A biomedical knowledge graph and transformer-based drug synergistic effect prediction method

Abstract

The application discloses a drug synergistic effect prediction method based on a biomedical knowledge graph and a Transformer, and relates to a drug synergistic effect prediction method. In order to solve the problem that a traditional drug combination discovery process mainly depends on clinical trials, which is not only time-consuming and laborious, but also high in cost and risky to patients, the application comprises the following steps: extracting drug data samples, generating a data set, dividing training and test sets, performing network training and testing; constructing a biomedical knowledge graph; converting a sequence list of a drug structure into a graph by using Rdkit; mining a subgraph of the knowledge graph by using a multi-hop subgraph mining network; learning feature representations of the knowledge network and the drug molecular graph by using a relation-aware Transformer, and performing fusion; performing synergistic effect prediction between drug pairs by using a multilayer perception machine; inputting drug pairs in a training set into the above model; inputting drug pairs in a test set into the prediction model to obtain a prediction result. The application belongs to the technical field of drug synergistic effect prediction.

Description

Technical Field

[0001] The invention relates to a drug synergistic effect prediction method, belonging to the technical field of drug synergistic effect prediction. Background Art

[0002] In the face of intractable diseases such as cancer, AIDS, and cardiovascular disease, relying solely on single drug therapy often fails to achieve satisfactory therapeutic effects. These diseases can trigger abnormalities in multiple cellular mechanisms, affecting tissues and organs throughout the body. In the field of cancer treatment, combination drug therapy strategies are becoming increasingly important, maximizing the therapeutic effect by combining multiple drugs. This approach is favored for several significant advantages: (1) it can improve the therapeutic effect; (2) it helps overcome drug resistance; and (3) it reduces the side effects caused by excessive drug dosage. For example, the combination of amiloride and hydrochlorothiazide has been shown to be effective in treating hypertension. Currently, this multi-drug combination therapy has become the standard for treating various complex diseases. However, the traditional drug combination discovery process mainly relies on clinical trials, which is not only time-consuming and labor-intensive, but also costly, and the potential risks to patients cannot be ignored. The emergence of high-throughput screening technology has enabled the accumulation of drug combination data. Researchers can now evaluate the interactions of a large number of drug combinations in various cell lines, thereby accelerating the discovery of new drug combinations. In today's era of booming biomedical data, it is almost impossible to comprehensively test all possible drug combinations. Therefore, it is imperative to develop an efficient and economical drug combination screening method.

[0003] To address this challenge, researchers have developed a range of computational methods to predict the effects of drug combinations, covering areas such as systems biology, mathematical modeling, dynamic simulation, and machine learning. Systems biology methods use biological knowledge to analyze biological networks, while mathematical methods rely on mathematical models and statistical tests. Dynamic models make predictions by simulating the dynamic changes of nodes in biological networks. Although these methods are suitable for processing small-scale data sets, they usually rely on existing knowledge. At the same time, machine learning methods, such as random forests, support vector machines, and naive Bayes, can process large-scale data sets and accelerate the process of identifying drug combinations.

[0004] In recent years, the development of deep learning technologies has brought new opportunities for drug combination prediction. These techniques can extract key features from large-scale datasets and have demonstrated their effectiveness in the biomedical field. However, while these methods have made progress in analyzing the molecular structures of drugs and cell lines, they often overlook information about drug interactions.

[0005] As a heterogeneous graph, knowledge graphs represent entities and their relationships through nodes and edges, providing a new perspective for discovering potential connections between entities. In recent years, knowledge graphs have achieved success in fields such as natural language processing and recommendation systems. Summary of the Invention

[0006] To address the problem that the traditional drug combination discovery process mainly relies on clinical trials, which is not only time-consuming and labor-intensive, but also costly and poses risks to patients, this invention proposes a drug synergy prediction method based on biomedical knowledge graphs and Transformer.

[0007] The technical solution adopted by the present invention to solve the above problems is: the specific steps of the present invention include:

[0008] Step 1: Extract drug data samples, generate data sets, divide them into training and test sets, and perform network training and testing;

[0009] Step 2: Use drug and other node information to construct a biomedical knowledge graph;

[0010] Step 3: Use Rdkit to convert the sequence representation of the drug structure into a graph;

[0011] Step 4: Use the multi-hop subgraph mining network to mine subgraphs of the knowledge graph;

[0012] Step 5: Use the relationship-aware Transformer to learn the feature representations of the knowledge network and the drug molecule graph and fuse them;

[0013] Step 6: Use a multi-layer perceptron to predict the synergistic effect between drug pairs;

[0014] Step 7: Input the drug pairs in the training set into the above model to train the model; after multiple iterations, the final prediction model is obtained;

[0015] Step 8: Input the drug pairs in the test set into the prediction model, obtain the prediction results, and analyze the results.

[0016] Furthermore, step 1 specifically includes:

[0017] The dataset used was randomly divided into a training set and a test set in a ratio of 4:1, with 80% being the training set and 20% being the test set. Both the training set and the test set contained each relationship type, and a 5-fold cross-validation process was performed on the dataset.

[0018] Furthermore, step 2 specifically includes:

[0019] The knowledge graph is represented as G = (E, R), where E represents the node set of the knowledge graph and R represents the relationship set of the knowledge graph;

[0020] The structure of the knowledge graph is represented by an entity-relationship-entity triple T = (h, r, t), where h, t∈E represent the head entity and the tail entity respectively, and r∈R represents the relationship between two entities.

[0021] Furthermore, step 4 specifically includes:

[0022] Step 401: Use the shortest path strategy to extract a knowledge subgraph for each drug or cell line, and filter out target nodes that are far from the source node;

[0023] Step 402: Use a K-hop subgraph extractor to construct a subgraph of each node by including related nodes and their adjacent nodes;

[0024] Step 403: input the sampled subgraph into the relation-aware graph converter module;

[0025] Step 404: Obtain the representation of the drug or cell line node as:

[0026] h i =MLP(BKG(i),MG(i)),

[0027] h c =BKG(c),

[0028] Among them, h i represents the drug node, h c Indicates cell lines.

[0029] Furthermore, in step 5, there are multiple different paths in the relationship path between two distant nodes. The shortest path is selected as the relationship between the two nodes:

[0030] Step 501: Relation-aware self-attention module;

[0031] Given any graph G = (V, E, R), for any node pair (h, t), there is a triple (h, r, t), where h, t ∈ V, r ∈ R, and the self-attention score is updated by the node features and relationship features;

[0032] Q(T)=f(h)W q ,

[0033] K(T)=f(t)W k ,

[0034]

[0035] Where f(h)=x h +r h→t and f(t)=x t +rt→h It is a trainable matrix used to generate query and key. φ(x t )=W v x t Used to calculate value;

[0036] Step 502: Relation-aware Transformer.

[0037] Build a new Transformer using the new self-attention mechanism,

[0038] α 1 =α 1-1 +Attention(α 1-1 ),

[0039] X l =MLP(X l-1 ),

[0040] α l Represents the nodes updated by the previous layer;

[0041] Step 503: Fuse features and make predictions;

[0042] For a drug synergistic pair (d i , d j , c k ), and their characterization is obtained through the above description (h i , h j , h k ), input it into the multi-layer perceptron for prediction,

[0043] P(i, j, k) = MLP(concat(h i , h j , h k )).

[0044] The beneficial effects of the present invention are as follows: the present invention proposes a drug synergy prediction method based on biomedical knowledge graph and Transformer, which can predict whether there is synergy between drug pairs; compared with DeepWalk, GCN, DeepSynergy, KGNN, GraphSynergy and KANSynergy, the model algorithm accuracy proposed by the present invention is better than the above six methods, and the model constructed by the present invention has improvements in ACC, AUPR and AUC. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of the present invention;

[0046] Figure 2 This is the framework diagram of the present invention DETAILED DESCRIPTION

[0047] Specific implementation method 1: Figure 1 and Figure 2 As shown in the figure, a drug synergy prediction method based on biomedical knowledge graph and Transformer includes the following steps:

[0048] Step 1: Extract drug data samples, generate a data set, divide it into training and test sets, and perform network training and testing; randomly divide the used data set into training set and 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 relationship type, and perform a 5-fold cross-validation on the data set;

[0049] Step 2: Use drug and other node information to construct a biomedical knowledge graph;

[0050] The knowledge graph is represented as G = (E, R), where E represents the node set of the knowledge graph and R represents the relationship set of the knowledge graph;

[0051] The structure of the knowledge graph is represented by an entity-relationship-entity triple T = (h, r, t), where h, t∈E represent the head entity and the tail entity respectively, and r∈R represents the relationship between two entities;

[0052] Step 3: Use Rdkit to convert the sequence representation of the drug structure into a graph;

[0053] Step 4: Use the multi-hop subgraph mining network to mine subgraphs of the knowledge graph; specifically,

[0054] Step 401: Use the shortest path strategy to extract a knowledge subgraph for each drug or cell line, and filter out target nodes that are far from the source node;

[0055] Step 402: Use a K-hop subgraph extractor to construct a subgraph of each node by including related nodes and their adjacent nodes;

[0056] Step 403: input the sampled subgraph into the relation-aware graph converter module;

[0057] Step 404: Obtain the representation of the drug or cell line node as:

[0058] h i =MLP(BKG(i),MG(i)),

[0059] h c =BKG(c),

[0060] Among them, h irepresents the drug node, h c indicates cell lines;

[0061] Step 5: Use the relationship-aware Transformer to learn the feature representations of the knowledge network and the drug molecule graph and fuse them. In the relationship path between two distant nodes, there are multiple different paths. The shortest path is selected as the relationship between the two nodes:

[0062] Step 501: Relation-aware self-attention module;

[0063] Given any graph G = (V, E, R), for any node pair (h, t), there is a triple (h, r, t), where h, t ∈ V, r ∈ R, and the self-attention score is updated by the node features and relationship features;

[0064] Q(T)=f(h)W q ,

[0065] K(T)=f(t)W k ,

[0066]

[0067] Where f(h)=x h +r h→t and f(t)=x t +r t→h It is a trainable matrix used to generate query and key. φ(x t )=W v x t Used to calculate value;

[0068] Step 502: Relation-aware Transformer.

[0069] Build a new Transformer using the new self-attention mechanism,

[0070] α 1 =α l-1 +Attention(α l-1 ),

[0071] X l =MLP(X l-1 ),

[0072] α l Represents the nodes updated by the previous layer;

[0073] Step 503: Fuse features and make predictions;

[0074] For a drug synergistic pair (di , d j , c k ), and their characterization is obtained through the above description (h i , h j , h k ), input it into the multi-layer perceptron for prediction,

[0075] P(i, j, k) = MLP(concat(h i , h j , h k ));

[0076] Step 6: Use a multi-layer perceptron to predict the synergistic effect between drug pairs;

[0077] Step 7: Input the drug pairs in the training set into the above model to train the model; after multiple iterations, the final prediction model is obtained;

[0078] Step 8: Input the drug pairs in the test set into the prediction model, obtain the prediction results, and analyze the results.

[0079] Among them, other nodes used include: target protein, cell line, tissue, and the heterogeneous information used includes: drug-protein interaction information, protein-protein interaction information, protein-cell line interaction information, and cell line-tissue interaction information.

[0080] SMILES (Simplified Molecular Input Line Entry System) is a commonly used representation of drug molecules using SMILES sequences. RDKit is a commonly used biochemical information toolkit. It provides a wide range of computational operations on 2D or 3D chemical molecules and can generate molecular descriptors for deep learning. RDKit can be used to convert sequence representations of drug molecules into 3D graphs.

[0081] This implementation considers the acquisition of knowledge graphs, including the DrugCombDB and Oncology-Screen datasets. These datasets contain the following information, as shown in Table 1: the DrugCombDB dataset contains 475 drugs, 76 cell lines, and 17,404 synergistic pairs; the Oncology-Screen dataset contains 21 drugs, 29 cell lines, and 1,044 synergistic pairs. Other biomedical knowledge data is obtained from multiple databases, including DrugBank, BindingDB, ChEMBL, PharmGKB, CCLE, and the human protein-protein interactome.

[0082] As shown in Table 1, the DrugCombDB dataset contains a total of 16,832 nodes, 4 types of relations, and a total of 250,256 triples. The Oncology-Screen dataset contains a total of 16.026 nodes, 4 types of relations, and a total of 226,475 triples.

[0083] Table 1

[0084]

[0085] Table 2

[0086]

[0087] Evaluation indicators

[0088] The drug pair interaction prediction method involved in the present invention is a binary classification problem. In the present invention, ACC, AUC and AUPR scores are used as evaluation indicators. The calculation formulas of these indicators are as follows:

[0089] ACC

[0090] Precision is used to measure the proportion of correctly classified samples among all positive and negative samples. The calculation formula is as follows:

[0091]

[0092] The definitions of TP, TN, FP, and FN are shown in Table 3.

[0093] Table 3

[0094]

[0095] AUC

[0096] 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.

[0097] AUPR

[0098] AUPR stands for Area under Precision / Recall curve, which is the area under the PR curve. The horizontal axis of the PR curve represents recall, and the vertical axis represents precision. The PR curve is easily affected by the sample distribution (the ratio of positive and negative samples in the training set). Therefore, AUPR can be used to measure the predictive performance of imbalanced datasets. An AUPR value closer to 1 indicates better model performance.

[0099] The experimental results evaluate the present invention and compare it with DeepWalk, GCN, DeepSynergy, KGNN, GraphSynergy and KANSynergy. The results are shown in Table 4:

[0100] Table 4

[0101]

[0102] The results show that the model algorithm described in the present invention is more accurate than the other six methods. Specifically, compared with other methods, the model described in the present invention has improvements in ACC, AUPR, and AUC.

[0103] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the present profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical content disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement of the above embodiments made according to the technical essence of the present invention, within the spirit and principles of the present invention, without departing from the content of the technical solution of the present invention, shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A drug synergy prediction method based on biomedical knowledge graph and Transformer, characterized by: The specific steps include: Step 1: Extract drug data samples, generate data sets, divide them into training and test sets, and perform network training and testing; Step 2: Use drug and other node information to construct a biomedical knowledge graph; Step 3: Use Rdkit to convert the sequence representation of the drug structure into a graph; Step 4: Use the multi-hop subgraph mining network to mine subgraphs of the knowledge graph; specifically, Step 401: Use the shortest path strategy to extract a knowledge subgraph for each drug or cell line, and filter out target nodes that are far from the source node; Step 402: Use a K-hop subgraph extractor to construct a subgraph of each node by including related nodes and their adjacent nodes; Step 403: input the sampled subgraph into the relation-aware graph converter module; Step 404: Obtain the representation of the drug or cell line node as: ， ， in, represents the drug node, indicates cell lines; Step 5: Use the relationship-aware Transformer to learn the feature representations of the knowledge network and the drug molecule graph and fuse them. In the relationship path between two distant nodes, there are multiple different paths. The shortest path is selected as the relationship between the two nodes: Step 501: Relation-aware self-attention module; Given any graph , for any node pair , there is a triple ,in , , update the self-attention score through node features and relationship features; ， ， ， ， ， in, and It is a trainable matrix used to generate query and key. , Used to calculate value; Step 502: Relation-aware Transformer. Build a new Transformer using the new self-attention mechanism, ， ， Represents the nodes updated by the previous layer; Step 503: Fuse features and make predictions; For a drug synergistic pair , their characterization is obtained by the above method , input it into the multi-layer perceptron for prediction, ； Step 6: Use a multi-layer perceptron to predict the synergistic effect between drug pairs; Step 7: Input the drug pairs in the training set into the model and train the model; after multiple iterations, the final prediction model is obtained; Step 8: Input the drug pairs in the test set into the prediction model, obtain the prediction results, and analyze the results.

2. The method for predicting drug synergy based on biomedical knowledge graph and Transformer according to claim 1, characterized in that: Step 1 specifically includes: The dataset used was randomly divided into a training set and a test set in a ratio of 4:1, with 80% being the training set and 20% being the test set. Both the training set and the test set contained each relationship type, and a 5-fold cross-validation process was performed on the dataset.

3. The method for predicting drug synergy based on biomedical knowledge graph and Transformer according to claim 1, characterized in that: Step 2 specifically includes: The knowledge graph is represented as ,in, Represents a set of nodes in the knowledge graph, Represents a set of relationships in a knowledge graph; The knowledge graph is organized into entity-relationship-entity triples Indicates that Represent the head entity and the tail entity respectively, Represents a relationship between two entities.

Images