Substructure-based drug interaction prediction method and system

Abstract

The application belongs to the field of drug mutual prediction, and provides a drug mutual interaction prediction method and system based on a substructure, which comprises the following steps: obtaining a drug structure graph of a drug to be predicted; encoding all atomic information in the drug structure graph and determining the type information of each chemical bond in the drug; screening a plurality of atomic features as representative substructures of the corresponding drug based on the type information of the chemical bond; learning the interaction between unknown substructures from the interaction between known substructures based on the association between the drug substructures, obtaining drug interaction perception representation by using a zigzag iterative network according to the representative substructures of the corresponding drug; and calculating the probability of the interaction between the drugs to be predicted based on the representative substructures of the corresponding drug and the drug interaction perception representation, so as to solve the problem that it is difficult to directly train and learn accurate SSI information, fully and accurately extract substructure features, and improve the final DDI prediction accuracy.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of drug mutual prediction, and particularly relates to a drug mutual interaction prediction method and system based on substructure. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] Existing drug mutual interaction (DDI) prediction work considering substructure can be divided into two categories, one based on explicit substructure and the other based on implicit substructure.

[0004] The method based on explicit substructure needs a predefined substructure list. This kind of method predicts DDI by directly mining the interaction between predefined substructures. This predefined substructure list can be a manually defined molecular fingerprint or a frequently occurring SMILE substring extracted from the dataset through sequence pattern mining. The advantage of the method based on explicit substructure is that the learning process of DDI can be limited to specific substructures. However, its disadvantages are also very obvious:

[0005] (1) The manually defined molecular fingerprint is limited by the professional knowledge of experts in this field and does not have enough flexibility to discover what is beyond human definition; and for drugs in the early development stage, the manually defined molecular fingerprint may not exist.

[0006] (2) The non-manually defined molecular fingerprint, i.e. the SMILE substring extracted by sequence pattern mining, may have incomplete word structure information, i.e. the information of key substrings may be truncated or even mismatched, thereby introducing a lot of noise to the model.

[0007] The method based on implicit substructure does not need a predefined substructure list. The model generally extracts self-adaptive substructures of size and shape through graph neural network (GCN or GAT). This method does not need prior knowledge and can mine undefined substructures. However, if the substructure features cannot be extracted accurately enough, the accumulated errors will affect the final DDI prediction. Existing methods based on implicit substructure focus on optimizing the substructure extraction method, which can improve the model performance and achieve good results. However, these works ignore the correlation between substructure-substructure interactions (SSI). Since existing DDI datasets, such as Drugbank, do not contain substructure interaction labels, it is difficult to directly train and learn accurate SSI information. SUMMARY

[0008] To solve at least one technical problem in the above background art, the present application provides a substructure-based drug interaction prediction method, which can learn more SSI by using the correlation between SSI, and model and learn more interactions between substructures by a chemical bond-assisted zigzag iterative network, thereby predicting drug interactions.

[0009] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0010] The first aspect of the present application provides a substructure-based drug interaction prediction method, comprising the following steps:

[0011] Obtaining a drug structure graph of a drug to be predicted;

[0012] Encoding all atomic information in the drug structure graph and determining the type information of each chemical bond in the drug;

[0013] Based on the atomic information and the type information of the chemical bond, screening a plurality of atomic features as representative substructures of the corresponding drug;

[0014] Based on the representative substructures of the corresponding drug, using a zigzag iterative network to learn the interaction between unknown substructures from the interaction between known substructures according to the correlation between drug substructures, and obtaining a drug interaction perception representation;

[0015] According to the representative substructures of the corresponding drug and the drug interaction perception representation, calculating the probability of interaction between the drug to be predicted.

[0016] The second aspect of the present application provides a substructure-based drug interaction prediction system, comprising:

[0017] A structure graph acquisition module for obtaining a drug structure graph of a drug to be predicted;

[0018] An encoding module for encoding all atomic information in the drug structure graph and determining the type information of each chemical bond in the drug;

[0019] A substructure screening module for screening a plurality of atomic features as representative substructures of the corresponding drug based on the atomic information and the type information of the chemical bond;

[0020] A zigzag iteration module for learning the interaction between unknown substructures from the interaction between known substructures according to the correlation between drug substructures based on the representative substructures of the corresponding drug by using a zigzag iterative network, and obtaining a drug interaction perception representation;

[0021] an interaction prediction module configured to calculate a probability of an interaction between the drugs to be predicted based on the representative substructures of the corresponding drugs and the drug interaction perception representation.

[0022] A third aspect of the present application provides a computer readable storage medium.

[0023] A computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the substructure-based drug interaction prediction method according to the first aspect.

[0024] A fourth aspect of the present application provides a computer device.

[0025] A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the substructure-based drug interaction prediction method according to the first aspect when executing the program.

[0026] Compared with the prior art, the present application has the following beneficial effects:

[0027] 1. The prior art does not contain the label of the interaction between substructures, and it is difficult to directly learn accurate SSI information. The present application can capture the difficult-to-learn SSI by learning the relatively fuzzy interaction between substructures from the relatively clear interaction between substructures through the zigzag iterative network of the chemical bonds of drugs.

[0028] 2. The accumulated error will affect the final DDI prediction because the substructure features cannot be extracted accurately enough. The present application introduces the bond type when encoding the substructure to better learn the chemical properties of the drug, designs a simple and effective substructure encoder to encode each substructure into a representative vector, better extracts the substructure features, and can fully extract the substructure features to improve the accuracy of the final DDI prediction.

[0029] The advantages of the additional aspects of the present application will be partially given in the following description, partially become obvious from the following description, or be understood through the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0030] The accompanying drawings, which form a part of the present application, are intended to provide further understanding of the present application, and the illustrative embodiments of the present application and their description serve the purpose of explaining the present application. The present application should not be limited by the improper interpretation of the accompanying drawings.

[0031] Figure 1 is a substructure-based drug interaction prediction block diagram according to an embodiment of the present application;

[0032] Figure 2is a flow chart of a drug interaction prediction method based on substructure of an embodiment of the present application;

[0033] Figure 3 is a definition of a chemical bond type of an embodiment of the present application;

[0034] Figure 4 is a schematic diagram of a zigzag iterative network update of an embodiment of the present application. DETAILED DESCRIPTION

[0035] The present application is further described below in conjunction with the accompanying drawings and embodiments.

[0036] It should be noted that the following detailed description is merely exemplary and is intended to provide further description of the present application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application pertains.

[0037] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, components, and / or groups thereof, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof.

[0038] As mentioned in the background section, since the existing DDI dataset does not contain substructure interaction labels, it is difficult to directly train and learn accurate SSI information. The present application proposes a zigzag iteration module (ZI) that learns relatively ambiguous SSI from relatively clear SSI by using the association between SSI, thereby capturing SSI that is difficult to learn directly.

[0039] The zigzag update process based on the zigzag iteration module (ZI) is as follows: if the interaction between substructure i and substructure j (,j) is unknown, but the interaction between substructure i' and j (',) and the interaction between i and j' (,j') are known, then (,j) can be updated by (',) and (,j').

[0040] The three interaction edges (,j), (,j'), and (',) mentioned here can form a "zigzag", so it is called zigzag update. In the substructure interaction matrix, there are many such "zigzags", so the substructure interaction (,j) can be updated from multiple angles.

[0041] To solve the problem that the accumulated error will affect the final DDI prediction due to the inability to extract the substructure features accurately and sufficiently, the present application proposes a CBAGNN (chemical bond assisted graph neural network) to obtain more representative substructures, which is a graph neural network considering the type of chemical bond. In organic chemistry, the chemical bond in the functional group (i.e. substructure) plays an important role in SSI. For example, double bonds can undergo epoxidation and dihydroxylation reactions, and aromatic bonds can undergo halogenation, nitration and sulfonation reactions. Therefore, it is assumed that if the model can identify the chemical bond in the substructure, better DDI prediction performance can be obtained. In addition, in order to encourage adjacent substructures (whose central atoms are adjacent) to learn unique representations, a layer of RGCN + 3 layers of GAT is designed as the architecture of CBAGNN. Specifically, CBAGNN uses RGCN to explicitly consider all chemical bonds in the first layer, and uses GAT to implicitly consider chemical bonds in the second and subsequent layers. In this way, the representations of adjacent substructures learned from CBAGNN are not only different at the central atom, but also different in the weight of the chemical bond.

[0042] Embodiment one

[0043] With reference to Figures 1-2 The present embodiment provides a substructure-based drug interaction prediction method, comprising the following steps:

[0044] S101: Obtain a plurality of drug structure graphs G; wherein in the drug structure graph G, the node is an atom, and the edge is a chemical bond between atoms.

[0045] S102: Preprocess the given drug structure graph G, and encode all atom information into a matrix M 0 ∈R n *d determine the type of each chemical bond in the given drug by Huffman coding;

[0046] Wherein n and d respectively represent the number of atoms and the length of the input vector;

[0047] As shown in Figure 3 determine the type of each chemical bond in the given drug by Huffman coding (a total of 16 types), the specific process is as follows: use 4 binary bits to encode the type of chemical bond in the drug molecule, the first two binary bits are used to encode four types of chemical bonds: single bond, double bond, triple bond and aromatic bond. The third binary bit is used to encode whether the chemical bond is a conjugated bond, and the fourth binary bit is used to encode whether the chemical bond is a ring bond.

[0048] In the present embodiment, the encoding is performed by Rdkit (an open source toolkit for chemical informatics).

[0049] S103: Based on the type information of each chemical bond, encode each atomic feature into a vector using a substructure extraction model, and select a number of atomic vectors as representative substructures of the corresponding drug;

[0050] The substructure extraction model adopts a chemical bond assisted graph neural network, and the chemical bond assisted graph neural network comprises one layer of RGCN and three layers of GAT.

[0051] S301: Merge the type information of each chemical bond into its corresponding atomic representation using one layer of RGCN of the chemical bond assisted graph neural network; the formula is as follows:

[0052]

[0053] wherein, is the representation of the ith atom at the lth layer, and the initial representation of the ith atom is M 0 obtained when constructing the input of GBAGNN, the i-th row, t is one of the bond types, τ is the set of bond types, N t (i) is the adjacent atom connected to atom i through bond type t, W t (l) and W (l) are all learnable weights in the network.

[0054] S302: Update the atomic representation using three layers of GAT of the chemical bond assisted graph neural network; the formula is as follows:

[0055]

[0056] wherein, K is the number of attention heads, N(i) is the adjacent atom connected to atom i, is the feature of the adjacent atom is the weight of the adjacent atom.

[0057] S303: After the atomic information is updated in S302, the representation M l ∈R n*d of all atoms is obtained, and a pooling operation (SAGPooling) is used to calculate the score s of each atomic feature.

[0058] Then, set a hyperparameter k to determine the number of atomic features to be retained, select k atomic features from the scores s of all atomic features as the representative substructures of the drug, and represent them using M (sel) .

[0059] M () =op(s,k)∈R k*d

[0060] Since each atom vector represents a substructure after the node update, a pooling layer is used to filter out k representative atom vectors as k substructures.

[0061] The above scheme has the advantages that the key type is introduced when encoding the substructure to better learn the chemical properties of the drug, a simple and effective substructure encoder is designed to encode each substructure into a representative vector, better extract substructure features, and fully extract substructure features to improve the accuracy of final DDI prediction.

[0062] S104: Based on the representative substructure of the corresponding drug, using the zigzag iterative network, learning the interaction between unknown substructures from the interaction between known substructures according to the association between drug substructures, obtaining drug interaction perception representation;

[0063] The embodiment is described in detail taking the interaction between two drug substructures as an example;

[0064] In order to learn the SSI between two drugs, the zigzag iteration module first receives the substructure features of the two drugs and as input, and obtains the interaction perception substructure and through interaction initialization, zigzag update and substructure update.

[0065] S401: Perform interaction initialization on the substructures of the two drugs and to obtain the interaction matrix F, specifically:

[0066] 1. Expand the dimensions of the substructures and to k*k*d, denoted as:

[0067]

[0068]

[0069] 2. Obtain the interaction matrix F through element multiplication and concatenation operation, denoted as:

[0070]

[0071]

[0072] F=mlp([F (dot) ‖F(cat) ])∈R k*k*d

[0073] Here, mlp() is a multilayer perceptron layer, where each element F in the interaction matrix F... i,j represent The i-th substructure and The interaction between the j-th substructures is considered, but the interaction here does not take into account other SSIs.

[0074] By considering F i,j Associations with other SSIs (e.g., F) i,k and f k,j To update F, k≠i,j) i,j .like Figure 4 As shown in the figure, the left side is the interaction matrix F, and the right side is an example of two interactions being updated. The three interaction edges (,j), (,j-1), (-1,j) (or (,j), (,j+2), (+2,j)) can form a zigzag shape. Each zigzag shape indicates that the interaction in the middle (such as (,j)) can be updated by two adjacent interactions (such as (i,j-1) and (-1,j)).

[0075] Before S402 and S403, first obtain the query, key, and value required by the attention mechanism:

[0076]

[0077]

[0078]

[0079] Among them W q W k W v All of these are trainable parameter matrices.

[0080] S402: Using the interaction matrix F i,j linear transformation As a query, calculate the i-th row of the interaction matrix F (i.e., ) and the j-th column (i.e. Attention score:

[0081]

[0082]

[0083] In the above calculation of the attention score H in the i-th rowm in the formula of evaluate the relevance of interaction (i, j) and (i, m), quantify the distinctness of interaction (i, m).

[0084] S403: update F using the i-th row of interaction values and its attention score (the regularization result of m ) and the j-th column of interaction values and its attention score (the regularization result of m ) together to update F i,j get

[0085]

[0086]

[0087]

[0088] S404: take as input, and use nonlinear transformation and softmax calculation to get the cross-updating representation of the two drugs:

[0089]

[0090]

[0091]

[0092] S405: splice the cross-updating representation of the two drugs and the original representation and the element multiplication result of the two representations:

[0093]

[0094]

[0095] S406: filter information with a self-gated updating mechanism to get the final interaction-aware representation:

[0096]

[0097]

[0098] S105: based on the substructure of the given drug and the interaction-aware representation, calculate the probability of the interaction existing between the given drugs:

[0099]

[0100] wherein, The dimension of is k*k, which is used to evaluate the importance of each SSI, and R is the matrix after encoding the interaction type r.

[0101] The advantage of the above scheme is that the zigzag iterative network assisted by the chemical bond of the drug can learn the interaction between relatively vague substructures from the interaction between relatively clear substructures, which can capture the SSI that is difficult to learn directly.

[0102] Embodiment two

[0103] The embodiment provides a drug interaction prediction system based on substructures, which comprises:

[0104] A structure graph acquisition module is configured to acquire a drug structure graph of a drug to be predicted.

[0105] An encoding module is configured to encode all atomic information in the drug structure graph and determine the type information of each chemical bond in the drug.

[0106] A substructure screening module is configured to screen a plurality of atomic features as representative substructures of the corresponding drug based on the atomic information and the type information of the chemical bond.

[0107] A zigzag iterative module is configured to learn the interaction between unknown substructures from the interaction between known substructures based on the representative substructures of the corresponding drug by using a zigzag iterative network according to the association between drug substructures, so as to obtain a drug interaction perception representation.

[0108] An interaction prediction module is configured to calculate the probability of the interaction between the drugs to be predicted based on the representative substructures of the corresponding drug and the drug interaction perception representation.

[0109] Embodiment three

[0110] The embodiment provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the steps in the drug interaction prediction method based on substructures in the above embodiment one.

[0111] Embodiment four

[0112] The embodiment provides a computer device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor realizes the steps in the drug interaction prediction method based on substructures in the above embodiment one when executing the program.

[0113] Those skilled in the art will appreciate that embodiments of the application can be readily used as software, hardware, or a combination of software and hardware. In a software embodiment, the methods can be tangibly embodied in a computer-readable storage medium having stored

[0114] The present application is described in reference to the drawings using a flowchart illustration and / or a block diagram of the method, apparatus (system) and computer program product according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing system or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart Figure 1 one or more functions specified in the flowchart block or blocks. Figure 1 one or more functions specified in the flowchart block or blocks.

[0115] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more functions specified in the flowchart block or blocks. Figure 1 one or more functions specified in the flowchart block or blocks.

[0116] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more functions specified in the flowchart block or blocks. Figure 1 one or more functions specified in the flowchart block or blocks.

[0117] Those skilled in the art will appreciate that implementing all or part of the methods described above in the embodiments can be accomplished by means of a computer program that directs relevant hardware to perform the steps. The program can be stored in a computer readable storage medium, which can be a magnetic disk, optical disk, Read-Only Memory (ROM) or Random Access Memory (RAM), etc.

[0118] The above descriptions are only the preferred embodiments of the present application, not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the scope of the present application.

Claims

1. A drug interaction prediction method based on substructure, characterized in that, Includes the following steps: Obtain the drug structure diagram of the drug to be predicted; Encode all atomic information in the drug structure diagram and determine the type of each chemical bond within the drug; Based on atomic information and chemical bond type information, several atomic features were selected as representative substructures of the corresponding drugs. Based on the representative substructures of the corresponding drugs, a zigzag iterative network is used to learn the interactions between unknown substructures from the interactions between known substructures, based on the associations between drug substructures, to obtain a drug interaction perception representation. Based on the representative substructure of the corresponding drug and the drug interaction perception representation, the probability of an interaction between the drugs to be predicted is calculated. The method utilizes a zigzag iterative network to learn the interactions between unknown substructures from the interactions between known substructures, based on the associations between drug substructures, to obtain a drug interaction perception representation, including the following steps: Interaction matrices are obtained by initializing the substructures of the two drugs to identify their interactions. A linear transformation of the interaction matrix is ​​used as a query to calculate the attention scores for rows and columns in the interaction matrix; Update the corresponding action matrix based on the attention scores in the rows and columns of the interaction matrix; Based on the updated corresponding action matrix, a self-gating mechanism is used to filter information to obtain the final drug interaction perception representation.

2. The drug interaction prediction method based on substructure as described in claim 1, characterized in that, Huffman coding is used to determine the type of each chemical bond within a drug.

3. The drug interaction prediction method based on substructure as described in claim 2, characterized in that, The method of determining the type of each chemical bond in a drug through Huffman coding includes: using 4 binary bits to encode the type of chemical bond in the drug molecule, the first two binary bits are used to encode four types of chemical bonds: single bond, double bond, triple bond and aromatic bond; the third binary bit is used to encode whether the chemical bond is a conjugated bond and the fourth binary bit is used to encode whether the chemical bond is a cyclic bond.

4. The drug interaction prediction method based on substructure as described in claim 1, characterized in that, A chemical bond-assisted graph neural network is used to screen atomic features. The chemical bond-assisted graph neural network includes a single layer RGCN and three layers GAT. The single layer RGCN is used to merge the type information of each chemical bond into its corresponding atomic representation, and the three layers GAT are used to update the atomic representation.

5. The drug interaction prediction method based on substructure as described in claim 4, characterized in that, The method of using a single-layer RGCN to merge the type information of each chemical bond into its corresponding atomic representation is as follows: in, It is the first The atom in the first Layer representation, It refers to the type of chemical bond. It is a set of chemical bond types. It is with atoms By chemical bond type Adjacent atoms connected to each other, and All of these are learnable weights in the network.

6. The drug interaction prediction method based on substructure as described in claim 4, characterized in that, After updating the atomic representation, a pooling operation is used to calculate the score of each atomic feature. The number of atomic features to be retained is determined according to the set hyperparameters. The corresponding number of atomic features are selected from the scores of all atomic features as the representative substructure of the drug.

7. A drug interaction prediction system based on substructure, characterized in that, include: The structure diagram acquisition module is used to acquire the drug structure diagram of the drug to be predicted. The encoding module is used to encode all atomic information in the drug structure diagram and determine the type of each chemical bond within the drug; The substructure screening module is used to screen out several atomic features as representative substructures of the corresponding drug based on atomic information and chemical bond type information. The zigzag iteration module is used to learn the interactions between unknown substructures from the interactions between known substructures based on the representative substructures of the corresponding drug using a zigzag iteration network, and obtain a drug interaction perception representation based on the associations between drug substructures. The interaction prediction module is used to calculate the probability that there is an interaction between the drugs to be predicted, based on the representative substructure of the corresponding drug and the drug interaction perception representation. The method utilizes a zigzag iterative network to learn the interactions between unknown substructures from the interactions between known substructures, based on the associations between drug substructures, to obtain a drug interaction perception representation, including the following steps: Interaction matrices are obtained by initializing the substructures of the two drugs to identify their interactions. A linear transformation of the interaction matrix is ​​used as a query to calculate the attention scores for rows and columns in the interaction matrix; Update the corresponding action matrix based on the attention scores in the rows and columns of the interaction matrix; Based on the updated corresponding action matrix, a self-gating mechanism is used to filter information to obtain the final drug interaction perception representation.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the substructure-based drug interaction prediction method as described in any one of claims 1-6.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the drug interaction prediction method based on substructure as described in any one of claims 1-6.

Images