A drug side effect prediction model based on a meta-path graph neural network

Abstract

The present invention relates to the technical field of drugs, and particularly to a drug side effect prediction model based on a meta-path graph neural network, comprising: an input module, a meta-path-based feature learning module, and a prediction module; Input module: In the input module, a heterogeneous information network based on multiple data sets is constructed, and these data sets involve heterogeneous biological entities and the associations between them; Meta-path-based feature learning module: In the meta-path-based feature learning module, L-layer meta-path-based information aggregation is used to obtain meaningful representations of heterogeneous nodes by fully capturing the complex semantics in the heterogeneous information network based on multiple data sets; Prediction module: The predicted drug side effect associations are obtained by a fully connected neural network for the representations of drugs and side effects.

Description

Technical Field

[0001] The present invention relates to the technical field of drugs, and particularly to a drug side effect prediction model based on a meta-path graph neural network. Background Art

[0002] The side effects of drugs have received increasing attention in the biomedical field. Accurately identifying drug side effects is crucial for drug development and drug safety monitoring. Although traditional pharmacological experiments can accurately detect drug side effects, the identification process is time-consuming, costly, and may result in incomplete side effect identification.

[0003] With the expansion of various biomedical databases, many computational methods have been developed for the drug side effect association (DSA) prediction task. The existing DSA prediction methods can be divided into three categories: (1) docking-based methods. Docking refers to the process of identifying the binding position of a ligand to a target. According to biomedical knowledge, when a drug binds to a specific protein, certain side effects will occur. Therefore, docking-based methods attempt to predict DSA by identifying the potential binding sites of drugs. (2) machine learning-based methods, which regard DSA prediction as a classification problem. By using known drug side effect associations as training data, traditional classification algorithms or deep learning-based classifiers are used to predict new DSA. (3) network-based methods, which regard DSA-related biological entities (such as drugs, targets, and side effects) as nodes and the connections between nodes as edges to construct a DSA-related network, and model DSA prediction as a link prediction problem.

[0004] Although researchers have proposed many methods for the DSA prediction task, three common drawbacks still need to be addressed. First, most existing methods fail to fully utilize various data that may affect the occurrence of drug side effects. According to pharmacokinetics (PK) and pharmacodynamics (PD), DSA may be affected by several biological processes involving drug-target interactions, drug-disease associations, and disease-protein associations. The corresponding datasets will help learn high-quality features of drugs and side effects and improve the performance of DSA prediction. Second, existing methods cannot capture the complex semantic associations between biological entities involved in the occurrence of drug side effects. The situation where different entities such as drugs, side effects, proteins, and diseases play different roles and have heterogeneous characteristics poses a great challenge to capturing the complex semantic associations between drugs and side effects. Last but not least, most existing methods lack interpretability of the derived DSA, that is, these methods usually can only provide the predicted DSA and cannot clearly describe the reasons for the derived predictions. However, interpretability is important for doctors to make some decisions, such as prescribing drugs to her or his patients. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a drug side effect prediction model based on a meta-path graph neural network.

[0006] A drug side effect prediction model based on a meta-path graph neural network of the present invention includes: an input module, a meta-path based feature learning module, and a prediction module;

[0007] Input module: In the input module, a heterogeneous information network based on multiple data sets is constructed, and these data sets involve heterogeneous biological entities and the associations between them;

[0008] Meta-path based feature learning module: In the meta-path based feature learning module, L-layer meta-path based information aggregation is utilized to obtain meaningful representations of heterogeneous nodes by fully capturing the complex semantics in the heterogeneous information network based on multiple data sets;

[0009] Prediction module: The predicted drug side effect association is obtained through a fully connected neural network for the representations of drugs and side effects.

[0010] Furthermore, four entities, namely drugs, proteins, side effects, and diseases, are used to construct a heterogeneous information network based on multiple data sets.

[0011] Furthermore, the meta-path based feature learning module includes L-layer meta-path based information aggregation, and each layer includes two parts, namely meta-path based information propagation and meta-path level attention mechanism;

[0012] Meta-path based information propagation: When a specific node and a meta-path are given, meta-path based information propagation can extract the information of the neighbors of this node based on the meta-path and aggregate it, and the node embedding of the previous layer is also introduced as a residual connection;

[0013] Meta-path level attention mechanism: When multiple meta-paths are given, through meta-path based information propagation, multiple embeddings of the same node can be obtained, and the meta-path level attention mechanism automatically learns the weights of these embeddings.

[0014] Furthermore, for the prediction module, given a pair of drugs and side effects, the prediction probability of their association is derived from a fully connected layer with a Sigmoid function.

[0015] The beneficial effects of the present invention compared with the prior art are as follows: 1. The docking-based methods attempt to predict DSA by identifying potential binding sites of drugs. However, the scarcity of drugs and proteins with known 3D structures limits the wide application of docking-based methods, which rely too much on the accurate 3D structures of target proteins and drug compounds. The present invention constructs a heterogeneous information network to mine the semantic types between entities and does not directly rely on the known 3D structures of drugs and proteins;

[0016] 2. For machine learning-based methods, DSA prediction is regarded as a classification problem. By using known drug side effect associations as training data, traditional classifiers or deep learning-based classifiers are learned to predict new DSA. However, due to limited training samples, especially positive samples, the performance of machine learning-based DSA prediction methods needs to be improved in practical applications; the sampling ratio of positive and negative samples in the dataset designed in the experiment of the present invention is 1:1, which solves the problem of sample imbalance;

[0017] 3. In network-based methods, DSA-related biological entities (such as drugs, targets, and side effects) are regarded as nodes, and the connections between nodes are regarded as edges to construct a DSA-related network. This type of method models DSA prediction as a link prediction problem and uses network-related tools, such as matrix operations and graph neural networks (GNNs), to predict potential DSA accordingly. However, since most similarity-based methods are designed based on homogeneous networks and do not fully consider the non-uniformity between nodes and edges, these methods usually cannot capture the complex semantic relationships between drugs and side effects, resulting in poor performance in predicting DSA; the present invention designs an effective meta-path-based fusion mechanism to capture the complex semantics in the HIN and obtain high-quality embeddings of drugs and side effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is the overall technical roadmap of the present invention;

[0019] Figure 2 is the working flowchart of the l-th layer of meta-path-based information aggregation. DETAILED DESCRIPTION OF THE INVENTION

[0020] The following combines the drawings and embodiments to further describe in detail the specific implementation manners of the present invention. The following embodiments are used to illustrate the present invention but are not used to limit the scope of the present invention.

[0021] In response to the background and invention content summarized above, the overall technical route adopted by the present invention is as Figure 1As shown, it is divided into three modules: an input module, a meta-path based feature learning module, and a prediction module. First, a heterogeneous information network (HIN) based on multiple datasets is constructed in the input module, and these datasets involve heterogeneous biological entities and the associations between them. Then, in the feature learning module, L-layer meta-path based information aggregation is utilized to obtain meaningful representations of heterogeneous nodes by fully capturing the complex semantics in the HIN. Finally, the predicted drug-side effect associations are obtained through a fully connected neural network for the representations of drugs and side effects.

[0022] According to the above overall technical route, the specific technical solutions adopted by the present invention are described as follows:

[0023] 1. Input module

[0024] In this study, four types of entities are considered to construct the HIN, namely drugs, proteins, side effects, and diseases. More specifically, the sets of drugs, proteins, side effects, and diseases are represented by V dr , V pr , V se , and V di respectively. Additionally, the sizes of V dr , V pr , V se , V di are represented by n dr , n pr , n se , n di respectively.

[0025] For these four types of entities, eight associations are considered, including four homogeneous connections and four heterogeneous connections. More specifically, the four homogeneous connections are drug-drug interactions (represented as ), protein-protein interactions (represented as ), side effect-side effect similarity associations (represented as ), and disease-disease similarity associations (represented as ). The four heterogeneous connections are drug-protein interactions (represented as ), drug-disease associations (represented as ), protein-disease associations (represented as ), and drug-side effect associations (represented as ). Note that for the two similarity associations (i.e., side effect-side effect similarity association and disease-disease similarity association), the range of the entry values in the matrix (i.e., A ss and A ee ) is [0, 1].

[0026] Given multiple biological datasets as input, the constructed HIN is represented as G=(V, E), where V is the set of four heterogeneous types of nodes (i.e., V = V dr ∪V pr ∪V se ∪V di ), and E represents the set of connection types (E = A dd ∪A pp ∪A ss ∪A ee ∪A dp ∪A de ∪A pe ∪A ds ).

[0027] In this implementation, one-hot encoding is used as the initial feature representation of nodes. Specifically, the initial embedding of nodes in the HIN is H (0) defined as follows:

[0028]

[0029] where, are the initial embeddings of drugs, proteins, side effects, and diseases. Based on H (0) , the designed meta-path-based aggregation mechanism will learn more meaningful node embeddings.

[0030] 2. Meta-Path-Based Feature Learning Module

[0031] In the meta-path-based feature learning module, a meta-path-based information aggregation mechanism is proposed to make full use of the complex semantics in the HIN to learn more meaningful node embeddings.

[0032] A meta-path, which is a basic element in the HIN, is a path defined on the network pattern. According to the definition of the meta-path, it is known that different meta-paths represent different semantic information contained in the HIN. Therefore, an effective feature learning method needs to make full use of multiple meta-paths to capture the complex semantics in the HIN.

[0033] The proposed meta-path-based feature learning module consists of L layers of meta-path-based information aggregation. In each layer, first, meta-path-based information propagation is performed on each node based on the selected meta-path, thereby collecting meaningful information from each important meta-path. Then, a meta-path-level attention mechanism is designed to consider the different contributions of meta-paths to updating node embeddings; the details of the meta-path-based information aggregation in the l-th layer will be explained. The workflow of the meta-path-based information aggregation in the l-th layer is as Figure 2 shown.

[0034] 21. Meta-Path-Based Information Propagation

[0035] In the meta-path-based information propagation, four types of nodes are treated equally, and the embedding update processes of different nodes are carried out in a similar way. Without loss of generality, taking the drug node i as an example to describe the idea of meta-path-based information propagation.

[0036] Given an instance of a meta-path, according to the meta-path, the end node is called the neighbor of the start node. The set of node neighbors according to a certain meta-path is defined as all neighbors of this node under this meta-path. Suppose in the implementation, a total of M meta-paths are considered and represented as {P1, …, P m , …, P M}, given a meta-path P m and a drug node i, the neighbor set of the drug node i under the meta-path P m is represented as: where each element represents the index of the neighbor, and is the total number of instances of the meta-path P m with the start node i; according to the meta-path P m , the semantic information of the meta-path P m used to update the embedding of drug i is contained in all neighbors of drug i. Therefore, in the l-th layer, the updated embedding of node i according to the meta-path P m is calculated as follows.

[0037]

[0038] where is the training parameter of the l-th layer of the meta-path P m , d is the dimension of the node embedding, the first term in ReLU(·) is the average contribution of all neighbors under the meta-path P m , and the second term is the embedding of node i here as a residual connection.

[0039] Through the above process, considering the semantics in the meta-path P m , the updated embedding of node i is obtained.

[0040] To effectively combine the contributions of all meta-paths, a meta-path-level attention mechanism is designed.

[0041] 22. Meta-path-level attention mechanism

[0042] Since different meta-paths represent different semantics between node pairs, the meta-path-level attention mechanism is used to capture the distinctiveness between meta-paths. In the l-th layer, under the meta-path P m , given the embedding of node i, a transformation is first performed, which is formally described as follows:

[0043]

[0044] Among them, the matrix W a ∈R d×d and the bias vector b a ∈R d are the training parameters in the non-linear activation function tanh(·). Then, the normalized attention scores for each meta-path are calculated through the following equation:

[0045]

[0046] where <·,·> represents the inner product operation of two vectors, and u a represents the trainable context vector of the hierarchical attention. Finally, the final embedding of the drug node i is obtained by the weighted average of the embeddings obtained according to M meta-paths:

[0047]

[0048] By utilizing L-layer meta-path-based information aggregation, the final representation of the node is obtained, denoted by H as follows:

[0049]

[0050] where H dr , H pr , H se and H di are the final representations of drugs, proteins, side effects, and diseases respectively; based on H dr and H se , the prediction of drug-side effect associations can be simply implemented.

[0051] 3. Prediction Module

[0052] Given a pair of a drug and a side effect, the prediction probability of an association between them is derived from a fully connected layer with a Sigmoid function. Formally, the input pair is the i-th drug i ∈ {1, …, n dr} and the j-th side effect j ∈ {1, …, n se}, which are finally represented by H dr (i) and H se (j) respectively. Then, the probability of an association between the i-th drug and the j-th side effect is calculated as follows:

[0053]

[0054] where [·||·] represents the concatenation of two vectors, and W and b are the training parameters of the prediction module.

[0055] To optimize the proposed framework, cross-entropy is used as the loss function. More specifically, during the training phase, assume that the input mini-batch consists of N samples, denoted as {(X1,y1),…,(X n ,y n ),…,(X N ,y N ),}. For the nth sample, X n =(i n ,j n ), and y n is the ground truth label of X n . Then, the loss of the entire framework is calculated as follows:

[0056]

[0057] By minimizing the loss function, all parameters in the proposed model are optimized by the Adam optimizer and backpropagated in an end-to-end manner.

[0058] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A drug side effect prediction model based on a meta-path graph neural network, characterized in that, Including: An input module, a meta-path based feature learning module, and a prediction module; Input module: In the input module, a heterogeneous information network based on multiple datasets is constructed, where these datasets involve heterogeneous biological entities and the associations between them; Meta-path based feature learning module: In the meta-path based feature learning module, L layers of meta-path based information aggregation are utilized to obtain meaningful representations of heterogeneous nodes by fully capturing the complex semantics in the heterogeneous information network based on multiple datasets; Prediction module: The predicted drug-side effect associations are obtained by a fully-connected neural network for the representations of drugs and side effects; Given a pair of a drug and a side effect, the predicted probability of an association between them is derived from a fully connected layer with a sigmoid function. Formally, the input pair is the i-th drug \(i\in\{1,\ldots,n\}\) dr} and the j-th side effect \(j\in\{1,\ldots,n\}\) se}, which are finally represented by \(H\) dr (i) and \(H\) se (j) respectively. Then, the probability of an association between the i-th drug and the j-th side effect is calculated as follows: where [·||·] represents the concatenation of two vectors, and W and b are the training parameters of the prediction module, To optimize the proposed framework, cross-entropy is used as the loss function. More specifically, during the training phase, assume that the input mini-batch consists of N samples, denoted as {(X1,y1),…,(X n ,y n ),…,(X N ,y N ),}. For the nth sample, X n =(i n ,j n ), y n is the true label of X n . Then, the loss of the entire framework is calculated as follows: By minimizing the loss function, all parameters in the proposed model are optimized by the Adam optimizer and backpropagated in an end-to-end manner.

2. The drug side effect prediction model based on the meta-path graph neural network according to claim 1, wherein, Four types of entities, namely drugs, proteins, side effects, and diseases, are used to construct the heterogeneous information network based on multiple datasets.

3. The drug side effect prediction model based on the meta-path graph neural network according to claim 2, wherein The meta-path based feature learning module includes L layers of meta-path based information aggregation, and each layer consists of two parts, namely meta-path based information propagation and meta-path level attention mechanism; Meta-path based information propagation: When a specific node and a meta-path are given, the meta-path based information propagation can extract the information of the neighbors of this node based on the meta-path and aggregate it. This module also introduces the node embedding of the previous layer as a residual connection; Meta-path level attention mechanism: When multiple meta-paths are given, multiple embeddings of the same node can be obtained through the meta-path based information propagation, and the meta-path level attention mechanism automatically learns the weights of these embeddings.

4. The drug side effect prediction model based on the meta-path graph neural network according to claim 3, wherein, For the prediction module, given a pair of a drug and a side effect, the predicted probability of their association is derived from a fully-connected layer with a Sigmoid function.

Images