Drug relocation method using neighborhood information and weighted fusion network

Abstract

The invention relates to a drug relocation method, in particular to a drug relocation method using neighborhood information and a weighted fusion network. The invention aims to solve the problems that the existing drug relocation method is a shallow prediction model, the characteristics of drugs and disease nodes cannot be deeply known, and the relocation result of the existing drug relocation method is biased to common infectious diseases. The method comprises the following steps: extracting a data sample of the action relationship of drugs on diseases; in the drug-to-disease effect relation graph, learning interaction characteristics between drugs and disease nearest neighbors by using a weighted graph convolutional network; aggregating the learned characteristics of the drugs and the diseases; hidden high-order nonlinear features are obtained through a multi-layer fusion network; inputting the action relationship of the drugs in the training set on the diseases into the model, training the model, and obtaining a final prediction model after multiple iterations; and inputting the drug pairs in the test set into the prediction model to obtain a prediction result. The invention belongs to the technical field of drug relocation.

Description

Technical Field

[0001] The invention relates to a drug relocation method and belongs to the technical field of drug relocation. Background Art

[0002] The development of new drugs is an expensive process. Due to inaccurate identification of drug targets and related reactions, only 10% of drugs that enter Phase I clinical trials are approved, and the rest of the drugs will not be used due to high toxicity or ineffectiveness. The high failure rate of drug clinical trials and the complexity of drug development have led to the continuous extension of the new drug development cycle. The average R&D cycle of a new drug is 8.1 years, and the average R&D cycle of most drugs is 4.6-11.9 years. The long cycle of drug development has led to an increase in drug development costs. Relevant studies have shown that the average total R&D cost of a new drug is $2-3 billion. Although researchers have invested a lot of time, money and material resources, the number of successfully developed new drugs has not increased effectively. According to relevant reports from the FDA, its Center for Drug Evaluation and Research (CDER) approved a total of 50 new therapies in 2021, including 36 new molecular entities and 14 biologics, which is lower than the number in 2020. In the past 10 years, CDER has approved only 24 drugs per year on average. It can be seen that the development of traditional drugs is expensive, time-consuming and difficult.

[0003] Drug repositioning, also known as new uses for old drugs, is a strategy used to discover old drugs or drugs under development that go beyond their originally approved indications and expand their scope of application and use. Drug repositioning technology can effectively shorten the drug development cycle, reduce research and development costs, and play an important role in promoting drug discovery. At present, drug migration based on deep learning can be roughly divided into four categories: migration methods based on known similarity associations, migration methods based on drug structures, and migration methods based on heterogeneous information.

[0004] Migration methods based on known similarity associations, the "similarity principle" refers to the fact that similar compounds tend to have similar biological activities. Therefore, similar drugs are more likely to be associated with the same disease. The first type of method relies on calculating the similarity between drugs and diseases. However, these methods are all shallow prediction models that cannot deeply understand the characteristics of drug and disease nodes, and the over-use of similarity also makes the repositioning results too biased towards common infectious diseases.

[0005] Drug structure-based transfer methods attempt to predict drug functions by learning the molecular structure of the drug. Such methods can model or represent drug structures well, but cannot fully utilize multi-source heterogeneous data, so their role in drug repositioning is limited.

[0006] Based on the migration method of heterogeneous information, in the knowledge graph composed of drugs and diseases, we can well calculate their similarities to identify effective therapeutic drugs for specific diseases. Previous studies have shown that drug-target networks, drug-drug networks, drug-disease networks and protein interaction networks play an important role in determining the relationship between drugs and diseases, providing new opportunities for drug discovery or drug repositioning. Summary of the invention

[0007] The present invention aims to solve the problem that the existing drug repositioning methods are shallow prediction models that cannot deeply understand the characteristics of drug and disease nodes, and their repositioning results tend to be common infectious diseases. A drug repositioning method using neighborhood information and weighted fusion network is proposed.

[0008] The technical solution adopted by the present invention to solve the above-mentioned problem is: the steps of the present invention include:

[0009] Step 1: Extract data samples of the relationship between drug effects and diseases, generate data sets, divide them into training sets and test sets, and perform network training and testing;

[0010] Step 2: In the drug-disease relationship graph, use the weighted graph convolutional network to learn the interaction characteristics between drugs and disease nearest neighbors;

[0011] Step 3: Aggregate the learned characteristics of drugs and diseases;

[0012] Step 4: Obtain hidden high-order nonlinear features through a multi-layer fusion network, and then combine them with the original features as new drug or disease features to predict drug repositioning;

[0013] Step 5: Input the drug-disease relationship in the training set into the above model, train the model, and obtain the final prediction model after multiple iterations;

[0014] Step 6: Input the drug pairs in the test set into the prediction model, obtain the prediction results, and perform result analysis.

[0015] Furthermore, step 1 specifically includes: randomly dividing the used data set into a training set and a test set in a ratio of 4:1, of which 80% is the training set and 20% is the test set; both the training set and the test set contain each action type; the action types are numbered from 0 to (n-1), where n represents the total number of categories, and the data set is subjected to a 5-fold cross-validation process.

[0016] Furthermore, step 2 specifically includes learning the interactions between drug nearest neighbors through a weighted linear aggregator, which expands the drug feature representation. The formula is as follows:

[0017]

[0018] In formula (1), r d is the feature representation of drug d, i and j are the nearest neighbors of drug d, ω is the weight matrix, W ij is the interaction weight between drug i and drug j; the weighted graph convolution operation aggregates the pairwise interactions between neighbors;

[0019] The characteristics of the acquired disease are expressed as:

[0020]

[0021] Furthermore, step 3 specifically includes: for a given drug-disease pair [d, v], using generalized matrix factorization to learn the association between the drug-disease pairs, the formula is as follows:

[0022] r d，v =F d ⊙F v (3)

[0023] In formula (3), r d，v Represents the association between drug-disease pairs, representing the embedding vectors of drug d and disease v, respectively, and aggregates drug and disease features:

[0024]

[0025]

[0026]

[0027] In formulas (4), (5) and (6), and are the representations of drugs and diseases obtained through the interactions between neighbors, and σ(·) is a nonlinear activation function.

[0028] Furthermore, step 4 specifically includes:

[0029] Step 401, multi-layer fusion network;

[0030] Design a multi-layer fusion network to capture hidden high-order nonlinear features and fuse them with the original embedded features to predict drug-disease associations. The formula is:

[0031] x 1+1 =αx0⊙(wx1+b)+(1-α)x1) (7),

[0032] h l+1 =f(w l h l +b l) (8),

[0033] In formulas (7) and (8), the initial input embeddings obtained from the encoder are x0, x1 and x 1+1 represents the input and output of the l+1 layer fusion network, w and b are the learnable weight matrix and bias; h l and h l+1 are the input and output of layer l, w l The weight matrix, b l is the bias vector, f(·) is the activation function of ReLU;

[0034] Step 402, deeply integrating the network;

[0035] The deep neural network and the multi-layer fusion network are combined. When combined, a parallel structure is adopted. The formula is as follows:

[0036] f fusion +f deep (9).

[0037] The beneficial effects of the present invention are as follows: the present invention proposes a migration method based on heterogeneous information, which can well calculate their similarities in a knowledge graph composed of drugs and diseases to identify effective therapeutic drugs for specific diseases; the present invention provides new opportunities for drug discovery or drug repositioning; the present invention expands the drug feature representation by learning the interactions between drug nearest neighbors through a weighted linear aggregator. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of the drug repositioning method using neighborhood information and weighted fusion network;

[0039] Figure 2 It is a framework diagram of the drug repositioning method using neighborhood information and weighted fusion network. DETAILED DESCRIPTION

[0040] Specific implementation method 1: Figure 1 and Figure 2 As shown, a drug repositioning method using neighborhood information and weighted fusion network includes the following specific steps:

[0041] Step 1, extracting drug-disease effect relationship data samples, generating a data set, dividing it into a training set and a test set, and performing network training and testing; randomly dividing the used data set into a training set and a test set at a ratio of 4:1, wherein 80% is a training set and 20% is a test set; both the training set and the test set contain each type of effect; the effect types are numbered from 0 to (n-1), where n represents the total number of categories, and the data set is subjected to a 5-fold cross validation process;

[0042] Step 2: In the drug-disease relationship graph, a weighted graph convolutional network is used to learn the interaction features between drugs and disease nearest neighbors; a weighted linear aggregator is used to learn the interaction between drug nearest neighbors and expand the drug feature representation. The formula is as follows:

[0043]

[0044] In formula (1), r d is the feature representation of drug d, i and j are the nearest neighbors of drug d, ω is the weight matrix, W ij is the interaction weight between drug i and drug j; the weighted graph convolution operation aggregates the pairwise interactions between neighbors;

[0045] The characteristics of the acquired disease are expressed as:

[0046]

[0047] Step 3: Aggregate the learned features of drugs and diseases; for a given drug-disease pair [d, v], use generalized matrix factorization to learn the association between drug-disease pairs, as follows:

[0048] r d，v =F d ⊙F v (3)

[0049] In formula (3), r d，v Represents the association between drug-disease pairs, representing the embedding vectors of drug d and disease v, respectively, and aggregates drug and disease features:

[0050]

[0051]

[0052]

[0053] In formulas (4), (5) and (6), and are the representations of drugs and diseases obtained through the interactions between neighbors, and σ(·) is a nonlinear activation function;

[0054] Step 4: Obtain hidden high-order nonlinear features through a multi-layer fusion network, and then combine them with the original features as new drug or disease features to predict drug repositioning;

[0055] Specifically include:

[0056] Step 401, multi-layer fusion network;

[0057] Design a multi-layer fusion network to capture hidden high-order nonlinear features and fuse them with the original embedded features to predict drug-disease associations. The formula is:

[0058] x 1+1 =ax0⊙(wx1+b)+(1-α)x1) (7),

[0059] h l+1 =f(w l h l +b l ) (8),

[0060] In formulas (7) and (8), the initial input embeddings obtained from the encoder are x0, x1 and x 1+1 represents the input and output of the l+1 layer fusion network, w and b are the learnable weight matrix and bias; h l and h l+1 are the input and output of layer 1, w l The weight matrix, b l is the bias vector, f(·) is the activation function of ReLU;

[0061] Step 402, deeply integrating the network;

[0062] The deep neural network and the multi-layer fusion network are combined. When combined, a parallel structure is adopted. The formula is as follows:

[0063] f fusion +f deep (9);

[0064] Step 5: Input the drug-disease relationship in the training set into the above model, train the model, and obtain the final prediction model after multiple iterations;

[0065] Step 6: Input the drug pairs in the test set into the prediction model, obtain the prediction results, and perform result analysis.

[0066] Dataset

[0067] The model is evaluated using three real-world datasets;

[0068] Fdataset includes 1933 drugs known in the Drug Library database, 593 drugs associated with diseases and 313 diseases from the OMIM database.

[0069] The Cdataset contains 2532 known drug-disease associations, including 663 drugs extracted from the Drugbank database and 409 diseases listed in the OMIM database.

[0070] LRSSL includes 3051 validated drug-disease associations, mainly including 763 drugs extracted from the drug library database and 681 diseases obtained from the MeSH database.

[0071] Table 1

[0072]

[0073]

[0074] Evaluation indicators

[0075] AUC and AUPR scores are used as evaluation indicators, and the calculation formula of the indicators is as follows:

[0076] True Positive Rate

[0077] TPR is used to measure the proportion of correctly classified positive samples to actual positive samples. The calculation formula is as follows:

[0078]

[0079] False Positive Rate

[0080] FPR is used to measure the proportion of misclassified positive samples to actual negative samples. The calculation formula is as follows:

[0081]

[0082] The definitions of TP, TN, FP, and FN are shown in Table 2.

[0083] Table 2

[0084]

[0085] AUC

[0086] AUC stands for Area under curve, which is the area under the ROC curve. ROC can reflect the classification ability. Its horizontal axis is the false positive rate (FPR) and the vertical axis is the true positive rate (TPR). The closer the AUC is to 1, the better the model result.

[0087] AUPR

[0088] AUPR stands for Area under Precision / Recall curve, which is the area under the PR curve. The horizontal axis of the PR curve is the recall rate, and the vertical axis is the precision rate. The PR curve is easily affected by the sample distribution (the ratio of positive and negative samples in the training sample), so AUPR can be used to measure the prediction performance of unbalanced data sets. The closer the AUPR value is to 1, the better the model performance.

[0089] The experimental results show that the present invention is compared with DRIMC, SCMFDD, NIMCGCN, BNNR and LAGCN. The results are shown in Table 3:

[0090] Table 3

[0091]

[0092] The results show that in Fdataset and Cdataset, the AUC obtained by BNNR is slightly higher than the model of the present invention. However, the AUPR of BNNR on all three data sets is lower than that of the model proposed by the present invention. Generally speaking, the unknown drug-disease pairs are much larger than the known drug-disease pairs, so the negative samples are much larger than the positive samples. For unbalanced data sets, AUPR is also an indispensable evaluation indicator. Compared with other models, the model of the present invention has better results.

[0093] The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement made to the above embodiments without departing from the content of the technical solution of the present invention, based on the technical essence of the present invention, within the spirit and principles of the present invention, still fall within the protection scope of the technical solution of the present invention.

Claims

1. A drug repositioning method using neighborhood information and weighted fusion network, characterized in that: The specific steps include: Step 1: Extract data samples of the relationship between drug effects and diseases, generate data sets, divide them into training sets and test sets, and perform network training and testing; Step 2: In the drug-disease relationship graph, use the weighted graph convolutional network to learn the interaction characteristics between drugs and disease nearest neighbors; Step 3: Aggregate the learned characteristics of drugs and diseases; Step 4: Obtain hidden high-order nonlinear features through a multi-layer fusion network, and then combine them with the original features as new drug or disease features to predict drug repositioning; Step 5: Input the drug-disease relationship in the training set into the above model, train the model, and obtain the final prediction model after multiple iterations; Step 6: Input the drug pairs in the test set into the prediction model, obtain the prediction results, and perform result analysis.

2. A drug repositioning method using neighborhood information and weighted fusion network according to claim 1, characterized in that: Step 1 specifically includes: randomly dividing the used data set into a training set and a test set in a ratio of 4:1, of which 80% is the training set and 20% is the test set; both the training set and the test set contain each action type; the action types are numbered from 0 to (n-1), where n represents the total number of categories, and the data set is subjected to a 5-fold cross-validation process.

3. The drug repositioning method using neighborhood information and weighted fusion network according to claim 1, characterized in that: Step 2 specifically involves learning the interactions between drug nearest neighbors through a weighted linear aggregator, which expands the drug feature representation. The formula is as follows: In formula (1), r d is the feature representation of drug d, i and j are the nearest neighbors of drug d, ω is the weight matrix, W ij is the interaction weight between drug i and drug j; the weighted graph convolution operation aggregates the pairwise interactions between neighbors; The characteristics of the acquired disease are expressed as:

4. The drug repositioning method using neighborhood information and weighted fusion network according to claim 1, characterized in that: Step 3 specifically includes: for a given drug-disease pair [d, v], use generalized matrix factorization to learn the association between the drug-disease pairs, the formula is as follows: r d,v =F d ⊙F v (3), In formula (3), r d,v Represents the association between drug-disease pairs, representing the embedding vectors of drug d and disease v, respectively, and aggregates drug and disease features: In formulas (4), (5) and (6), and are the representations of drugs and diseases obtained through the interactions between neighbors, and σ(·) is a nonlinear activation function.

5. The drug repositioning method using neighborhood information and weighted fusion network according to claim 1, characterized in that: Step 4 specifically includes: Step 401, multi-layer fusion network; Design a multi-layer fusion network to capture hidden high-order nonlinear features and fuse them with the original embedded features to predict drug-disease associations. The formula is: x 1+1 =αx0⊙(wx1+b)+(1-α)x1) (7), h l+1 =f(w l h l +b l ) (8), In formulas (7) and (8), the initial input embeddings obtained from the encoder are x0, x1 and x 1+1 represents the input and output of the l+1 layer fusion network, w and b are the learnable weight matrix and bias; h l and h l+1 are the input and output of layer l, w l The weight matrix, b l is the bias vector, f(·) is the activation function of ReLU; Step 402, deeply integrating the network; The deep neural network and the multi-layer fusion network are combined. When combined, a parallel structure is adopted. The formula is as follows: f fusion +f deep (9)。

Images