of drug d i and drug d j .

[0051] Preferably, the probability obtaining module comprises:

[0052]

[0053] wherein, represents the output feature of the lth layer, represents the output feature of the (l-1)th layer, W (l) is the weight matrix of the lth layer, b (l) is the bias term of the lth layer, and ReLU(·) is an activation function.

[0054] Preferably, the probability obtaining module further comprises: training the LDA-GNN model and the LDA-DHG model with an embedding constraint strategy:

[0055] The lncRNA-drug association data, the dimensionality-reduced lncRNA similarity matrix, and the drug similarity matrix are used as edges and nodes of a graph structure and input into an LDA-GNN model constructed based on GAT and GCN. The node update method is as follows:

[0056]

[0057] in, This represents the feature vector of node j in layer 0 of the LDA-GNN model. Let d(i) represent the feature vector of node i in the first layer of the LDA-GNN model, d(i) represent the degree of node i, d(j) represent the degree of node j, N(i) represent the set containing all neighbors of node i, and w 1 This represents the parameter matrix that GCN needs to learn;

[0058] The lncRNA-drug association data, the dimensionality-reduced lncRNA similarity matrix, and the drug similarity matrix are used as edges and nodes of a graph structure and input into an LDA-DHG model constructed based on GAT and GCN. The node update method is as follows:

[0059]

[0060] in, This represents the feature of node i in the l-th layer of the LDA-DHG model. This represents the feature of node i in the (l+1)th layer of the LDA-DHG model. Agg is the aggregation operation, and MLP is a multilayer perceptron used for feature updating.

[0061] Introducing an attention mechanism to calculate the attention coefficient between nodes:

[0062]

[0063] Among them, e ij Let be the attention coefficient of node i to its neighbor node j. Let i be the feature representation of the k-th layer node. Let W be the feature representation of node j in layer k, and let a be the shared attention mechanism for calculating the attention coefficient; (k) Let σ represent the parameter matrix that needs to be learned in the k-th layer, and let σ represent the activation function.

[0064] The attention coefficient is normalized using the Softmax function:

[0065]

[0066] Among them, a ij To note the normalized value of the coefficients, e imThe attention coefficient of the node i to all its neighbor nodes m∈N i ;

[0067] Weighted summation:

[0068]

[0069] Preferably, the probability acquisition module further comprises:

[0070]

[0071] Wherein, y i represents the true label of the i-th sample, y i ' represents the predicted value of the i-th sample, x a represents the embedding vector of the a-th associated pair, y b represents the embedding vector of the b-th associated pair, and α and β are hyperparameters, N is the different classification results, and n is the dimension of each vector.

[0072] Compared with the prior art, the present application has the following beneficial effects:

[0073] The present application introduces an embedding constraint strategy, which solves the shortcomings of existing methods in exploring the geometric relationship between molecular entities. By ensuring the proximity of related nodes in the embedding space, the prediction performance and discrimination ability of the model are significantly improved, and the quality of the learned representation is improved. By introducing the embedding constraint strategy, the present application has achieved significant performance improvement in the LDA-GNN and LDA-DHG two graph neural network models, and has shown higher AUC, AUPR and Pre@k indicators in the LDRA prediction task. This distance constraint ensures that the confirmed molecular entities such as LncRNA, drugs and targets are closer in the feature space, so as to realize more accurate prediction.

[0074] Experimental results show that the embedding constraint strategy has good robustness under different data scales, and is suitable for small and large scale data sets, and has strong potential for wide application. The present application optimizes the embedding space, which helps to accurately distinguish positive and negative samples, thereby improving the biological interpretability of the prediction results, supporting more in-depth molecular pathological analysis and target identification, and helping more effective drug discovery and disease research. BRIEF DESCRIPTION OF DRAWINGS

[0075] In order to more clearly illustrate the technical solutions of the present application, the following briefly introduces the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0076] Figure 1A flow chart of a drug-lncRNA relationship prediction method based on embedding constraint for an embodiment of the present application is shown in the figure.

[0077] Figure 2 A line chart of two models in the LDRAs prediction task after introducing the embedding constraint strategy for an embodiment of the present application is shown in the figure, wherein (a) and (b) are line charts of AUC and AUPR values of the two models on data sets 1, 2 and 3, respectively.

[0078] Figure 3 A sample distance distribution diagram of two models in the LDRAs prediction task after introducing the embedding constraint for an embodiment of the present application is shown in the figure.

[0079] Figure 4 An embedding similarity diagram of a pair of associations of two models in the LDRAs prediction task after introducing the embedding constraint for an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0080] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0081] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0082] Embodiment One

[0083] As shown in the figure, the present application provides a drug-lncRNA relationship prediction method based on embedding constraint, comprising: Figure 1 S1, collect lncRNA-drug association data and pre-process, and construct a data set;

[0084] S2, based on the data set, construct an lncRNA-drug association matrix, and extract an lncRNA-drug bipartite graph;

[0085] S3, based on the lncRNA-drug association matrix, obtain an lncRNA similarity matrix and a drug similarity matrix, and perform dimensionality reduction processing on the lncRNA similarity matrix and the drug similarity matrix;

[0086]

[0087] ​S4. Input the lncRNA-drug bipartite graph and the dimension-reduced lncRNA similarity matrix and drug similarity matrix into the LDA-GNN model and LDA-DHG model with embedding constraint strategy to obtain the probability scores of lncRNA and drug association.

[0088] S5. Based on probability scores, predict the relationship between drugs and lncRNAs.

[0089] The following explanation uses a specific LncRNA-drug association prediction task as an example:

[0090] Input / output data description:

[0091] Download the lncRNA-drug association datasets from http: / / www.jianglab.cn / D-lnc / index.jsp, and remove duplicate associations. The Validated dataset contains 4691 lncRNAs, 48 ​​drugs, and 4791 associations; the cMap dataset contains 129 lncRNAs, 1279 drugs, and 15804 associations; and the GEO dataset contains 2360 lncRNAs, 115 drugs, and 28487 associations.

[0092] (1) L={l1,l2,...,l m} represents the lncRNA dataset, where D = {d1, d2, ..., d} n} represents the drug dataset, m represents the number of LncRNAs in dataset L, and n represents the number of drugs in dataset D;

[0093] (2) Construct an association matrix to represent the association information between LncRNA and drugs, Y ij =0 indicates lncRNA l i and drug d j Relationship unknown, Y ij =1 indicates lncRNAl i and drug d j There is a correlation. Known correlations are treated as positive samples, and an equal number of unknown correlations are treated as negative samples.

[0094] (3) Based on the lncRNA-drug association matrix G=(U,V,E), extract the lncRNA-drug bipartite graph, where U={l1,l2,....,l m} represents the set of lncRNA nodes, V = {d1, d2, ..., d...} n Let} be the set of drug nodes, and let E be the set of edges consisting of all node pairs satisfying A(i,j)=1, representing lncRNAl i and drug d jThere is an undirected edge between them;

[0095] (4) According to the lncRNA-drug association matrix, the Gaussian interaction for feature kernel function widely used in calculating vector similarity is used to calculate the lncRNA similarity matrix and the drug similarity matrix, respectively, wherein,

[0096] The similarity calculation formula of lncRNA is:

[0097] G l (l i ,l j )=exp(-r l ||Y(l i )-Y(l j )|| 2 )

[0098] The similarity calculation formula of drug is:

[0099] G d (d i ,d j )=exp(-r d ||Y(d i )-Y(d j )|| 2 )

[0100] Wherein, G l (l i ,l j ) and G d (d i ,d j ) are the Gaussian similarity of lncRNAl i and lncRNAl j and the Gaussian similarity of drug d i and drug d j , r l and r d are the scale functions of the Gaussian kernel function in the control of lncRNA and drug similarity, used to control the "smoothness" of the similarity function; exp(·) is the natural exponential function, Y(l i ), Y(l j ), Y(d i ) and Y(d j ) are the feature vector representations of lncRNAl i , lncRNAl j , drug d i and drug d j , respectively, and ||Y(l i )-Y(l j )|| 2 represents the lncRNAli Euclidean distance between the feature vectors of drug d j i Euclidean distance between the feature vectors of drug d j 2 Euclidean distance between the feature vectors of drug d i j Euclidean distance between the feature vectors of drug d

[0101] For the LDA-DHG model, two methods of global ranking and local ranking are used to construct the similarity matrix between nodes. Global ranking:

[0102]

[0103] Global ranking ranks all elements of the similarity matrix and calculates the percentile value of each element, where N is the total number of nodes, and is the indicator function, where the similarity S(i,j) is the similarity value between node i and node j, and S(i,k) is the similarity value between node i and other nodes k∈{1,N};

[0104] Local ranking:

[0105]

[0106] Local ranking ranks each element in each row and calculates the relative ranking of each element in the row, where M is the number of elements in each row of the matrix.

[0107] (5) The row vectors of the lncRNA similarity matrix are taken as the feature vectors of lncRNA, and the row vectors of the drug similarity matrix are taken as the feature vectors of drug, and PCA dimension reduction processing is performed thereon;

[0108] (6) Input the lncRNA-drug bipartite graph, the dimension-reduced lncRNA similarity matrix, and the drug similarity matrix;

[0109] (7) Input the lncRNA feature h i and the drug feature h j obtained by neural network training into the deep neural network, and the deep neural network includes L layers, and the calculation process of each layer is:

[0110]

[0111] wherein, i.e. the initial input is the spliced feature of the node pair, represents the output feature of the lth layer, represents the output feature of the (l-1)th layer, W (l) is the weight matrix of the lth layer, and b (l) ​​​Bias term for the first layer, ReLU(·) is the activation function, and the number of layers is set to 3. After the neural network transformation, the input is output by the activation function to obtain the probability score of the lncRNA and drug association.

[0112] Model training:

[0113] In order to prove the generalization ability of the embedding constraint strategy, the strategy is introduced into the currently commonly used model for drug-lncRNA relationship prediction. Since there are few existing methods in this field, the present invention selects two models with better performance, LDA-GNN based on graph attention network (GAT) and graph convolution network (GCN), and dynamic heterogeneous graph LDA-DHG based on lncRNA-lncRNA and drug-drug similarity network construction. The embedding constraint strategy is introduced into the existing two GNN models for predicting drug-lncRNA relationship, specifically by designing a regularization method when updating the node embedding of the network, so that the embedding vectors of the node pairs known to have a relationship are closer. The following is a detailed introduction to the network architecture of the two models:

[0114] (1) LDA-GNN model is composed of one layer of graph convolution network (GCN) and multiple layers of graph attention layer (GAT). One layer of GCN is used to aggregate the information of neighboring nodes to update its own features. The lncRNA-drug association data, the lncRNA similarity matrix and the drug similarity matrix after dimensionality reduction are input into the LDA-GNN model as the edges and nodes of the graph structure, and the node update method is:

[0115]

[0116] where σ represents the activation function, represents the feature vector of node j in the 0th layer (input layer) of the LDA-GNN model, represents the feature vector of node i in the 1st layer (input layer) of the LDA-GNN model, d(i) represents the degree of node i, d(j) represents the degree of node j, N(i) represents the set containing all neighbors of node i, and w 1 represents the parameter matrix that GCN needs to learn.

[0117] (2) LDA-DHG model is based on GAT and GCN network, where the features of nodes are updated by aggregating the features of neighboring nodes. The lncRNA-drug association data, the lncRNA similarity matrix and the drug similarity matrix after dimensionality reduction are input into the LDA-DHG model as the edges and nodes of the graph structure, and the node update method is:

[0118]

[0119] where, represents the feature of node i in the l-th layer in the LDA-DHG model, represents the feature of node i in the l+1-th layer in the LDA-DHG model, N(i) represents the set containing all neighbors of node i, Agg is the aggregation operation, and MLP is the multi-layer perception used for feature updating.

[0120] (3) GAT, although similar to GCN, updates the node features by aggregating the information of neighboring nodes, but GAT allows different weights to be assigned to different neighboring nodes in the aggregation process, and introduces an attention mechanism, so that more important nodes get more attention. Taking node i and node j as an example, the attention coefficient between them is calculated:

[0121]

[0122] where, e ij is the attention coefficient of node i to neighbor node j, is the feature representation of node i in the k-th layer, is the feature representation of node j in the k-th layer, a is the shared attention mechanism for calculating the attention coefficient, which is initialized randomly and the parameters are updated through training, W (k) represents the parameter matrix to be learned in the k-th layer, and σ represents the activation function, and LeakyReLU is selected as the activation function.

[0123] (4) In order to compare the attention coefficients between different nodes, the attention coefficients are normalized using the Softmax function:

[0124]

[0125] where, a ij is the normalized value of the attention coefficient, e ij is the attention coefficient of node i to neighbor node j, e im is the attention coefficient of node i to all its neighbor nodes m∈N i , which is used for softmax normalization.

[0126] After calculating the attention coefficients between node i and its neighbor nodes, different weights can be assigned to the neighbor nodes of node i, and the weighted sum of the neighbor nodes is calculated as the input of the next layer:

[0127]

[0128] Distance constraint strategy:

[0129] The present application is improved based on the existing distance constraint loss strategy, by utilizing the known relationship of the associated pairs in the data set, the proximity between the biological related nodes in the feature space is explicitly enhanced, so as to improve the accuracy and generalization ability of drug-lncRNA relationship prediction. In the feature updating process of LDA-GNN model and LDA-DHG model, the distance constraint strategy is introduced respectively, the associated pairs which have been proved to exist relationship are set as positive samples, and other associated pairs are set as negative samples. By reducing the distance between the embedding vectors of the positive samples in the training process, the effective separation of the positive and negative samples in the feature space is promoted. In the embedding vector updating process, the feature representation of the lncRNA and drug nodes with known associated relationship will become more similar, so as to further optimize the feature distribution and enhance the biological interpretability and prediction performance of the model. The distance constraint ensures that the molecular entities of lncRNA and drug are closer in the feature space, specifically in the loss function:

[0130]

[0131] Wherein, y i represents the true label of the i th sample, y i ' represents the predicted value of the i th sample, x a represents the embedding vector of the a th associated pair, y b represents the embedding vector of the b th associated pair, alpha and beta are hyperparameters, N is different classification results, that is, whether there is an association, and n is the dimension of each vector.

[0132] Different hyperparameters will affect the prediction performance of the model, in order to achieve the best performance, the present application tries different hyperparameters, and finally selects a group of hyperparameters with the best prediction performance of the model, which are alpha 0.5, beta 0.5, dropout 0.2, epoch 40, batchsize 32 and learning rate 0.001.

[0133] The effect of the present application is further illustrated by experiment, the prediction performance of the five-fold cross-validation algorithm is as shown in the following table. Figures 2-4

[0134] In summary, the embedding constraint strategy is introduced in the present application, which solves the shortcomings of the existing method in exploring the geometric relationship between molecular entities. By ensuring the proximity of related nodes in the embedding space, the prediction performance and discrimination ability of the model are significantly improved. By introducing the embedding constraint strategy, the present application has achieved significant performance improvement in the LDA-GNN and LDA-DHG two kinds of graph neural network models, and has shown higher AUC, AUPR and Pre@k indexes in the LDRA prediction task.

[0135] ​The experimental results show that the embedding constraint strategy has good robustness under different data scales, is suitable for small-scale and large-scale data sets, and has strong application potential.

[0136] Embodiment two

[0137] The application also provides an embedding constraint-based drug-lncRNA relationship prediction system for implementing the embedding constraint-based drug-lncRNA relationship prediction method of the foregoing embodiments, and the system comprises a collection module, a two-part graph extraction module, a similarity matrix acquisition module, a probability acquisition module and a prediction module.

[0138] The collection module is used for collecting lncRNA-drug association data and pre-processing, and constructing a data set.

[0139] The two-part graph extraction module is used for constructing an lncRNA-drug association matrix based on the data set, and extracting an lncRNA-drug two-part graph.

[0140] The similarity matrix acquisition module is used for acquiring an lncRNA similarity matrix and a drug similarity matrix based on the lncRNA-drug association matrix, and performing dimension reduction processing on the lncRNA similarity matrix and the drug similarity matrix.

[0141] The probability acquisition module is used for inputting the lncRNA-drug two-part graph and the dimension-reduced lncRNA similarity matrix and drug similarity matrix into LDA-GNN models and LDA-DHG models with embedding constraint strategies, and acquiring probability scores of lncRNA and drug association.

[0142] The prediction module is used for realizing drug-lncRNA relationship prediction based on the probability scores.

[0143] Further, the similarity matrix acquisition module comprises:

[0144] G l (l i ,l j )=exp(-r l ||Y(l i )-Y(l j )|| 2 )

[0145] G d (d i ,d j )=exp(-r d ||Y(d i )-Y(dj )|| 2 )

[0146] where G l (l i ,l j ) and G d (d i ,d j ) are the Gaussian similarities between lncRNA l i and lncRNA l j and drug d i and drug d j , respectively, r l and r d are the scale functions controlling the Gaussian kernel functions in lncRNA and drug similarities, respectively, exp(·) is the natural exponential function, Y(l i ), Y(l j ), Y(d i ) and Y(d j ) are the feature vector representations of lncRNA l i , lncRNA l j , drug d i and drug d j , respectively, ||Y(l i )-Y(l j )||2 2 represents the squared Euclidean distance between the feature vectors of lncRNA l i and lncRNA l j ; and ||Y(d i )-Y(d j )||2 2 represents the squared Euclidean distance between the feature vectors of drug d i and drug d j .

[0147] The probability obtaining module comprises:

[0148]

[0149] wherein, x represents the output feature of the lth layer, x represents the output feature of the (l-1)th layer, W (l) is the weight matrix of the lth layer, b (l) is the bias term of the lth layer, and ReLU(·) is an activation function.

[0150] The probability obtaining module further comprises: training the LDA-GNN model and the LDA-DHG model embedded with the constraint strategy:

[0151] The lncRNA-drug association data, the lncRNA similarity matrix and the drug similarity matrix after dimension reduction are input into the LDA-GNN model constructed based on GAT and GCN as edges and nodes of a graph structure, and the node update mode is as follows:

[0152]

[0153] wherein, represents the feature vector of node j in the 0th layer of the LDA-GNN model, represents the feature vector of node i in the 1st layer of the LDA-GNN model, d(i) represents the degree of node i, d(j) represents the degree of node j, N(i) represents a set containing all neighbors of node i, w 1 represents a parameter matrix to be learned by GCN;

[0154] The lncRNA-drug association data, the lncRNA similarity matrix and the drug similarity matrix after dimension reduction are input into the LDA-DHG model constructed based on GAT and GCN as edges and nodes of a graph structure, and the node update mode is as follows:

[0155]

[0156] wherein, represents the feature of node i in the lth layer of the LDA-DHG model, represents the feature of node i in the (l+1)th layer of the LDA-DHG model, Agg is an aggregation operation, and MLP is a multilayer perceptron used for feature updating;

[0157] The attention mechanism is introduced to calculate the attention coefficient between nodes:

[0158]

[0159] wherein, e ij is the attention coefficient of node i to neighbor node j, is the feature representation of node i in the kth layer, is the feature representation of node j in the kth layer, a is a shared attention mechanism for calculating the attention coefficient; W (k) represents a parameter matrix to be learned in the kth layer, and σ represents an activation function;

[0160] The attention coefficient is normalized by using a Softmax function:

[0161]

[0162] wherein, a ij is the normalized value of the attention coefficient, e im is the attention coefficient of node i to all neighbor nodes m∈Ni a coefficient of attention of the i-th sample;

[0163] a weighted sum:

[0164]

[0165] The probability acquisition module further comprises:

[0166]

[0167] wherein y i represents a true label of the i-th sample, y i represents a predicted value of the i-th sample, x a represents an embedding vector of the a-th associated pair, y b represents an embedding vector of the b-th associated pair, and α and β are hyperparameters, N is a different classification result, and n is a dimension of each vector.

[0168] The above-described embodiments are merely intended to describe the preferred modes of the present application, and are not intended to limit the scope of the present application. Various modifications and improvements to the technical solutions of the present application made by those of ordinary skill in the art without departing from the design spirit of the present application shall fall within the scope of protection of the present application as defined by the claims.

Claims

1. A method for predicting drug-lncRNA relationships based on embedding constraints, characterized in that, The method includes: S1. Collect lncRNA-drug association data and preprocess them to construct a dataset; S2. Based on the dataset, construct the lncRNA-drug association matrix and extract the lncRNA-drug bipartite graph; S3. Based on the lncRNA-drug association matrix, obtain the lncRNA similarity matrix and the drug similarity matrix, and perform dimensionality reduction processing on the lncRNA similarity matrix and the drug similarity matrix; S4. Input the lncRNA-drug bipartite graph, the dimension-reduced lncRNA similarity matrix, and the drug similarity matrix into the LDA-GNN model and the LDA-DHG model with embedding constraint strategy to obtain the probability scores of lncRNA and drug association. S5. Based on the probability scores, predict the relationship between drugs and lncRNAs; S4 includes: lncRNA features obtained through neural network training h i and drug characteristics h j The input to the deep neural network consists of L layers, and the computation process of each layer is as follows: in, That is, the initial input is the concatenated features of node pairs. This represents the output feature of the l-th layer. W represents the output feature of the (l-1)th layer. (l) Let b be the weight matrix of the l-th layer. (l) The term is the bias term of the l-th layer, ReLU(·) is the activation function, and the number of layers is set to 3. After the neural network transformation, the input activation function outputs the probability score of the association between lncRNA and drug.

2. The drug-lncRNA relationship prediction method based on embedding constraints according to claim 1, characterized in that, S3 includes: G l (l i ,l j )=exp(-r l ||Y(l i )-Y(l j )|| 2 ); G d (d i ,d j )=exp(-r d ||Y(d i )-Y(d j )|| 2 ); Among them, G l (l i ,l j ) and G d (d i ,d j () are lncRNAl i With lncRNAl j Gaussian similarity and drug d i With drug d j Gaussian similarity, r l and r d These are the scaling functions controlling the Gaussian kernel function in lncRNA and drug similarity, respectively, exp(·) is the natural exponential function, and Y(l i ), Y(l j ), Y(d i ) and Y(d j () are lncRNAl i lncRNAl j Drug d i and drug d j The eigenvector representation, ||Y(l i )-Y(l j )|| 2 Indicates lncRNAl i With lncRNAl j The squared Euclidean distance between the eigenvectors; ||Y(d i )-Y(d j )|| 2 Indicates drug d i With drug d j The squared Euclidean distance between the eigenvectors.

3. The drug-lncRNA relationship prediction method based on embedding constraints according to claim 1, characterized in that, S4 further includes: training the LDA-GNN model and the LDA-DHG model with embedding constraint policies. The lncRNA-drug association data, the dimensionality-reduced lncRNA similarity matrix, and the drug similarity matrix are used as edges and nodes of a graph structure and input into an LDA-GNN model constructed based on GAT and GCN. The node update method is as follows: in, This represents the feature vector of node j in layer 0 of the LDA-GNN model. Let d(i) represent the feature vector of node i in the first layer of the LDA-GNN model, d(i) represent the degree of node i, d(j) represent the degree of node j, N(i) represent the set containing all neighbors of node i, and w 1 This represents the parameter matrix that GCN needs to learn; The lncRNA-drug association data, the dimensionality-reduced lncRNA similarity matrix, and the drug similarity matrix are used as edges and nodes of a graph structure and input into an LDA-DHG model constructed based on GAT and GCN. The node update method is as follows: in, This represents the feature of node i in the l-th layer of the LDA-DHG model. This represents the feature of node i in the (l+1)th layer of the LDA-DHG model. Agg is the aggregation operation, and MLP is a multilayer perceptron used for feature updating. Introducing an attention mechanism to calculate the attention coefficient between nodes: Among them, e ij Let be the attention coefficient of node i to its neighbor node j. Let i be the feature representation of the k-th layer node. Let W be the feature representation of node j in layer k, and let a be the shared attention mechanism for calculating the attention coefficient; (k) Let σ represent the parameter matrix that needs to be learned in the k-th layer, and let σ represent the activation function. The attention coefficient is normalized using the Softmax function: Among them, a ij To note the normalized value of the coefficients, e im For node i, consider all its neighboring nodes m∈N i Attention coefficient; Weighted summation:

4. The drug-lncRNA relationship prediction method based on embedding constraints according to claim 3, characterized in that, S4 further includes: Among them, y i Let y represent the true label of the i-th sample. i ' represents the predicted value of the i-th sample, x a Let y represent the embedding vector of the a-th association pair. b Let represent the embedding vector of the b-th association pair, where α and β are hyperparameters, N represents the different classification results, and n is the dimension of each vector.

5. A drug-lncRNA relationship prediction system based on embedding constraints, used to implement the drug-lncRNA relationship prediction method based on embedding constraints according to any one of claims 1-4, characterized in that, The system includes: an acquisition module, a bipartite graph extraction module, a similarity matrix acquisition module, a probability acquisition module, and a prediction module; The acquisition module is used to acquire lncRNA-drug association data and preprocess it to construct a dataset; The bipartite graph extraction module is used to construct an lncRNA-drug association matrix and extract an lncRNA-drug bipartite graph based on the dataset. The similarity matrix acquisition module is used to acquire the lncRNA similarity matrix and the drug similarity matrix based on the lncRNA-drug association matrix, and to perform dimensionality reduction processing on the lncRNA similarity matrix and the drug similarity matrix; The probability acquisition module is used to input the lncRNA-drug bipartite graph and the dimension-reduced lncRNA similarity matrix and the drug similarity matrix into the LDA-GNN model and the LDA-DHG model with embedding constraint strategy to obtain the probability score of lncRNA and drug association. The prediction module is used to predict the relationship between drugs and lncRNAs based on the probability scores. The probability acquisition module includes: lncRNA features obtained through neural network training h i and drug characteristics h j The input to the deep neural network consists of L layers, and the computation process of each layer is as follows: in, That is, the initial input is the concatenated features of node pairs. This represents the output feature of the l-th layer. W represents the output feature of the (l-1)th layer. (l) Let b be the weight matrix of the l-th layer. (l) The term is the bias term of the l-th layer, ReLU(·) is the activation function, and the number of layers is set to 3. After the neural network transformation, the input activation function outputs the probability score of the association between lncRNA and drug.

6. The drug-lncRNA relationship prediction system based on embedding constraints according to claim 5, characterized in that, The similarity matrix acquisition module includes: G l (l i ,l j )=exp(-r l ||Y(l i )-Y(l j )|| 2 ); G d (d i ,d j )=exp(-r d ||Y(d i )-Y(d j )|| 2 ); Among them, G l (l i ,l j ) and G d (d i ,d j () are lncRNAl i With lncRNAl j Gaussian similarity and drug d i With drug d j Gaussian similarity, r l and r d These are the scaling functions controlling the Gaussian kernel function in lncRNA and drug similarity, respectively, exp(·) is the natural exponential function, and Y(l i ), Y(l j ), Y(d i ) and Y(d j () are lncRNAl i lncRNAl j Drug d i and drug d j The eigenvector representation, ||Y(l i )-Y(l j )|| 2 Indicates lncRNAl i With lncRNA l j The squared Euclidean distance between the eigenvectors; ||Y(d i )-Y(d j )|| 2 Indicates drug di and drug d j The squared Euclidean distance between the eigenvectors.

7. The drug-lncRNA relationship prediction system based on embedding constraints according to claim 5, characterized in that, The probability acquisition module further includes training the LDA-GNN model and the LDA-DHG model with embedding constraint policies. The lncRNA-drug association data, the dimensionality-reduced lncRNA similarity matrix, and the drug similarity matrix are used as edges and nodes of a graph structure and input into an LDA-GNN model constructed based on GAT and GCN. The node update method is as follows: in, This represents the feature vector of node j in layer 0 of the LDA-GNN model. Let d(i) represent the feature vector of node i in the first layer of the LDA-GNN model, d(i) represent the degree of node i, d(j) represent the degree of node j, N(i) represent the set containing all neighbors of node i, and w 1 This represents the parameter matrix that GCN needs to learn; The lncRNA-drug association data, the dimensionality-reduced lncRNA similarity matrix, and the drug similarity matrix are used as edges and nodes of a graph structure and input into an LDA-DHG model constructed based on GAT and GCN. The node update method is as follows: in, This represents the feature of node i in the l-th layer of the LDA-DHG model. This represents the feature of node i in the (l+1)th layer of the LDA-DHG model. Agg is the aggregation operation, and MLP is a multilayer perceptron used for feature updating. Introducing an attention mechanism to calculate the attention coefficient between nodes: Among them, e ij Let be the attention coefficient of node i to its neighbor node j. Let i be the feature representation of the k-th layer node. Let W be the feature representation of node j in layer k, and let a be the shared attention mechanism for calculating the attention coefficient; (k) Let σ represent the parameter matrix that needs to be learned in the k-th layer, and let σ represent the activation function. The attention coefficient is normalized using the Softmax function: Among them, a ij To note the normalized value of the coefficients, e im For node i, consider all its neighboring nodes m∈N i Attention coefficient; Weighted summation:

8. The drug-lncRNA relationship prediction system based on embedding constraints according to claim 7, characterized in that, The probability acquisition module further includes: Among them, y i Let y represent the true label of the i-th sample. i ' represents the predicted value of the i-th sample, x a Let y represent the embedding vector of the a-th association pair. b Let represent the embedding vector of the b-th association pair, where α and β are hyperparameters, N represents the different classification results, and n is the dimension of each vector.

Images