A drug feature representation method based on interactive representation learning

Abstract

The present invention provides a drug feature representation method based on interactive representation learning. Through the bidirectional cyclic interaction mechanism of heterogeneous graph representation learning and molecular graph representation learning, the multi-scale drug feature representation is fully integrated and utilized, so that the multi-scale drug features of the drug are complementary and integrated, and the function of obtaining the drug feature representation from the multi-scale drug description data is realized. The present invention targets the rich drug description data, recycles the macroscopic and microscopic features of the drug, and simultaneously carries out heterogeneous graph representation learning and molecular graph representation learning of the drug, fully utilizing the feature interaction of the drug at different scales, integrating the macroscopic dimension data and microscopic dimension data of the drug, and obtaining the drug feature representation from the drug description data, which has important guiding significance for downstream drug discovery and drug design.

Description

Technical Field

[0001] The present invention belongs to the field of graph representation learning and biomedicine technology, and specifically relates to a drug feature representation method based on interactive representation learning. Background Art

[0002] Drugs are substances that humans use to prevent, treat, and diagnose diseases. In reality, there is a large amount of drug-related data information, such as biomedical texts, electronic health records, reporting systems, and public databases (DrugBank, PubChem, KEGG, etc.), which contain a rich variety of drug physical and chemical data, drug target data, drug-drug interaction data, drug metabolic pathway data, drug protease data, and drug structural data. These data can reflect the multi-scale characteristics of drugs. For example, drug target data and protease data can reflect certain specific functional characteristics of drugs, while drug structural data contains the biological activity of drugs and can explain the characteristics of drugs from a microscopic perspective. Therefore, the description data of drugs is roughly divided into macro-dimensional data and micro-dimensional data. Macro-dimensional data includes biochemical description data such as drug targets, proteases, and metabolic pathways, and micro-dimensional data includes the characteristics and structural information of multi-atoms and multi-bonds inside drug molecules. The macro-feature representation and micro-feature representation of drugs represent description information from different perspectives, and there is a certain degree of complementarity and redundancy. Obtaining the feature representation of drugs from multi-scale drug description data has important guiding significance for downstream drug discovery and drug design. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a drug feature representation method based on interactive representation learning, which is used to obtain drug feature representation from multi-scale drug description data.

[0004] The technical solution adopted by the present invention to solve the above technical problems is: a drug feature representation method based on interactive representation learning, comprising the following steps:

[0005] S1: Collect multi-scale descriptive data of drugs;

[0006] S2: Extracting multi-scale features of drugs, including using heterogeneous graph representation to learn to extract macroscopic features of drugs, and using molecular graph representation to learn to extract microscopic features of drugs;

[0007] S3: Use a bidirectional recurrent interactive learning mechanism with multi-scale data to interactively learn multi-scale feature representations of drugs.

[0008] According to the above scheme, in step S1, the specific steps are:

[0009] S11: Obtain drug interaction data, drug target data, and drug enzyme data from the drug database DrugBank, and use the ID mapping server to map the drug target to the drug database KEGG DRUG to obtain the drug metabolic pathway data;

[0010] S12: Obtain drug-protein interaction data from the drug database BioGRID, construct the above drug description data and various types of relationships between them into a heterogeneous drug association network, and obtain macro-dimensional data of drugs;

[0011] S13: Obtain the canonical SMILES representation of each drug molecule from the drug database PubChem and perform preprocessing, convert the SMILES string into the representation of the molecular structure graph and graph data structure in turn, use nodes and edges to represent atoms and chemical bonds respectively, and obtain the microscopic dimensional data of the drug.

[0012] According to the above scheme, in step S2, the specific steps of extracting the macroscopic characteristics of the drug are:

[0013] S21: Constructing representation learning models for heterogeneous graphs;

[0014] Assume that the set of all node types is O, and each node v i ∈V is divided into specific node types; the set of all link types is R, and each link e∈E is divided into a specific link type; let the heterogeneous network describing the drug G = (V, E) be related to the node type mapping function ε: V→O and the link type mapping function ψ: E→R; the drug interaction data, target, enzyme and metabolic pathway characteristics are constructed as a heterogeneous association network of drugs;

[0015] The heterogeneous graph representation learning model of drugs is used to learn a feature representation, i.e., a mapping function, for each node:

[0016]

[0017] Assume |R|>1, E r Contains all links of type r∈R, then:

[0018] E=∪ r∈R E r ,

[0019] Divide the heterogeneous network into multiple sub-networks according to the type of links:

[0020] G r =(V,E r );

[0021] A unified low-dimensional vector representation is learned for each node on each link type, and the macro feature representation of the node is obtained by integration;

[0022] Let node v i The set of neighbor nodes on link type r is N i，r , the initial vector representation of each node on link type r is obtained by random initialization Let aggregator function be an aggregation function, used to aggregate information using methods including averaging and pooling; update function be an update function;

[0023] Aggregate and update the neighbor information of node vi on link type r to obtain the k-th order vector represented as The update formula is defined as follows:

[0024]

[0025]

[0026] After K iterations, node v i The vector representation u on link type r i，r Equivalent to Let M be the number of link types, and the splicing node v i The vector representation of all link types is used to obtain the feature vector representation U of the node. i :

[0027]

[0028] S22: A graph neural network framework based on the attention mechanism is used to aggregate information of different types of edges through the attention mechanism to extract macro-feature representations of drug nodes;

[0029] Assume w r and W r For link type r, it is a trainable parameter, and the self-attention mechanism is used to calculate the feature vector representation U of the node i The coefficients of the linear combinations of That is, link type r to node v i The attention weights are:

[0030]

[0031] Assume D r is a trainable transformation matrix, then node v i The vector representation on link type r is:

[0032]

[0033] Then node vi The macroscopic features of drugs learned on the heterogeneous graph are represented as:

[0034]

[0035] Furthermore, in step S21, the update function includes:

[0036] (a) Inner product, which is used to multiply the features of a node and its neighbors to update the node hidden state. Let ⊙ represent the element-by-element multiplication operator:

[0037]

[0038] (b) Gated recurrent unit, which is used to select a gated recurrent unit to obtain a vector representation with strong expressive power. Let GRU represent the gated recurrent unit:

[0039]

[0040] (c) Multilayer perceptron, which is used to concatenate the information of the node and its neighbor nodes and input them into the multilayer perceptron:

[0041]

[0042] Furthermore, in step S2, the specific steps of extracting the microscopic characteristics of the drug are:

[0043] S25: Constructing representation learning models for molecular graphs;

[0044] Let the node feature be x v , The link feature is e v，w ∈E′, then the molecular graph describing the drug is:

[0045] G′=(V′, E′);

[0046] The microscopic molecular graph representation learning model is used to learn the microscopic feature representation of the molecular graph for each drug node, that is, to learn a function:

[0047] g δ ::V′→Z;

[0048] S26: Adopt message passing mechanism to update the feature representation of nodes and links simultaneously;

[0049] In the message passing mechanism, the link hidden state h(e v，w ), node hidden state h(v), link message m(e v，w ) and node message m(v), assuming that the message passing process iterates T times, the current message passing iteration depth is t, and each iteration is based on the node’s intermediate message mt (v) and the linked intermediate message m t (e v，w ) Update the current hidden state h of the node t (v) and the current hidden state h of the link t (e v，w );

[0050] Assume that the neighbor nodes of node v in graph G′ are N(v), and through x v Assign the vector representation h of the initial hidden state of each node 0 (v) by Assign the vector representation h of the initial hidden state of each link 0 (e u，v ), the message vector is obtained by aggregating the neighbor nodes of node v, and the message vector is combined with the hidden state of node v at the previous moment to update the current hidden state h of each node in the molecular graph t (v):

[0051]

[0052] h t (v) = update(m t (v) h t-1 (v)) (10);

[0053] Through the node hidden state h t (v) Subtract the reverse link information to get the linked message vector m t (e v，w );Let the weight matrix of the fully connected network be W m , the message vector of the link is input into the fully connected network and added to the initial state of the link. After conversion by the activation function σ(·), the current hidden state h of each link in the molecular graph is updated. t (e v，w ):

[0054] m t (e v，w )=h t (v)-h t-1 (e w，v ) (11),

[0055] h t (e v，w )=σ(h 0 (e v，w )+W m ·m t (e v，w )) (12);

[0056] S27: Construct a message passing neural network to learn the vector representation of drug molecular graphs and obtain the microscopic feature representation of drugs;

[0057] Use the readout function to calculate the vector representation z of the entire molecular graph G′ i :

[0058]

[0059] After obtaining the final vector representation h(v) of each node v, the vector representation z of the entire drug molecule graph is obtained through the readout function. i , and obtain the microscopic characteristics of the drug.

[0060] Furthermore, in step S27, the update function includes:

[0061] (a) Summation, which is used to add all node vector representations to obtain the vector representation of the molecular graph:

[0062]

[0063] (b) Multilayer perceptron, which is used to input the node vector representation h(v) into the multilayer perceptron and add the output results to obtain the vector representation of the molecular graph:

[0064]

[0065] (c) Multi-scale readout function, used to convert the vector representation h of each node into t (v) The result is input into a multi-layer perceptron and summed up. Finally, the output results of each iteration are concatenated to obtain the vector representation of the molecular graph:

[0066]

[0067] Furthermore, in step S3, the specific steps are:

[0068] In the model of heterogeneous graph representation learning, the drug micro-representation learned in each layer of iteration interacts bidirectionally with the drug macro-representation in each layer of iteration. The number of iterations K of heterogeneous graph representation learning is equal to the number of iterations T of molecular graph representation learning.

[0069] The drug-drug interaction network G is a subnetwork of the heterogeneous network. d Each drug node on the network G obtains a positive sample set P and a negative sample set N. d The 1st to kth order neighbor nodes of a drug node on are positive samples of the drug node, otherwise they are negative samples;

[0070] Assume that the microscopic feature dimension of the molecular graph is m. In the process of learning the heterogeneous graph representation of drugs, the molecular graph representation of drugs is regarded as a type of link R′=R∪m. According to the information aggregation and update of nodes described in formulas (1) and (2), the drug-drug interaction network G is added. d The aggregation and updating process of the molecular feature representation of drug nodes in i In G d The neighbor nodes in i，m , the micro-feature representation h obtained through iterative learning k (v i ) is assigned to each drug node v i Vector representation on link type m Get drug node v i Vector representation at the kth iteration on link type m

[0071]

[0072]

[0073] Assume r′∈R′, M′=M+1, after K iterations, the splicing node v i The vector representation u on each link type i，r The following comprehensive vector representation is obtained:

[0074]

[0075] Using the attention mechanism of formulas (7) and (8), in the bidirectional loop interaction mode, we can get the node v i The node vector learned on the drug heterogeneous graph is

[0076] Let C = |P| + |N|, and It uses the discriminator and The mutual information estimation function of the model is used, and the heterogeneous graph representation learning mapping function of the drug k-th iteration is The mapping function for molecular graph representation learning is Adopt the contrastive learning objective function, use the drug feature representation learned in each layer of iteration for contrastive learning, and design the objective function of the k-th layer iteration

[0077]

[0078] set up represents the empirical probability distribution, x 1 , x 2 is the input sample, x2 ′ is sampled at The input sample, sp(·) is the softplus function, and the mutual information estimation function is defined for:

[0079]

[0080] Define the mutual information estimation function for:

[0081]

[0082] After K iterations, the objective function of the bidirectional loop interaction is This enables two-way interaction between macro and micro representation learning.

[0083] A computer storage medium stores a computer program executable by a computer processor, wherein the computer program executes a drug feature representation method based on interactive representation learning.

[0084] The beneficial effects of the present invention are:

[0085] 1. The present invention discloses a drug feature representation method based on interactive representation learning, which combines graph representation learning and attention mechanism, designs a multi-scale drug feature extraction method based on heterogeneous graph and molecular graph representation learning, and proposes an interactive learning mechanism for multi-scale drug features, which fully integrates the features of drugs at different scales and realizes the function of obtaining drug feature representation from multi-scale drug description data.

[0086] 2. The present invention targets heterogeneous associated data of drugs, and learns the macro-feature representation of drugs through heterogeneous node aggregation and attention mechanism to eliminate the noise impact brought by multi-feature data; for the molecular structure data of drugs, the message passing mechanism between atoms and chemical bonds is used to learn the feature representation of molecular graphs.

[0087] 3. The present invention targets abundant drug description data, recycles the macroscopic and microscopic features of drugs, and simultaneously carries out heterogeneous graph representation learning and molecular graph representation learning of drugs, fully utilizing the feature interactions of drugs at different scales, integrating the macroscopic dimensional data and microscopic dimensional data of drugs, and obtaining the feature representation of drugs from drug description data, which has important guiding significance for downstream drug discovery and drug design. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 It is a multi-scale drug description data graph of an embodiment of the present invention.

[0089] Figure 2 It is a representation learning model diagram of a heterogeneous graph according to an embodiment of the present invention.

[0090] Figure 3 It is a representation learning model diagram of the molecular graph according to an embodiment of the present invention.

[0091] Figure 4 It is a schematic diagram of a bidirectional cyclic interaction representation learning mode according to an embodiment of the present invention.

[0092] Figure 5 is a flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0093] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0094] See also Figure 5 A method for representing drug features based on interactive representation learning according to an embodiment of the present invention comprises the following steps:

[0095] First, the preparation of drug-related data is introduced, and then the extraction scheme of multi-scale drug features and the interaction mechanism of multi-scale drug features are explained.

[0096] 1. Collect multi-scale description data of drugs

[0097] The data used in the present invention mainly come from the rich drug description data in public drug databases (such as DrugBank, PubChem, KEGG DRUG and BioGRID, etc.), including drug interaction data, drug target data, drug metabolic pathway data, drug protease data and drug structure data, etc.

[0098] First, drug-drug interaction data, drug target data, and drug enzyme data were obtained from DrugBank, and the drug targets were mapped to the KEGG DRUG database using the ID mapping server to obtain drug metabolic pathway data. Secondly, drug-protein interaction data were obtained from the BioGRID database. The above drug description data and various types of relationships between them were constructed into a heterogeneous drug association network to obtain macro-dimensional drug data. Then, the standardized SMILES representation of each drug molecule was obtained from PubChem. For the preprocessing of the SMILES string, the RDKit toolkit software was used to convert the SMILES string into a molecular structure graph. The drug molecular structure graph was converted into a representation of a graph data structure, with nodes and edges representing atoms and chemical bonds, respectively, to obtain micro-dimensional drug data. Finally, multi-scale drug description data (including 3535 drugs, 2411 targets, 285 enzymes, and 314 metabolic pathways) were obtained.

[0099] 2. Extracting multi-scale features of drugs

[0100] In view of the numerous drug description data collected, the present invention adopts a graph representation learning method to extract feature representations that reflect the multi-scale characteristics of drugs, including a drug macro-feature extraction method based on heterogeneous graph representation learning and a drug micro-feature extraction method based on molecular graph representation learning.

[0101] ① First, it is necessary to integrate the macro-dimensional data of drugs and eliminate the noise impact caused by multi-feature data. To this end, the macro-feature representation of drugs is learned through heterogeneous node aggregation and attention mechanism.

[0102] ② Secondly, microscopic dimensional data such as the molecular graph of drugs provide a new perspective. It is necessary to analyze the structural characteristics of atoms and chemical bonds in molecules and explore the message transmission mechanism between atoms and chemical bonds to learn the feature representation of molecular graphs, that is, the microscopic feature representation of drugs.

[0103] For the extraction of macroscopic features of drugs, the present invention designs a representation learning model for heterogeneous graphs, adopts a graph neural network framework based on an attention mechanism, and aggregates information of different types of edges through the attention mechanism to learn the macroscopic feature representation of drug nodes. For the extraction of microscopic features of drugs, the present invention designs a representation learning model for molecular graphs, adopts a message passing mechanism, simultaneously updates the feature representation of nodes and links, constructs a message passing neural network to learn the vector representation of drug molecular graphs, and finally obtains the microscopic feature representation of drugs.

[0104] (1) Extracting macroscopic features of drugs based on heterogeneous graph representation learning

[0105] The present invention designs a representation learning model for heterogeneous graphs, which can learn the macro-feature representation of drug nodes from the heterogeneous association network of drugs. Figure 2 As shown in the figure, the model adopts a graph neural network framework based on the attention mechanism, which aggregates information of different types of edges through the attention mechanism to learn the feature representation of drug nodes.

[0106] 1) Heterogeneous node aggregation mechanism

[0107] First, we describe a heterogeneous network of drugs G = (V, E), which is associated with a node type mapping function ε: V → O and a link type mapping function ψ: E → R. O and R represent the set of all node types and the set of all link types, respectively. i ∈V all belong to a specific node type. Similarly, each link e∈E is classified as a specific link type. This project constructs the drug interaction data, target, enzyme and metabolic pathway characteristics into a heterogeneous association network of drugs.

[0108] The heterogeneous graph representation learning model of drugs aims to learn a feature representation for each node, that is, to learn a mapping function. V→F. In the present invention, E=∪ r∈R E r , E r Contains all links of type r∈R (|R|>1). The heterogeneous network is divided into multiple sub-networks according to the type of links, namely G r =(V,E r ). Then, a unified low-dimensional vector representation is learned for each node on each link type, and finally integrated to obtain the macro feature representation of the node.

[0109] Node v on link type r i The k-th order vector representation of It is obtained by aggregating and updating the neighbor information of the node on link type r. The update formula is defined as follows:

[0110]

[0111]

[0112] Among them, N i，r is the set of neighbor nodes of node vi on link type r. The initial vector of each node on link type r is represented by It is obtained by random initialization. The aggregator function is an aggregation function, which can aggregate information by averaging, pooling, etc. The update function is an update function. In this invention, three design schemes of update functions are proposed:

[0113] (a) Inner Product: One idea for updating a node’s hidden state is to multiply the features of the node and its neighbors’ aggregate.

[0114]

[0115] Here, ⊙ represents the element-wise multiplication operator.

[0116] (b) Gated recurrent unit: Gated recurrent unit is selected to obtain a vector representation with stronger expressive power.

[0117]

[0118] Here, GRU stands for the Gated Recurrent Unit.

[0119] (c) Multilayer Perceptron: The information of the node and its neighboring nodes is concatenated and input into the multilayer perceptron.

[0120]

[0121] After K iterations, node v iThe vector representation u on link type r i，r Equivalent to Then concatenate the nodes v i The vector representation of all link types finally obtains the comprehensive vector representation Ui of the node.

[0122]

[0123] Where M is the number of link types.

[0124] 2) Attention Mechanism

[0125] After getting the node's feature vector representation U i After that, the self-attention mechanism is used to calculate U i The coefficients of the linear combinations of That is, link type r to node v i The attention weights are:

[0126]

[0127] Among them, w r and W r For link type r, it is a trainable parameter. Therefore, for node v i The vector on link type r can be expressed as:

[0128]

[0129] Among them, D r is a trainable transformation matrix.

[0130] Finally, node vi learns the macroscopic features of the drug on the heterogeneous graph as

[0131] (2) Extraction of drug micro-features based on molecular graph representation learning

[0132] The present invention designs a molecular graph-oriented representation learning model to learn the microscopic feature representation of drugs from the molecular structure data of drugs. Figure 3 As shown, unlike heterogeneous graph representation learning, this model adopts a message passing mechanism to simultaneously update the feature representations of nodes and links, and constructs a message passing neural network to learn the vector representation of drug molecule graphs.

[0133] 1) Message passing mechanism of molecular graph

[0134] First, describe the molecular graph of drugs G′=(V′, E′), where the node features are x v , The link feature is e v，w∈E′. The microscopic molecular graph representation learning model aims to learn the microscopic feature representation of the molecular graph for each drug node, that is, to learn a function, g δ :V′→Z. In the message passing mechanism, the link hidden state h(e v，w ), node hidden state h(v), link message m(e v，w ) and node message m(v). The message passing process iterates T times, and each iteration is based on the intermediate message m of the node and link. t (v), m t (e v，w ) Update the current hidden state h of nodes and links t (v) h t (e v，w ), t is the current message passing iteration depth.

[0135] The iterative process of message passing is as follows: First, the current hidden state h of each node in the molecular graph is obtained t (v). Specifically, the neighbor nodes of node v are aggregated to obtain a message vector. Afterwards, the message vector is combined with the hidden state of node v at the previous moment to update the hidden state of the node. The update formula is as follows:

[0136]

[0137] h t (v) = update(m t (u), h t-1 (v)) (10);

[0138] Where N(v) represents the neighbor nodes of node v in the graph G′. The aggregate function is an aggregation function, which can be aggregated by averaging, pooling, etc. The update function is an update function, which can be the update function proposed above. The vector representation of the initial hidden state of each node is h 0 (v) by x v The initial hidden state of each link is represented by the vector h 0 (e u，v )Depend on Assignment obtained.

[0139] Then, we need to get the current hidden state h of each link in the molecular graph t (e v，w ). Since the updated node hidden state h has been obtained in the above process t (v), so the linked message vector m t (e v，w ) can be represented by the node hidden state h t(v) is obtained by subtracting the information of the reverse link. Then the message vector of the link is input into the fully connected network and added to the initial state of the link. After conversion through the activation function, the current hidden state of the link is updated. The update formula is as follows:

[0140] m t (e v，w )=h t (v)-h t-1 (e w，v ) (11),

[0141] h t (e v，w )=σ(h 0 (e v，w )+W m ·m t (e v，w )) (12);

[0142] Among them, W m is the weight matrix of the fully connected network, and σ(·) is the activation function.

[0143] 2) Readout function of molecular graph

[0144] Use the readout function to calculate the vector representation zi of the entire molecular graph G′:

[0145]

[0146] After obtaining the final vector representation h(v) of each node v, the vector representation z of the entire drug molecule graph is i The microscopic characteristics of the drug are obtained through the readout function. The present invention proposes three design schemes of the readout function:

[0147] (a) Summation: Simply add up the vector representations of all nodes to get the vector representation of the molecular graph.

[0148]

[0149] (b) Multilayer Perceptron: After the node vector representation h(v) is input into the multilayer perceptron, the output results are added together to obtain the vector representation of the molecular graph.

[0150]

[0151] (c) Multi-scale readout function: In each iteration, the vector representation h of each node is t (v) The output is fed into a multi-layer perceptron and summed up. Finally, the output results of each iteration are concatenated to obtain the vector representation of the molecular graph.

[0152]

[0153] 3. Interactive learning of multi-scale feature representation of drugs

[0154] The macroscopic and microscopic feature representations of drugs represent descriptive information from different perspectives, and there is a certain degree of complementarity and redundancy. The macroscopic feature representation of drugs reflects certain functional properties of drugs, while the microscopic feature representation of drugs contains the biological activity of drug structures. It is crucial to make full use of the feature interactions of drugs at different scales to obtain the feature representation of drugs. After extracting the macroscopic and microscopic feature representations of drugs, a bidirectional cyclic interactive learning mechanism of multi-scale data is used, such as Figure 4 As shown, the macroscopic and microscopic characteristics of drugs are cyclically utilized, and a multi-scale feature representation of drugs is obtained based on a heterogeneous graph representation learning model and a molecular graph representation learning model. A bidirectional cyclic interaction mode of drug characteristics is proposed, and the macroscopic and microscopic characteristics of drugs are combined to realize a bidirectional interaction between macroscopic and microscopic representation learning in each iteration.

[0155] Bidirectional cyclic interaction is a combination of macro and micro features of drugs, and bidirectional interaction between macro and micro representation learning is achieved in each iteration. On the one hand, the macro features of drugs can provide the topological structure information of nodes for the micro feature learning of drugs, which helps to guide the learning of more effective molecular graph representations and achieve macro-to-micro interaction. On the other hand, by using the molecular graph feature representation of drugs as the initialization information of the heterogeneous graph representation learning model and constraining the heterogeneous graph representation learning process, micro-to-macro interaction can be achieved.

[0156] In the heterogeneous graph representation learning model, the drug micro-representation learned in each layer of iteration interacts with the drug macro-representation in each layer of iteration. Therefore, in this interactive mode, the number of iterations of heterogeneous graph representation learning is equal to the number of iterations of molecular graph representation learning, that is, K = T. Specifically, in the sub-network of the heterogeneous network (i.e., the drug-drug interaction network G d ) each drug node on G gets a positive sample set P and a negative sample set N. d The neighbor nodes of order 1 to k on the node are positive samples of the drug node, otherwise they are negative samples.

[0157] First, in the process of learning the heterogeneous graph representation of drugs, the molecular graph representation of drugs is regarded as a type of link type, that is, R′=R∪m, where m represents the microscopic feature dimension of the molecular graph. According to the information aggregation and update of nodes described in formulas (1) and (2), the drug-drug interaction network G is added. d The aggregation and updating process of the molecular feature representation of the drug node in . So we get the drug node v i On the link type m (i.e., the microscopic feature dimension of the molecular graph), the vector representation at the kth iteration is

[0158]

[0159]

[0160] Among them, N i，m Represents drug node v i In G d Neighbor nodes in each drug node v i Vector representation on link type m The micro-feature representation h obtained by iterative learning k (v i ) is assigned a value.

[0161] After K iterations, the node v i The vector representation u on each link type i，r After concatenation, we finally get the following comprehensive vector representation:

[0162]

[0163] In which, r′∈R′, M′=M+1. The attention mechanism of formulas (7) and (8) is also used. Finally, in the bidirectional recurrent interaction mode, node v i The node vector learned on the drug heterogeneous graph is represented as:

[0164]

[0165] The present invention adopts a contrastive learning objective function, uses the drug feature representation learned in each layer of iteration to perform contrastive learning, and designs the objective function of the kth layer iteration

[0166]

[0167] Among them, C=|P|+|N|. and It uses the discriminator and Mutual information estimation function for modeling. and The heterogeneous graph representation of the drug k-th iteration and the molecular graph representation of the learned mapping function are represented respectively.

[0168] For the mutual information estimation function The definition is as follows:

[0169]

[0170] For the mutual information estimation function The definition is as follows:

[0171]

[0172]

[0173] in, represents the empirical probability distribution. 1 , x 2 is the input sample, x 2 ′ is sampled at The input sample of , sp(·) is the softplus function.

[0174] After K iterations, the objective function of the bidirectional loop interaction is This enables two-way interaction between macro and micro representation learning.

[0175] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

Claims

1. A drug feature representation method based on interactive representation learning, characterized in that: The following steps are involved: S1: Collect multi-scale descriptive data of drugs; S2: Extracting multi-scale features of drugs, including using heterogeneous graph representation learning to extract macroscopic features of drugs, and using molecular graph representation learning to extract microscopic features of drugs; the specific steps of extracting macroscopic features of drugs are: S21: Constructing a representation learning model for heterogeneous graphs; Assume that the set of all node types is O, and each node v i ∈V is divided into specific node types; the set of all link types is R, and each link e∈E is divided into a specific link type; let the heterogeneous network describing the drug G = (V, E) be related to the node type mapping function ε: V→O and the link type mapping function ψ: E→R; the drug interaction data, target, enzyme and metabolic pathway characteristics are constructed as a heterogeneous association network of drugs; The heterogeneous graph representation learning model of drugs is used to learn a feature representation, i.e., a mapping function, for each node: Assume |R|>1, E r Contains all links of type r∈R, then: E=U r∈R AND r , Divide the heterogeneous network into multiple sub-networks according to the type of links: G r =(V,E r ); A unified low-dimensional vector representation is learned for each node on each link type, and the macro feature representation of the node is obtained by integration; Let node v i The set of neighbor nodes on link type r is N i,r , the initial vector representation of each node on link type r is obtained by random initialization Let aggregator function be an aggregation function, used to aggregate information using methods including averaging and pooling; update function be an update function; Link node v on type r i The neighbor information on link type r is aggregated and updated to obtain the k-th order vector represented as The update formula is defined as follows: After K iterations, the vector representation u of node vi on link type r i,r Equivalent to Let M be the number of link types, and the splicing node v i The vector representation of all link types is used to obtain the feature vector representation U of the node. i : S22: A graph neural network framework based on the attention mechanism is used to aggregate information of different types of edges through the attention mechanism to extract macro-feature representations of drug nodes; Assume w r and W r For link type r, it is a trainable parameter, and the self-attention mechanism is used to calculate the feature vector representation U of the node i The coefficients of the linear combinations of That is, link type r to node v i The attention weights are: Assume D r is a trainable transformation matrix, then node v i The vector representation on link type r is: Then node v i The macroscopic features of drugs learned on the heterogeneous graph are represented as: S3: Use a bidirectional recurrent interactive learning mechanism with multi-scale data to interactively learn multi-scale feature representations of drugs.

2. The method for drug feature representation based on interactive representation learning according to claim 1, characterized in that: In the step S1, the specific steps are: S11: Obtain drug interaction data, drug target data, and drug enzyme data from the drug database DrugBank, and use the ID mapping server to map the drug target to the drug database KEGG DRUG to obtain the drug metabolic pathway data; S12: Obtain drug-protein interaction data from the drug database BioGRID, construct the above drug description data and various types of relationships between them into a heterogeneous drug association network, and obtain macro-dimensional data of drugs; S13: Obtain the canonical SMILES representation of each drug molecule from the drug database PubChem and perform preprocessing, convert the SMILES string into the representation of the molecular structure graph and graph data structure in turn, use nodes and edges to represent atoms and chemical bonds respectively, and obtain the microscopic dimensional data of the drug.

3. The method for drug feature representation based on interactive representation learning according to claim 1, characterized in that: In step S21, the update function includes: (a) Inner product, which is used to multiply the features of a node and its neighbors to update the node hidden state. Let ⊙ represent the element-by-element multiplication operator: (b) Gated recurrent unit, which is used to select a gated recurrent unit to obtain a vector representation with strong expressive power. Let GRU represent the gated recurrent unit: (c) Multilayer perceptron, which is used to concatenate the information of the node and its neighbor nodes and input them into the multilayer perceptron:

4. The method for drug feature representation based on interactive representation learning according to claim 1, characterized in that: In step S2, the specific steps of extracting the microscopic characteristics of the drug are: S25: Constructing representation learning models for molecular graphs; Let the node feature be Xv, The link feature is Then the molecular graph describing the drug is: G′=(V′, E′); The microscopic molecular graph representation learning model is used to learn the microscopic feature representation of the molecular graph for each drug node, that is, to learn a function: g δ :V′→Z; S26: Adopt message passing mechanism to update the feature representation of nodes and links simultaneously; In the message passing mechanism, the link hidden state h(e v，w ), node hidden state h(v), link message m(e v，w ) and node message m(v), assuming that the message passing process iterates T times, the current message passing iteration depth is t, and each iteration is based on the node’s intermediate message m t (v) and the linked intermediate message m t (e v，w ) Update the current hidden state h of the node t (v) and the current hidden state h of the link t (e v，w ); Let the neighbor nodes of node v in graph G be N(v), through x v Assign the vector representation h of the initial hidden state of each node 0 (v) by Assign the vector representation h of the initial hidden state of each link 0 (e u,v ), the message vector is obtained by aggregating the neighbor nodes of node v, and the message vector is combined with the hidden state of node v at the previous moment to update the current hidden state h of each node in the molecular graph t (v): h t (v)=update(m t (v),h t-1 (v)) (10); Through the node hidden state h t (v) Subtract the reverse link information to get the linked message vector m t (e v,w );Let the weight matrix of the fully connected network be W m , the message vector of the link is input into the fully connected network and added to the initial state of the link. After conversion by the activation function σ(·), the current hidden state h of each link in the molecular graph is updated. t (e v，w ): m t (e v，w )=h t (v)-h t-1 (e w，v ) (11), h t (e v，w )=σ(h 0 (e v，w )+W m ·m t (e v，w )) (12); S27: Construct a message passing neural network to learn the vector representation of drug molecular graphs and obtain the microscopic feature representation of drugs; Use the readout function to calculate the vector representation z of the entire molecular graph G′ i : After obtaining the final vector representation h(v) of each node v, the vector representation z of the entire drug molecule graph is obtained through the readout function. i , and obtain the microscopic characteristics of the drug.

5. A drug feature representation method based on interactive representation learning according to claim 4, characterized in that: In step S27, the form of the readout function includes: (a) Summation, which is used to add all node vector representations to obtain the vector representation of the molecular graph: (b) Multilayer perceptron, which is used to input the node vector representation h(v) into the multilayer perceptron and add the output results to obtain the vector representation of the molecular graph: (c) Multi-scale readout function, used to convert the vector representation h of each node into t (v) The result is input into a multi-layer perceptron and summed up. Finally, the output results of each iteration are concatenated to obtain the vector representation of the molecular graph:

6. A drug feature representation method based on interactive representation learning according to claim 5, characterized in that: In the step S3, the specific steps are: In the model of heterogeneous graph representation learning, the microscopic representation of drugs learned in each layer of iteration is bidirectionally interacted with the macroscopic representation of drugs in each layer of iteration. The number of iterations K of heterogeneous graph representation learning is equal to the number of iterations T of molecular graph representation learning. The drug-drug interaction network G is a subnetwork of the heterogeneous network. d Each drug node on the network G obtains a positive sample set P and a negative sample set N. d The 1st to kth order neighbor nodes of a drug node on are positive samples of the drug node, otherwise they are negative samples; Assume that the microscopic feature dimension of the molecular graph is m. In the process of learning the heterogeneous graph representation of drugs, the molecular graph representation of drugs is regarded as a type of link R′=R∪m. According to the information aggregation and update of nodes described in formulas (1) and (2), the drug-drug interaction network G is added. d The aggregation and updating process of the molecular feature representation of drug nodes in i In G d The neighbor nodes in i，m , the micro-feature representation h obtained through iterative learning k (v i ) is assigned to each drug node v i Vector representation on link type m Get drug node v i Vector representation at the kth iteration on link type m Assume r′∈R′, M′=M+1, after K iterations, the splicing node v i The vector representation u on each link type i，r The following comprehensive vector representation is obtained: Using the attention mechanism of formulas (7) and (8), in the bidirectional loop interaction mode, we can get the node v i The node vector learned on the drug heterogeneous graph is Let C = |P| + |N|, and It uses the discriminator and The mutual information estimation function of the model is used, and the heterogeneous graph representation learning mapping function of the drug k-th iteration is The mapping function for molecular graph representation learning is Adopt the contrastive learning objective function, use the drug feature representation learned in each layer of iteration for contrastive learning, and design the objective function of the k-th layer iteration Let P represent the empirical probability distribution, x1, x2 are input samples, and x2′ is sampled at The input sample, sp(·) is the softplus function, and the mutual information estimation function is defined for: Define the mutual information estimation function for: After K iterations, the objective function of the bidirectional loop interaction is This enables two-way interaction between macro and micro representation learning.

7. A computer storage medium, characterized in that: A computer program executable by a computer processor is stored therein, and the computer program executes a drug feature representation method based on interactive representation learning as described in any one of claims 1 to 6.

Images