A drug discovery method based on graph representation and deep learning

Abstract

The present invention discloses a drug discovery method based on graph representation and deep learning, which is applied to drug-target interaction prediction and drug-drug interaction prediction. The method combines graph representation learning and deep learning, integrates data related to drugs and targets based on DrugBank database, KEGG database and PharmGKB database, and constructs two knowledge graphs: KG-DTI knowledge graph and KG-DDI knowledge graph. Then, an improved graph representation learning model (MHRW2Vec model) is used to obtain a feature vector reflecting the interaction relationship between drugs and targets. Finally, the obtained feature vector is input into the improved TBAN model for drug-target interaction prediction and drug-drug interaction prediction. Experiments show that the method of the present invention can more deeply discover the relationship between a drug and its potential neighborhood, which is conducive to improving the accuracy of drug-target interaction and drug-drug interaction prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of biological information processing, and in particular to a drug discovery method based on graph representation and deep learning. Background Art

[0002] Traditional new drug development faces the dilemma of low throughput, long cycle, high cost and high risk. Before it is put into use, its toxicity and side effects need to be confirmed. The application of intelligent computing technology in drug discovery not only greatly shortens the time of new drug development, but also improves the success rate. Efficient drug discovery methods can quickly and effectively help doctors find drugs for symptomatic treatment, which has important practical significance.

[0003] Drug discovery technologies based on intelligent computing mostly extract features from the drug's targets, side effects, chemical properties, fingerprints and other properties. Then statistical methods, various machine learning algorithms and deep learning algorithms are used to classify and predict tasks. However, due to the lack of relevant biological information and cumbersome feature engineering, the features obtained by traditional methods either form a large and sparse binary matrix or a dense and small similarity matrix, which makes the model less robust and less scalable.

[0004] In addition, biological data involves multiple databases with different data structures and diverse data, making data processing difficult. Summary of the invention

[0005] The purpose of the present invention is to provide a drug discovery method based on graph representation and deep learning.

[0006] The present invention uses a knowledge graph to process data, where nodes in the graph represent different entities, such as drugs, diseases, protein targets, substructures, side effects, and pathways. The task of discovering new associations between drugs and targets, and between drugs and drugs can be expressed as a link prediction task. Graph embedding technology provides state-of-the-art link prediction results by capturing drugs and their potential neighborhoods, projecting each node in the graph into a dense vector. These feature vectors can be used for other downstream tasks, such as drug-target interaction prediction, drug-drug interaction prediction, etc.

[0007] The present invention provides a drug discovery method based on graph representation and deep learning, comprising the following steps:

[0008] Step 1: Extract bioinformatics data from DrugBank database, KEGG database, and PharmGKB database, and perform entity naming mapping;

[0009] Step 2: Integrate the bioinformatics data from the sub-databases of the KEGG database (including the KEGG-drug database, the KEGG-genes database, the KEGG-disease database, and the KEGG-pathway database) to construct the KG-DTI knowledge graph for drug-target interaction prediction; integrate the data from the DrugBank database, the PharmGKB database, and the KEGG-drug database to construct the KG-DDI knowledge graph for drug-drug interaction prediction;

[0010] Step 3: Construct the MHRW2Vec graph representation model, which includes two components: the MHRW random walk model and the Word2Vec model. Use the MHRW2Vec model to perform graph representation learning on the knowledge graph in step 2.

[0011] Step 4: Combine the text convolutional neural network and the bidirectional long short-term memory network, and introduce the attention mechanism to build the TBAN model. Use the batch normalization strategy to accelerate the convergence of the model. At the same time, introduce Gaussian noise in each layer of the network to optimize the performance of the TBAN model. After the training, the prediction model is obtained;

[0012] Step 5: Use the prediction model to make predictions.

[0013] Furthermore, the drug discovery method based on graph representation and deep learning provided by the present invention also has the following characteristics: in step one, the bioinformatics data includes but is not limited to drug-target interactions, drug-drug interactions, chemical properties, diseases, and target proteins.

[0014] Furthermore, the drug discovery method based on graph representation and deep learning provided by the present invention also has the following characteristics: in step one, the entity naming mapping includes the following methods at the same time: obtaining known drug-drug interaction pairs; mapping the drug ID numbers of the PubChem database, PharmGKB database and KEGG database into the drug ID numbers of the DrugBank database; mapping the drug target gold standard data set to the entity nodes of the KG-DTI knowledge graph.

[0015] Furthermore, the drug discovery method based on graph representation and deep learning provided by the present invention also has the following characteristics: the drug-related entities contained in the KG-DTI knowledge graph and the KG-DDI knowledge graph are: drugs, genes, proteins, pathways, enzymes and phenotypes; the KG-DTI knowledge graph does not contain bioinformatics data on drug-target interactions; the KG-DDI knowledge graph does not contain bioinformatics data on drug-drug interactions. The relevant program algorithm is specifically: remove the triples containing the "kegg_vocabulary:target" string in the KG-DTI knowledge graph, and remove the triples containing the "drugbank_vocabulary:ddi-interactor-in" string and the "kegg_vocabulary:Interaction" string in the KG-DDI knowledge graph.

[0016] Furthermore, the drug discovery method based on graph representation and deep learning provided by the present invention also has the following characteristics: in step three, in the MHRW2Vec model, the MHRW model constructs a Markov chain with a given probability distribution as a static distribution according to any given probability distribution, and then starts sampling after the Markov chain reaches convergence (the probability of each point being visited obeys the static distribution) to generate a node sequence. At this time, the obtained sampling set is approximately considered to obey the given distribution π; the Word2Vec model converts the sampling nodes of the MHRW model into feature vectors, and maps the feature vectors to the data sets of drug-target interactions and drug-drug interactions.

[0017] Furthermore, the drug discovery method based on graph representation and deep learning provided by the present invention also has the following characteristics: the MHRW model introduces the Metropolis-Hasting algorithm into random walks to generate node sequences in an unbiased sampling manner; the Metropolis-Hasting algorithm constructs a Markov chain so that the stationary state distribution of the Markov chain is the expected distribution, and the probability of the constructed Markov chain is expressed as:

[0018]

[0019] In formula (1), p x,y is the probability of transferring from node x to node y, is the acceptance probability, indicating whether to accept the newly generated state when constructing the Markov chain; 1-∑ z≠x q x,y is the sum of the probabilities of not accepting the new state, that is, the probability of staying at the current node; μ(·) is the static state of the expected distribution;

[0020] Since u(·) is uniformly distributed, μ(y) = μ(x), and the transition probability of the MHRW model is expressed as:

[0021]

[0022] In formula (2), p u,v is the transition probability from node u to node v, k u is the degree of node u, is to select the minimum of (node ​​u / node v) and 1.

[0023] Furthermore, the drug discovery method based on graph representation and deep learning provided by the present invention also has the following characteristics: in step 4, the loss function used in the training process of the TBAN model structure is a cross entropy loss function:

[0024]

[0025] In the formula (11), L is the loss function value, N is the number of samples, and y i represents the label of sample i, the positive example is 1, the negative example is 0, and p i Indicates the probability that sample i is predicted to be a positive class;

[0026] After calculating the loss function value, we get the gap between the predicted value of the network model and the true value, and then use the gradient descent and chain derivation rule to update the neural network parameters. The parameter update is shown in the following formula:

[0027]

[0028]

[0029] In formula (12) and formula (13), α is the learning rate, J is the loss function, is the weight between nodes, For bias.

[0030] Furthermore, the drug discovery method based on graph representation and deep learning provided by the present invention also has the following characteristics: Step 4 includes:

[0031] 1) Take the feature vector of size m×n as the input of the TBAN model, where m is the number of samples and n is the dimension of the vector. The batch normalization strategy is used in the text convolutional neural network to accelerate the model convergence;

[0032] 2) Gaussian noise is introduced into each layer of the network to improve the performance of the TBAN model. After that, the feature vector passes through a bidirectional long short-term memory network with an attention mechanism, and then passes through a fully connected layer and a Softmax layer to output the classification result;

[0033] 3) Negative samples were randomly selected from unknown interactions, with a positive-to-negative sample ratio of 1:5, and the classifier hyperparameters were optimized through ten-fold cross validation.

[0034] Compared with the prior art, the present invention has the following beneficial effects: :

[0035] The drug discovery method based on graph representation and deep learning used in the present invention can achieve the best performance on both small-scale and large-scale datasets, which shows the scalability and versatility of the method. In addition, the present invention finds the most important local information through the attention mechanism, which is consistent with the domain knowledge with certain explanatory properties. Case studies show that the method of the present invention has good performance for real drug prediction and has the following advantages:

[0036] (1) The present invention uses a semi-supervised method to extract drug and target related data from databases such as DrugBank and KEGG databases. On the other hand, a list of drug-target and drug-drug identifier combinations is compiled. In addition, the present invention also completes the ID number mapping of various databases and classifies drug targets.

[0037] (2) Integrate multi-source heterogeneous databases to construct a knowledge graph. The knowledge graph in the present invention integrates relevant data from databases such as DrugBank database, KEGG database, PharmGKB database, etc., including drugs, chemical properties, diseases, proteins, substructures, genes, side effects and pathways. The knowledge graph constructed with these diverse biological data helps to obtain the potential correlation between drugs and other entities (targets, drugs and genes). This large-scale data set also makes the model more scalable and robust.

[0038] (3) An improved knowledge representation model is designed: the MHRW2Vec model. This model is used for unbiased sampling, learns low-order representations of entities and relations, and generates feature vectors for each node of the knowledge graph, which can reflect the local and global characteristics of the knowledge graph.

[0039] (4) An improved deep learning classification model, TBAN model, was designed. This model was used to predict drug-target interaction and drug-drug interaction. The network structure can capture the local relationship values ​​in drug and target features, learn the relationship between drug targets, and thus better represent this information.

[0040] (5) Compared with traditional drug discovery methods, the present invention integrates information from multiple data sources and is an automatic, unsupervised intelligent prediction system that shows good robustness and scalability on both large-scale and small-scale data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a basic implementation flow chart of the present invention.

[0042] Figure 2 It is a structural diagram of the TBAN model in the present invention.

[0043] Figure 3 It is a graph of the AUC change curve, AUPR change curve, and loss function change curve obtained in each round in the embodiment of the present invention.

[0044] Figure 4 is the area under the receiver operating characteristic curve of the ten-fold cross validation in the embodiment of the present invention.

[0045] Figure 5 It is the area under the PR curve of the ten-fold cross validation in the embodiment of the present invention.

[0046] Figure 6 is a balanced F score curve of ten-fold cross validation in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the following embodiments are combined with the accompanying drawings to specifically illustrate the technical solutions of the present invention.

[0048] See also Figures 1 to 6 The embodiment relates to a drug discovery method based on graph representation and deep learning. The specific steps of the method are as follows:

[0049] Step 1: Extract bioinformatics data related to drugs and targets from databases such as DrugBank, KEGG, and PharmGKB (including bioinformatics data including drug-target interactions, drug-drug interactions, chemical properties, diseases, target proteins, etc.), compile a list of drug-target and drug-drug identifier combinations, map the ID numbers of other databases to the ID numbers of the DrugBank database, and classify the drugs.

[0050] Step 2: Construct a KG-DTI knowledge graph for drug-target interaction prediction based on the KEGG-drug database, KEGG-genes database, KEGG-disease database, and KEGG-pathway database; construct a KG-DDI knowledge graph for drug-drug interaction prediction based on the DrugBank database, KEGG-drug database, and PharmGKB database; remove the triplets containing the "drugbank_vocabulary:ddi-interactor-in" string and the "kegg_vocabulary:Interaction" string in the KG-DDI knowledge graph, and remove the triplets containing the "kegg_vocabulary:target" string in the KG-DTI knowledge graph.

[0051] Step 3: Construct the MHRW2Vec graph representation model, which includes two components: the MHRW random walk model and the Word2Vec model. Use the MHRW2Vec graph representation model to learn the graph representation of the knowledge graph. For each vertex v∈V, generate a graph walk sequence P with a depth of d and a root node of v. v Finally, generate N such as v r →e li →v li The graph walk sequence shown, where e li Represents vertex v r An edge of .

[0052] The MHRW model constructs a Markov chain with a given probability distribution as the static distribution, and then starts sampling after the Markov chain reaches convergence to generate a node sequence. Specifically, the MHRW model introduces the Metropolis_Hasting algorithm into random walks to generate node sequences in an unbiased sampling manner; the Metropolis_Hasting algorithm constructs a Markov chain so that the static state distribution of the Markov chain is the expected distribution, and the probability of the constructed Markov chain is expressed as:

[0053]

[0054] In formula (1), p x,y is the probability of transferring from node x to node y, is the acceptance probability, indicating whether to accept the newly generated state when constructing the Markov chain; 1-∑ z≠x q x,y is the sum of the probabilities of not accepting the new state, that is, the probability of staying at the current node; μ(·) is the static state of the expected distribution;

[0055] Since u(·) is uniformly distributed, μ(y) = μ(x), and the transition probability of the MHRW model is expressed as:

[0056]

[0057] In formula (2), p u,v is the transition probability from node u to node v, k u is the degree of node u, is to select the minimum of (node ​​u / node v) and 1.

[0058] The Word2Vec model converts the sampled nodes of the MHRW model into feature vectors and maps the feature vectors into the datasets of drug-target interaction and drug-drug interaction.

[0059] The sampled node sequence data becomes the input of the Skip-Gram model in the Word2Vec model, which uses the given words to predict the context. Specifically, given a sequence of training words w1,w2,w3,...,w t and window c, the goal of the Skip-Gram model is to maximize the following average log probability:

[0060]

[0061] In formula (3), T represents the position of the central word, c is the size of the sliding window, and w t is the tth word;

[0062] In formula (3), p(w t+j |w t ) is calculated using the Softmax function, and its expression is as follows:

[0063]

[0064] In formula (4), w t is the tth word, v wi is the input vector of the ith word, v′ wo is the output vector of the oth word, and V is the complete vocabulary;

[0065] Use the gradient descent method to minimize the loss function and substitute the calculation result of the Softmax function into formula (5):

[0066]

[0067] In formula (5), w i is the i-th word, v wiis the input vector of the ith word, v′ w is the output vector of word w, and V is the complete vocabulary;

[0068] The gradient update is shown in formula (6):

[0069]

[0070] In formula (6), w j is the jth word, v′ j is the output vector of word j, and V is the complete vocabulary;

[0071] During training, all entities and relations are projected into a low-dimensional feature space, and semantically similar entities are placed close to each other.

[0072] Step 4: Combine the text convolutional neural network and the bidirectional long short-term memory network, and introduce the attention mechanism to build the TBAN model. Input the feature vector into the TBAN model training and obtain the prediction model after training.

[0073] The present invention first randomly selects negative samples from the undetermined drug-target interaction set and drug-drug interaction set, so that the positive and negative sample ratio of the model is 1:5. The feature vector first passes through a four-layer one-dimensional convolutional neural network, and the convolution kernel sizes are 128, 32, 32, and 32, respectively. The convolution layers all use the maximum pooling method to extract features. Each layer of the network adds a batch normalization strategy to reduce overfitting, and the activation function of the four-layer structure uses the RELU function. This method increases the robustness of the system.

[0074] The extracted feature vectors are then subjected to average pooling and maximum pooling operations respectively, and two types of feature vectors are obtained and then vector concatenation is performed. The feature vector is then passed through a bidirectional long short-term memory network that integrates an attention mechanism, and the classification result is output through a fully connected layer and a Softmax layer. The attention mechanism takes the weighted sum of the vectors of all sequences as a new feature vector, and the dimension of the feature vector does not change in this layer. It is defined as follows:

[0075] M=tanh(H) (7)

[0076] α=Softmax(ω T M) (8)

[0077] r=Hα T (9)

[0078] h * =tanh(r) (10)

[0079] Where H is the output matrix of the bidirectional long short-term memory network d ωis the dimension of the feature vector, ω is the feature vector, ω T is the transpose of the eigenvector. The size of ω, α, r is d ω ,t,d ω Finally, we get the output vector h * .

[0080] The loss function used in the training process of the TBAN model is the cross entropy loss function, as shown in formula (11).

[0081]

[0082] In formula (11), L is the loss function value, N is the number of samples, and y i represents the label of sample i, the positive example is 1, the negative example is 0, and p i Indicates the probability that sample i is predicted to be a positive class;

[0083] After calculating the loss function value, the gap between the predicted value and the true value can be obtained, and then the neural network parameters are updated using gradient descent and chain derivation rule. The parameter updates are shown in equations (12) and (13).

[0084]

[0085]

[0086] In formulas (12) and (13), α is the learning rate, J is the loss function, is the weight between nodes, For bias.

[0087] During the training process, the TBAN model can better fit the data by continuously calculating the loss value and updating the parameters. After the training, the drug-target interaction and drug-drug interaction prediction models are obtained.

[0088] Step 5: Use the prediction model to make predictions.

[0089] The above embodiments are only used to help understand the method and core idea of ​​the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A drug discovery method based on graph representation and deep learning, characterized in that: The following steps are involved: Step 1: Extract bioinformatics data from DrugBank database, KEGG database, and PharmGKB database, and perform entity naming mapping; Step 2: Integrate the bioinformatics data from the sub-databases of the KEGG database to construct the KG-DTI knowledge graph for drug-target interaction prediction; integrate the data from the DrugBank database, PharmGKB database, and KEGG-drug database to construct the KG-DDI knowledge graph for drug-drug interaction prediction; Step 3: Construct an MHRW2Vec graph representation model, which includes two components: an MHRW random walk model and a Word2Vec model. Use the MHRW2Vec model to perform graph representation learning on the knowledge graph in step 2. Step 4: Combine the text convolutional neural network and the bidirectional long short-term memory network, and introduce the attention mechanism to build the TBAN model. Use the batch normalization strategy to accelerate the convergence of the model. At the same time, introduce Gaussian noise in each layer of the network to optimize the performance of the TBAN model. After the training, the prediction model is obtained; Step 5: Use the prediction model to make predictions; Among them, the loss function used in the training process of the TBAN model structure is the cross entropy loss function: In the formula (11), L is the loss function value, N is the number of samples, and y i represents the label of sample i, the positive example is 1, the negative example is 0, and p i Indicates the probability that sample i is predicted to be a positive class; After calculating the loss function value, we get the gap between the predicted value of the network model and the true value, and then use the gradient descent and chain derivation rule to update the neural network parameters. The parameter update is shown in the following formula: In formula (12) and formula (13), α is the learning rate, J is the loss function, is the weight between nodes, is bias; Step 4 specifically includes the following steps: 1) Take the feature vector of size m×n as the input of the TBAN model, where m is the number of samples and n is the dimension of the vector. The batch normalization strategy is used in the text convolutional neural network to accelerate the model convergence; 2) Gaussian noise is introduced into each layer of the network to improve the performance of the TBAN model. After that, the feature vector is output as a classification result through a bidirectional long short-term memory network fused with an attention mechanism, a fully connected layer, and a Softmax layer; 3) Negative samples were randomly selected from unknown interactions, with a positive-to-negative sample ratio of 1:5, and the classifier hyperparameters were optimized through ten-fold cross validation.

2. The drug discovery method based on graph representation and deep learning according to claim 1, characterized in that: In step 1, the bioinformatics data includes drug-target interactions, drug-drug interactions, chemical properties, diseases, and target proteins.

3. The drug discovery method based on graph representation and deep learning as claimed in claim 1, characterized in that: In step 1, the entity naming mapping includes the following methods: obtaining known drug-drug interaction pairs; mapping the drug ID numbers of the PubChem database, PharmGKB database, and KEGG database into the drug ID numbers of the DrugBank database; and mapping the drug target gold standard data set to the entity nodes of the KG-DTI interaction knowledge graph.

4. The drug discovery method based on graph representation and deep learning as claimed in claim 1, characterized in that: In step 2, the drug-related entities contained in the KG-DTI knowledge graph and the KG-DDI knowledge graph include: drugs, genes, proteins, pathways, enzymes, and phenotypes; The KG-DTI knowledge graph does not contain biological information data on drug-target interactions; The KG-DDI knowledge graph does not contain bioinformatics data on drug-drug interactions.

5. The drug discovery method based on graph representation and deep learning as claimed in claim 1, characterized in that: In step 3, in the MHRW2Vec model, the MHRW model constructs a Markov chain with a given probability distribution as a static distribution according to any given probability distribution, and then starts sampling after the Markov chain reaches convergence to generate a node sequence; The Word2Vec model converts the sampling nodes of the MHRW model into feature vectors, and maps the feature vectors into the data sets of drug-target interaction and drug-drug interaction.

6. The drug discovery method based on graph representation and deep learning according to claim 5, characterized in that: The MHRW model introduces the Metropolis_Hasting algorithm into random walks to generate node sequences in an unbiased sampling manner; The Metropolis_Hasting algorithm constructs a Markov chain so that the stationary state distribution of the Markov chain is the expected distribution. The probability of the constructed Markov chain is expressed as: In formula (1), p x,y is the probability of transferring from node x to node y, is the acceptance probability, indicating whether to accept the newly generated state when constructing the Markov chain; 1-Σ z≠x q x,y is the sum of the probabilities of not accepting the new state, that is, the probability of staying at the current node; μ(·) is the static state of the expected distribution; Since u(·) is uniformly distributed, μ(y) = μ(x), and the transition probability of the MHRW model is expressed as: In formula (2), p u,v is the transition probability from node u to node v, k u is the degree of node u, is to select the minimum of (node ​​u / node v) and 1.

Images