A drug efficacy prediction method based on graph neural network and omics information
Through the drug efficacy prediction method based on graph neural network, drug transcriptome data is integrated and structural relationship diagram is constructed, and the existing drug screening methods are solved, and efficient and accurate drug efficacy prediction is achieved.
Patent Information
- Application Number
- CN202210206033.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-04
- Publication Date
- 2025-05-02
- Estimated Expiration
- 2042-03-04
AI Technical Summary
Existing drug screening methods have problems such as high cost, overly idealized dependence on structural information and targeting ideas, and it is difficult to effectively screen out compounds with drug efficacy.
A drug efficacy prediction method based on graph neural network and omics information is adopted to integrate drug transcriptome data to construct a structural relationship diagram between drugs and biological processes, and a graph neural network is used to extract drug characteristics and perform functional clustering to predict the functions of the compounds to be tested.
This method can effectively reduce the cost of drug screening, avoid the disadvantages of relying on structural information, and improve the accuracy and efficiency of drug efficacy prediction.
Smart Images
Figure CN114649097B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of drug screening, and in particular to a drug efficacy prediction method based on graph neural network and omics information. Background Art
[0002] Drug development is a field with high risks, long cycles and high costs. According to statistics, the development of a new drug usually costs 500 million to 1 billion US dollars, takes 10 to 15 years or even longer, and is highly accidental and blind. Drug screening is the basis of new drug development. Finding suitable compounds as candidate drugs for preclinical experiments from tens of millions of molecules is a project like finding a needle in a haystack.
[0003] The current drug screening methods can be divided into two categories: First, phenotypic screening. Phenotype is the sum of the characteristics of an individual with a specific genotype under certain environmental conditions. Drug phenotypic screening is a screening method based on phenotypic changes under specific environments. Phenotypic screening of drugs has gone through a long time. Traditional drug phenotypic screening is more based on the screening of animal disease models, by observing the changes in the phenotype of the body caused by drugs. With the advancement of technology, the tiny elements that make up living organisms, such as cells, proteins, genes, etc., can be studied at a deeper level. Specifically, it can be divided into screening of animal disease models and in vitro drug screening. Second, computer-aided drug design. Computer-aided drug design is a method based on computer chemistry to design and optimize lead compounds through computer simulation, calculation and budget of the relationship between drugs and receptor biomacromolecules. Computer-aided drug design is actually to optimize and design lead compounds by simulating and calculating the interaction between receptors and ligands. Specifically, it can be divided into direct drug design that directly designs specific ligands based on the three-dimensional structure of receptor protein molecules, and indirect drug design based on ligands when the protein structure of the drug target is unclear.
[0004] However, the above-mentioned drug screening methods also have certain defects: for example, although phenotypic screening does not depend on the structure of the drug, it also faces huge challenges. The first is that the ultimate goal of our drug screening is to apply it to humans, and human diseases are a very complex process. The first difficulty is how to determine a phenotype for screening. In addition, our current phenotypic screening systems are mainly based on cells, fruit flies, nematodes, zebrafish, etc. These animals can be screened at a high throughput, but these animals are also quite different from human diseases. So we can summarize the challenges of phenotypic screening as the contradiction between complex diseases and simple means.
[0005] For example, computer-aided drug design has the disadvantage of being too dependent on structure and targeting. When screening drugs, especially first-in-class drugs, we always prefer drugs with clear targets, clear structures and targeting. Whether it is molecular docking in direct drug design or pharmacophore models and quantitative structure-activity relationships in indirect drug design, the shortcomings of the above methods can be attributed to the dependence on compound structure and targeting. First of all, there are challenges in the method itself. One is the accuracy of protein structure information in public databases, which is the primary condition for ensuring accurate results. Although researchers have high-end crystallographic hardware, data processing and structure optimization software in structural studies, false positives and inconsistent data may still occur. Second, the method depends on protein structure, and the method of analyzing protein structure is mainly through X-ray crystallography, NMR spectroscopy (nuclear magnetic resonance, NMR), nuclear magnetic resonance and cryo-electron microscopy. These experiments are limited by cost and time, and have certain experimental difficulties. X-ray crystallography can only analyze protein structure when the protein can be crystallized. However, most proteins are difficult to crystallize, accounting for 60% of currently approved drugs. Nuclear magnetic resonance is also only applicable to small molecule proteins. Furthermore, this method is based on an ideal state, that is, a disease can be treated by a target, and a target can be precisely targeted by a drug. However, we know that most diseases, especially various internal medicine diseases, are a systematic and comprehensive pathophysiological progression, not driven by a single gene or a single target. On the other hand, it is difficult for a compound to target only one target.
[0006] In this regard, the scientific research community is also working hard to overcome these problems, such as using human model animals and 3D culture technology, all of which aim to make the screened phenotypes more and more similar to complex human diseases. There have been a series of advances, especially in the fields of regenerative medicine, tumors, and immunity. However, complex phenotypes mean more investment, and the high price of cutting-edge phenotypic models means that only a very small number of groups can use them. Therefore, even if a screening system for complex phenotypes is established, the previous high investment has limited the drug screening methods of many practitioners. Summary of the invention
[0007] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a drug efficacy prediction method based on graph neural network and omics information. The present invention is based on drug transcriptome data, integrates massive drug transcriptome data through big data processing methods, extracts drug features through graph neural network, and performs functional clustering on existing drugs. Then, the compound to be predicted is compared with the existing drugs after clustering, and the function of the compound to be tested is predicted by this method.
[0008] To achieve the above object, the technical solution adopted by the present invention is:
[0009] The present invention provides a method for predicting drug efficacy based on graph neural network and omics information, comprising the following steps:
[0010] 1) Collection: Collect existing drug transcriptome data, unify the transcriptome data through gene set enrichment analysis, and obtain unified transcriptome data;
[0011] 2) Constructing a structural relationship graph: Converting unified transcriptome data into biological processes as node features, and representing relationships as edges, to construct a structural relationship graph between existing drugs and biological processes;
[0012] 3) Node learning representation: Combined with the idea of structural relationship graph representation learning, the node features of step 2) are vectorized using the candidate method to obtain the node feature vector;
[0013] 4) Graph Neural Network: Based on the constructed structural relationship graph, the node features are aggregated and updated. The similarity between the test compound and the existing drugs is compared through the extracted node feature vectors of the drug, so as to predict and screen the function of the test compound.
[0014] As a preferred embodiment of the drug efficacy prediction method of the present invention, the specific steps of collecting existing drug transcriptome data are as follows:
[0015] ① Collection of existing drug information:
[0016] The first step is to comprehensively collect information on existing drugs to ensure the normal operation of subsequent graph neural networks, and to integrate the drug catalog on drugbank;
[0017] ②Mining of transcriptome information:
[0018] According to the integrated drug catalog, data retrieval is planned in GEO and Array Express databases.
[0019] High-throughput data is often accompanied by massive amounts of information. For example, the current one-time sequencing depth can reach more than 10G. Although the transcriptome data is rich, there are problems such as the large time span of each experiment and inconsistent measurement standards (different samples and sequencing methods). Complex data must be processed in a unified manner. At the same time, how to apply big data processing methods to process these data, how to select and set the graph neural network algorithm, etc. are important issues and are also the technical key points of the present invention.
[0020] As a preferred embodiment of the drug efficacy prediction method of the present invention, the selected methods in step 3) include DEEPWALK, LINE, Metapath2Vec++, and GATNE.
[0021] As a preferred implementation of the drug efficacy prediction method of the present invention, the method for aggregating and updating node features in step 4) includes GAT and HAN.
[0022] As a preferred embodiment of the drug efficacy prediction method of the present invention, the drug to be tested includes acacetin.
[0023] As a preferred implementation of the drug efficacy prediction method of the present invention, the node features are divided into two representation modes, Up and Down, according to whether the biological process is increased or decreased.
[0024] As a preferred embodiment of the drug efficacy prediction method of the present invention, when Metapath2Vec++ is used as a candidate method, according to Drug--Biological Process--Drug (V bp-ap ) and Drug--Biological Process--Drug(V bp-down ) The two metapaths perform random walks on the node features, and the sequences obtained by random walks are used as training corpus. The training corpus is collected according to the positive and negative sampling optimization method of metapath2Vec++, and the nodes are vectorized.
[0025] As a preferred embodiment of the drug efficacy prediction method of the present invention, when HAN is selected to aggregate and update the node features, the constructed structural relationship graph is a heterogeneous graph G with three types of node attributes. bp-ap ) and Drug--Biological Process--Drug(V bp-down ) Two meta-paths split the heterogeneous graph G into two heterogeneous graphs G up , G down ; Import the learned vector as the node feature, calculate the node attention and semantic attention, and obtain the final node feature vector.
[0026] Compared with the prior art, the present invention has the following beneficial effects:
[0027] In phenotypic screening, although effective screening models can relatively well restore the disease environment of the human body, they also have the problem of high cost. Although virtual screening solves the cost problem, its over-reliance on structural methods and over-idealized targeting ideas have become its inevitable disadvantages. The present invention provides a drug efficacy prediction method based on graph neural networks and omics information. By informatizing phenotypic data, it avoids the high cost problem of phenotypic screening and solves the problem of virtual screening over-reliance on structural information. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 This is a visualization diagram of the implementation of Example 1;
[0029] Figure 2 for Figure 1 A partial enlarged view of
[0030] Figure 3 The fasting blood glucose changes of rats in each group;
[0031] Figure 4 The graphs of blood glucose changes after meals in each group of rats;
[0032] Figure 5 This is a graph of insulin resistance test results. DETAILED DESCRIPTION
[0033] In order to better illustrate the purpose, technical solutions and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.
[0034] In the following examples, the experimental methods used are conventional methods unless otherwise specified, and the materials, reagents, etc. used are all commercially available unless otherwise specified.
[0035] The present invention realizes the prediction of the efficacy of compounds based on graph neural network and omics information. Among them, graph neural network belongs to the category of artificial intelligence (AI).
[0036] Example
[0037] This example uses transcriptomes of some marketed drugs and the test compound Acacetin to be implemented ( Figure 1 , Figure 2 for illustration).
[0038] ①Convert the gene set information of the transcriptome into biological process through GSEA analysis.
[0039] ② The drug and all the biological processes involved and their relationships are constructed into a graph structure G = (V, E). The node V of the biological process is divided into two representation modes: Up and Down according to whether the biological process is up or down.
[0040] ③Use Metapath2Vec++ graph learning representation method, according to Drug--Biological Process--Drug (V bp-ap ) and Drug--Biological Process--Drug(V bp-down) Two meta-paths perform random walks on all nodes, and the sequences obtained by random walks are used as training corpus. The training corpus is collected according to the positive and negative sampling optimization method of metapath2Vec++, and the nodes are vectorized.
[0041] Metapath2Vec++ is a graph representation learning method developed based on the idea of Word2Vec in Natural Language Processing (NLP). The Word2Vec method is to traverse all the words in the corpus as the central word, select the words above and below the central word as the positive samples of the adjacent words according to the set window length, and define the labels of other words as negative samples. The learning parameters are maximized according to the conditional probability of positive and negative samples, and all words are word vectorized.
[0042] Metapath2Vec++ uses the sequence obtained by random walking the graph structure according to the semantic relationship (in this embodiment, the relationship between drugs and biological processes) as the corpus, traverses the nodes therein, and vectorizes the nodes in the graph according to the Wrod2Vec method.
[0043] ④ HAN (Heterogeneous Graph Attention Network) is used for graph neural network analysis. According to ②, the graph G constructed by the present invention is a heterogeneous graph G with three types of node attributes. According to the two meta-paths in ③, the heterogeneous graph G is split into two heterogeneous graphs G ap , G down Import the vector learned in ③ as node features. Calculate node attention (node-level) and semantic attention (semantic-level). Obtain the final node feature vector. Use this feature to compare, predict, and screen drug efficacy.
[0044] HAN is a graph attention network (GAT) based on the attention mechanism to learn heterogeneous graphs, and the graph attention network (GAT) is a space-based graph convolution network (GCN).
[0045] The GCN graph convolutional network generalizes the convolution operation from traditional data (such as images) to graph data. Its core idea is to learn a function mapping f(.) through which the nodes in the graph can aggregate their own features with the features of their neighbor nodes to generate new representations.
[0046] The difference between the GAT graph attention network and the graph convolutional network is that the attention mechanism is introduced, which is a method of learning the weights of different neighbor nodes while aggregating the features of neighbor nodes around the node. The weight is the attention parameter. Compared with GCN, which treats all neighbors of a node equally, the attention mechanism can assign different attention scores to each neighbor, thereby identifying more important neighbor nodes.
[0047] Graph neural network algorithms such as GCN and GAT are still based on isomorphic graphs, while this implementation case constructs a heterogeneous graph that contains the semantic relationship between drugs and biological processes. HAN is a graph neural network based on GAT for learning heterogeneous graphs. In the meta-path under a specific semantic relationship, each node performs GAT, aggregates the features of neighboring nodes, and constructs node-level attention features. In a heterogeneous graph, a node usually contains multiple types of semantics, and the node features of specific semantics only contain information about one aspect of the node. To learn more complex node features, it is necessary to fuse multiple semantic information based on meta-paths. In this embodiment, node-level features are first aggregated according to the meta-paths under the two semantics of drug-ascending biological processes and drug-downregulating biological processes, and then the features under the two semantics are aggregated again as the final features of the node through the semantic attention learning mechanism, and this feature is used to predict and screen drug efficacy. Figure 1 The dark dots in the middle left continuous section are biological processes that are upregulated, and the light dots are biological processes that are downregulated; Figure 1 The reunion part is some marketed drugs and acacetin, Figure 2 for Figure 1 In the local enlarged image, the point Metformin has a high similarity with the vector of the test compound. The prediction results show that acacetin may have the effect of controlling blood sugar.
[0048] ⑤Verify the prediction results through in vivo and in vitro experiments.
[0049] Anti-hyperglycemic test: Male SD rats (5-7 weeks old; weight 140-150 g) fed with a high-fat, high-fructose (HFF) diet were used to test the anti-hyperglycemic effect of Acacetin at a single dose of 150 mg / kg. The same dose of distilled water was added as the control group, and HFF+Acacetin was used as the experimental group.
[0050] refer to Figure 3-4 The fasting and postprandial blood sugar changes of rats in each group are shown, which indicate that acacetin has a significant effect in controlling blood sugar.
[0051] Insulin resistance test: L6 rat skeletal muscle cells were treated with free fatty acids (FFA) (0.75mM) for 24 hours, and the changes in glucose uptake in rat skeletal muscle cells under insulin resistance were evaluated by the fluorescent indicator 2-NBDG after acacetin treatment. The results showed that acacetin increased glucose uptake (reference Figure 5 ). The above experiments verified the results of acacetin in controlling blood sugar.
[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the essence and scope of the technical solution of the present invention.
Claims
1. A drug efficacy prediction method based on graph neural network and omics information, characterized in that: The following steps are involved: 1) Collection: Collect existing drug transcriptome data, unify the transcriptome data through gene set enrichment analysis, and obtain unified transcriptome data; 2) Constructing a structural relationship graph: Converting unified transcriptome data into biological processes as node features, and representing relationships as edges, to construct a structural relationship graph between existing drugs and biological processes; 3) Node learning representation: Combined with the idea of structural relationship graph representation learning, the node features of step 2) are vectorized using the candidate method to obtain the node feature vector; 4) Graph neural network: Based on the constructed structural relationship graph, the node features are aggregated and updated. The similarity between the test compound and the existing drugs is compared through the extracted node feature vectors of the drug, so as to predict and screen the function of the test compound; Specific steps for collecting existing drug transcriptome data: ① Collection of existing drug information: The first step is to comprehensively collect information on existing drugs to ensure the normal operation of subsequent graph neural networks, and to integrate the drug catalog on the drug bank; ②Mining of transcriptome information: According to the integrated drug catalog, data retrieval is planned in GEO and Array Express databases; In step 3), the selected methods include DEEP WALK, LINE, Metapath2Vec++, and GATNE; The methods for aggregating and updating node features in step 4) include GAT and HAN.
2. The drug efficacy prediction method according to claim 1, characterized in that: The drug to be tested includes acacetin.
3. The drug efficacy prediction method according to claim 1, characterized in that: Node features are divided into two representation modes: Up and Down, depending on whether the biological process is upregulated or downregulated.
4. The drug efficacy prediction method according to claim 1, characterized in that: When Metapath2Vec++ is used as the candidate method, the node features are randomly walked according to the two meta-paths of Drug--Biological Process--Drug and Drug--Biological Process--Drug. The sequences obtained by random walk are used as training corpus. The training corpus is collected according to the positive and negative sampling optimization method of metapath2Vec++, and the nodes are vectorized.
5. The drug efficacy prediction method according to claim 1, characterized in that: When HAN is used to aggregate and update node features, the constructed structural relationship graph is a heterogeneous graph G with three types of node attributes. According to the two meta-paths Drug--BiologicalProcess--Drug and Drug--Biological Process--Drug, the heterogeneous graph G is split into two heterogeneous graphs G up , G down ; Import the learned vector as the node feature, calculate the node attention and semantic attention, and obtain the final node feature vector.