Drug-based cancer inhibition sensitivity prediction method and device based on graph neural network

Abstract

Disclosed in the present invention are a drug-based cancer inhibition sensitivity prediction method and device based on a graph neural network. The method comprises: acquiring a candidate drug and cancer cell line original data; extracting drug information features of the candidate drug, and expanding the drug information features; constructing a drug molecule feature vector according to the drug information features, and constructing a drug adjacency matrix, a node feature matrix, and an edge feature matrix by taking atoms of a drug as nodes and chemical bonds as edges; constructing and iterating a graph neural network model to obtain a drug graph feature; obtaining a drug characterization after the drug graph feature and the drug molecule feature vector are aggregated; acquiring and storing the expression quantities of gene expression, gene mutation, gene methylation and gene copy number, in a cancer cell line, of a gene sequence, and extracting a cancer cell line characterization; and performing feature fusion on the drug characterization and the cancer cell line characterization to obtain a drug-cancer cell line instance pair, inputting the instance pair into a drug-based cancer inhibition sensitivity prediction model, and predicting to obtain an IC50 value of the candidate drug in a cancer cell line environment.

Description

A method and device for predicting drug cancer inhibition sensitivity based on graph neural network Technical Field

[0001] The present invention mainly relates to the field of artificial intelligence drug research and development, and specifically to a method and device for predicting drug cancer inhibition sensitivity based on graph neural networks. Background Art

[0002] Thanks to the rapid development of deep learning algorithms and the emergence of a large number of cancer cell lines and their whole genome information, data-driven cancer drug sensitivity prediction is an inevitable technological trend to accelerate cancer drug research and development. It can effectively alleviate the problems of long development cycles, high costs, and low efficiency of traditional cancer drugs. It can provide key guidance for doctors' clinical drug use and contribute to personalized cancer treatment. Drug anticancer sensitivity prediction based on deep learning is currently a hot topic. This type of method inputs drug and cancer cell line data into a deep learning model to form a "drug-cancer cell line" instance pair, and uses a deep learning algorithm to extract features and mine the internal connections between the instance pairs to predict drug anticancer sensitivity. However, the current deep learning method for drug anticancer sensitivity has the following limitations:

[0003] (1) Drug data acquisition is difficult, drug feature information is incomplete, and drug features are sparse, making drug feature extraction not rich enough.

[0004] (2) Cancer cell lines are difficult to characterize, and being represented by only a small amount of cancer cell line omics data will lead to insufficient feature extraction of cancer cell lines.

[0005] Therefore, there is an urgent need to propose a method for predicting drug tumor suppressor sensitivity to obtain richer information about drugs and cancer cell lines and extract more accurate "drug-cancer cell line" characterization.

[0006] Summary of the Invention

[0007] In response to the shortcomings of the existing technology, the present invention proposes a method and device for predicting drug cancer inhibition sensitivity based on graph neural network.

[0008] According to a first aspect of an embodiment of the present invention, a method for predicting drug cancer inhibition sensitivity based on a graph neural network is provided, the method comprising:

[0009] Obtain original data on drug candidates and cancer cell lines;

[0010] Extract drug information features of candidate drugs and expand them; construct drug molecular feature vectors based on drug information features, and construct drug adjacency matrix, node feature matrix, and edge feature matrix using drug atoms as nodes and chemical bonds as edges;

[0011] A graph neural network model is constructed based on the drug adjacency matrix, node feature matrix, and edge feature matrix. The graph neural network model is iterated to obtain drug graph features. The drug graph features are aggregated with the drug molecule feature vector to obtain the drug representation.

[0012] Obtain gene expression, gene mutation, gene methylation, and gene copy number of gene sequences in cancer cell lines, and extract cancer cell line characteristics;

[0013] Perform feature fusion on drug representation and cancer cell line representation to obtain drug-cancer cell line instance pairs;

[0014] The drug-cancer cell line example pairs are input into the drug tumor suppressor sensitivity prediction model to predict the IC50 value of the candidate drug in the cancer cell line environment.

[0015] According to a second aspect of an embodiment of the present invention, a device for predicting drug cancer inhibition sensitivity based on a graph neural network is provided, the device comprising: a multi-source data integration module, a storage module, and a processing module;

[0016] The multi-source data integration module is used to obtain raw data of candidate drugs and cancer cell lines; extract and expand the drug information features of candidate drugs; construct drug molecular feature vectors based on drug information features, and construct drug adjacency matrices, node feature matrices, and edge feature matrices using drug atoms as nodes and chemical bonds as edges;

[0017] The storage module includes: a drug information database and a cancer cell line information database;

[0018] Drug information database, used to store drug molecule feature vectors, drug adjacency matrix, node feature matrix and edge feature matrix;

[0019] Cancer cell line information database, used to store gene expression, gene mutation, gene methylation, and gene copy number expression levels of gene sequences in cancer cell lines;

[0020] The processing modules include: drug characterization extraction unit, cancer cell line characterization extraction unit, drug-cancer cell line instance pair acquisition unit, and drug tumor suppression sensitivity prediction unit;

[0021] Drug characterization extraction unit: A graph neural network model is constructed based on the drug adjacency matrix, node feature matrix, and edge feature matrix. The graph neural network model is iterated to obtain drug graph features. The drug graph features are aggregated with the drug molecular feature vector to obtain the drug characterization.

[0022] Cancer cell line characterization extraction unit, which extracts cancer cell line characterization by obtaining the expression levels of gene expression, gene mutation, gene methylation, and gene copy number of gene sequences in cancer cell lines;

[0023] a drug-cancer cell line instance pair acquisition unit, configured to perform feature fusion on drug representations and cancer cell line representations to obtain drug-cancer cell line instance pairs;

[0024] The drug tumor suppressor sensitivity prediction unit predicts the IC50 value of the candidate drug in the cancer cell line environment by inputting the drug-cancer cell line instance pair into the drug tumor suppressor sensitivity prediction model.

[0025] According to a third aspect of an embodiment of the present invention, an electronic device is provided, comprising one or more processors, for implementing the above-mentioned graph neural network-based drug cancer inhibition sensitivity prediction method.

[0026] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which a program is stored. When the program is executed by a processor, it is used to implement the above-mentioned drug cancer inhibition sensitivity prediction method based on graph neural network.

[0027] The beneficial effects of the present invention are as follows: the present invention provides a method and device for predicting drug tumor suppressor sensitivity based on graph neural network, which extracts drug information features of candidate drugs and expands them to address the difficulties in obtaining drug data, incomplete drug feature information, and sparse drug features, integrating major drug information data sources, breaking the bottlenecks of few drug information features from a single data source, missing drug information features from some drugs, fully capturing drug feature information, and increasing the dimension of drug information features; in view of the difficulty in characterizing cancer cell lines, the method accurately represents cancer cell lines by the expression of gene expression, gene mutation, gene methylation, and gene copy number in cancer cell lines through gene sequences, and encodes them through Transformer networks and feedforward neural networks to extract cancer cell line characterization. The present invention performs feature fusion on drug characterization and cancer cell line characterization to obtain drug-cancer cell line example pairs, and significantly improves the prediction accuracy of drug tumor suppressor sensitivity through more accurate and richer drug-cancer cell line example pairs. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0029] FIG1 is a flow chart of a method for predicting drug cancer inhibition sensitivity based on a graph neural network according to an embodiment of the present invention;

[0030] FIG2 is a flowchart of obtaining drug characterization according to an embodiment of the present invention;

[0031] FIG3 is a flowchart of a method for obtaining a characterization of a cancer cell line according to an embodiment of the present invention;

[0032] FIG4 is a schematic diagram of a device for predicting drug cancer inhibition sensitivity based on a graph neural network according to an embodiment of the present invention;

[0033] FIG5 is a schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0035] It should be noted that, unless there is any conflict, the features in the following embodiments and implementations may be combined with each other.

[0036] In order to solve the problems of troublesome drug data acquisition, incomplete drug feature information, sparse drug features and difficulty in characterizing cancer cell lines, the embodiment provides a method and device for predicting drug tumor suppressor sensitivity based on graph neural network, which integrates major drug information data sources, breaks the bottleneck of few drug information features of a single data source, missing some drug information features, etc., fully captures drug feature information, and increases the dimension of drug information features; accurately represents cancer cell lines through the expression of gene expression, gene mutation, gene methylation, and gene copy number of gene sequences in cancer cell lines, and encodes them through Transformer and feedforward neural networks. The present invention significantly improves the prediction results of drug tumor suppressor sensitivity by extracting more accurate and richer "drug-cancer cell line" representations.

[0037] FIG1 shows a flow chart of a method for predicting drug cancer inhibition sensitivity based on a graph neural network according to an embodiment of the present invention. The method specifically includes the following steps:

[0038] Step S1, obtaining original data of candidate drugs and cancer cell lines.

[0039] Step S2: extract the drug information features of the candidate drugs and expand them; construct the drug molecular feature vector based on the drug information features, and construct the drug adjacency matrix, node feature matrix and edge feature matrix with the drug atoms as nodes and chemical bonds as edges.

[0040] Specifically, step S2 includes the following sub-steps:

[0041] In step S201, the SMILES sequence of the molecule is used as the input standard to obtain drug information features from various drug information data sources (such as DeepChem, RDKit, and PubChem). By reading the official documents of each drug information data source, the specific meaning of each dimension in the drug information features is clarified.

[0042] Among them, the extractable drug information features include the atomic characteristics of the drug (atom type, formal charge, hybrid orbital, etc.), the chemical bond characteristics of the drug (bond type, homocyclicity, conjugation, etc.), and the molecular characteristics of the drug (molecular weight, number of valence electrons, etc.).

[0043] Taking the compound cyclobutane as an example, its SMILES sequence is "C1CCC1". Through "MolGraphConvFeaturizer", the atomic feature of the first "C" atom is obtained as "[1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]". The feature dimension is 30, of which dimensions 1-10 represent the atom type, dimension 11 represents the formal charge number, dimensions 12-14 represent hybrid orbitals, dimensions 15-16 represent hydrogen bond donors or acceptors, dimension 17 indicates whether the atom belongs to an aromatic ring, dimensions 18-24 represent the degree of the atom, and dimensions 25-30 represent the number of hydrogen atoms connected to the atom. Using the MolGraphConvFeaturizer, we can obtain the edge feature "[1,0,0,0,1,0,1,0,0,0,0]" for the edge between the first and second C atoms. The feature dimension is 11, with dimensions 1-4 representing the bond type of the edge, dimension 5 indicating whether the two atoms connected by the edge are in the same ring, dimension 6 indicating whether the edge is conjugated, and dimensions 7-11 representing the stereo configuration of the bond. Using the rdkit function, we can obtain molecular features for "C1CCC1," such as its molecular weight of 56.0626, its average molecular weight ignoring hydrogen of 48.0440, and its oil-water partition coefficient (LogP) of 1.5604.

[0044] Step S202 , compile a standard document of extractable drug information features, and perform deduplication, completion, and fusion on the drug information features obtained in step S201 .

[0045] Step S203 , based on the drug information obtained in step S202 , with drug atoms as nodes and chemical bonds as edges, construct a drug adjacency matrix A, a node feature matrix V, an edge feature matrix E, and a drug molecule feature vector Molecule_Feature.

[0046] Therefore, the drug adjacency matrix A is given by a matrix form of dimension (shape, shape), where "shape" represents the number of nodes of the drug (ie, the number of atoms of the drug).

[0047] The node feature matrix V is given in the form of a matrix with dimension (shape, dim1), where "dim1" represents the number of features of each node of the drug (ie, the number of features of each atom of the drug).

[0048] The edge feature matrix E is given by an adjacency matrix of dimensions (shape, shape, dim2), where dim2 represents the number of features per edge for the drug (i.e., the number of features of the drug's chemical bonds). Furthermore, when no edge exists between two nodes, the corresponding coordinate elements in the edge feature matrix are filled with zeros.

[0049] The drug molecular feature vector Molecule_Feature is given in the form of a feature vector with a dimension of (1, dim3), where "feature 3" represents the molecular feature dimension of the drug.

[0050] For the drug adjacency matrix A, since it is fixed, it can be directly obtained from the benchmark data source.

[0051] As the construction methods of the node feature matrix V and the edge feature matrix E are the same, the construction of the node feature matrix is ​​taken as an example to illustrate.

[0052] For the convenience of expression, this example only takes two data sources and converts the node feature matrix V that can be obtained by data source A and data source B into A , V B Respectively expressed in the following forms:

[0053] If the feature F B1 It does not appear in data source A. In this example, feature F B1 Add directly to V A At the end, it can be expressed as follows:

[0054] If the feature F B1 F in data source A A1 is a representation of the same property, when the characteristic F A1 With F B1 When the data is discretized, in this example, F A1 With F B1The feature that represents more in the above is used as the representation of the property. For example, the "Hybridization" attribute of the atom is only represented by MolGraphConvFeaturizer extractor as "sp", "sp2", "sp3", while the DMPNNFeaturizer extractor represents it as "sp", "sp2", "sp3", "sp3d", "sp3d2". Therefore, for the "Hybridization" attribute, this example uses the attribute extracted by the DMPNNFeaturizer extractor as the final attribute; when the feature F A1 With F B1 When the data is non-discretized, this example will A1 With F B1 Direct connection as a representation of this property can be expressed in the following form:

[0055] For the drug molecular feature vector Molecule_Feature, different molecular features can be obtained from different data sources, and these molecular features can be encoded into the drug molecular feature vector Molecule_Feature through encoding methods such as "one-hot".

[0056] Step S204: The drug adjacency matrix A, the node feature matrix V, the edge feature matrix E, and the drug molecule feature vector Molecule_Feature are transferred to the drug information database for storage.

[0057] As shown in Figure 2, step S3 builds a graph neural network model based on the drug adjacency matrix, node feature matrix and edge feature matrix, and iterates the graph neural network model to obtain drug graph features; after the drug graph features are aggregated with the drug molecular feature vectors, the drug representation is obtained.

[0058] It should be noted that in this example, the drug information features and the corresponding drug adjacency matrix A, node feature matrix V, and edge feature matrix E are obtained from the drug information database. A graph neural network model is built and iterated, performing message passing and aggregation via node and edge updates to obtain the drug graph feature (Graph_Feature). After aggregation with the drug molecule feature vector (Molecule_Feature) in the drug information database, the final drug representation (Whole_Feature) is obtained.

[0059] Specifically, step S3 includes the following sub-steps:

[0060] Step S301 is to build and iterate the graph neural network model, perform message transmission and aggregation in the form of node update and edge update, and obtain the drug graph feature (Graph_Feature).

[0061] Among them, the graph neural network model is updated in the form of node update, and the node feature update is performed in the lth layer of the graph neural network model. The expression is as follows:

[0062] In the formula, the edge feature matrix is ​​a matrix of dimension (shape, shape, dim2), where "dim2" represents the number of features of each edge of the drug, that is, the number of features of the drug chemical bond; during the node update process, the edge feature matrix is ​​split into "dim2" (shape, shape, 1) edge feature matrices; V (l+1) represents the node feature matrix of the l+1 layer, Updata represents the update function, V (l) represents the node feature submatrix of the lth layer, σ is the activation function, W represents the ith edge feature submatrix after the edge feature matrix on the lth layer is split. i (l+1) and Represents the graph neural network model parameters.

[0063] Update the graph neural network model in an edge-update manner, and perform edge feature update at the lth layer of the graph neural network model. The expression is as follows:

[0064] Where, Represents the edge e between node i and node j in the l+1 layer ij The update variable, σ is the activation function, V i (l+1) represents the feature vector of node i in the l+1 layer, V j (l+1) represents the feature vector of node j in the l+1 layer, and represents the graph neural network model parameters, represents the edge feature matrix between node i and node j in the l+1 layer, Updata represents the update function, Represents the edge feature matrix between node i and node j in layer l.

[0065] Step S303: Calculate the drug graph feature (Graph_Feature), which is expressed as follows:

[0066] Among them, |V| represents the number of nodes, V i (L) Represents the feature vector of node i in the Lth layer, where L is the number of layers of the graph neural network model.

[0067] Step S304: Aggregate the drug graph feature with the drug molecule feature vector to obtain the drug representation. The expression is as follows: Whole_Feature = Concat(Graph_Feature,Molecule_Feature)

[0068] Where Whole_Feature represents drug characterization, Graph_Feature represents drug graph features, and Molecule_Feature represents drug molecular features.

[0069] As shown in FIG3 , step S4 is to obtain the expression levels of gene expression, gene mutation, gene methylation, and gene copy number of the gene sequence in the cancer cell line, and extract the cancer cell line characteristics.

[0070] Specifically, step S4 includes the following sub-steps:

[0071] Step S401, obtaining information related to the expression levels of gene sequences in cancer cell lines. For cancer cell line C1, the expression level of sequence X c1 It can be expressed as follows:

[0072] Where n represents the number of pathogenic genes in cancer cell lines, X c1,gi Indicates the gene expression of the i-th pathogenic gene in the cancer cell line (express c1,gi ), gene mutation (mutat c1,gi ), gene methylation (methy c1,gi ), gene duplication level copy c1,gi expression level.

[0073] Step S402: the expression sequence X of the cancer cell line C1 is c1 Input into the Transformer network and output representation vector Z c1 , the expression is as follows:

[0074] In particular, this example uses a multi-head attention mechanism in the Transformer network. Indicates the expression sequence X of cancer cell line C1 c1 The representation obtained after the i-th attention head, Softmax is the activation function, are the Q, K, and V matrices of the Transformer network, Z c1 For X c1 Output representation after the Transformer network.

[0075] Step S403: characterize the vector Z c1 Input into the feedforward neural network to obtain the cancer cell line representation Cell_Feature, which is expressed as follows: Cell_Feature=σ(Z c1 W+b)

[0076] Where σ is the activation function, W and b represent the parameters of the feedforward neural network model

[0077] Step S5: perform feature fusion on the drug representation and the cancer cell line representation to obtain a drug-cancer cell line instance pair, which is expressed as follows: Mol_Cell_Feature = Whole_Feature || Cell_Feature

[0078] Where Mol_Cell_Feature represents a drug-cancer cell line instance pair, Whole_Feature represents a drug characterization, and Cell_Feature represents a cancer cell line characterization.

[0079] Step S6: input the drug-cancer cell line example pair into the drug tumor suppressor sensitivity prediction model to predict the IC50 value of the candidate drug in the cancer cell line environment.

[0080] Furthermore, the drug tumor suppression sensitivity prediction model is a classifier. The drug-cancer cell line example pair is input into the classifier to predict the drug tumor suppression sensitivity. The IC50 value of the drug in the cancer cell line environment is predicted, thereby quantifying the inhibitory ability of the drug on the cancer cell line. The expression is as follows: Pred = Softmax (Mol_Cell_Feature * W + b)

[0081] Where Mol_Cell_Feature represents the drug-cancer cell line instance pair, and W and b are the parameters of the drug tumor inhibitory sensitivity prediction model.

[0082] In summary, the present invention provides a method for predicting drug tumor suppressor sensitivity based on graph neural network, which extracts drug information features of candidate drugs and expands them, integrating major drug information data sources, breaking the bottlenecks of few drug information features of a single data source, missing drug information features of some drugs, etc., fully capturing drug feature information, and increasing drug information feature dimensions; in view of the difficulty in characterizing cancer cell lines, the expression of gene expression, gene mutation, gene methylation, and gene copy number of gene sequences in cancer cell lines is used to accurately represent cancer cell lines, and is encoded by Transformer network and feedforward neural network to extract cancer cell line characterization. The present invention performs feature fusion on drug characterization and cancer cell line characterization to obtain drug-cancer cell line instance pairs, and significantly improves the prediction accuracy of drug tumor suppressor sensitivity through more accurate and richer drug-cancer cell line instance pairs.

[0083] The embodiment of the present invention further provides a device for predicting drug cancer inhibition sensitivity based on a graph neural network, as shown in FIG4 , the device includes: a multi-source data integration module, a storage module, and a processing module;

[0084] The multi-source data integration module is used to obtain raw data of candidate drugs and cancer cell lines; extract and expand the drug information features of candidate drugs; construct drug molecular feature vectors based on drug information features, and construct drug adjacency matrices, node feature matrices, and edge feature matrices using drug atoms as nodes and chemical bonds as edges;

[0085] The storage module includes: a drug information database and a cancer cell line information database;

[0086] Drug information database, used to store drug molecule feature vectors, drug adjacency matrix, node feature matrix and edge feature matrix;

[0087] Cancer cell line information database, used to store gene expression, gene mutation, gene methylation, and gene copy number expression levels of gene sequences in cancer cell lines;

[0088] The processing modules include: drug characterization extraction unit, cancer cell line characterization extraction unit, drug-cancer cell line instance pair acquisition unit, and drug tumor suppression sensitivity prediction unit;

[0089] Drug characterization extraction unit: A graph neural network model is constructed based on the drug adjacency matrix, node feature matrix, and edge feature matrix. The graph neural network model is iterated to obtain drug graph features. The drug graph features are aggregated with the drug molecular feature vector to obtain the drug characterization.

[0090] Cancer cell line characterization extraction unit, which extracts cancer cell line characterization by obtaining the expression levels of gene expression, gene mutation, gene methylation, and gene copy number of gene sequences in cancer cell lines;

[0091] a drug-cancer cell line instance pair acquisition unit, configured to perform feature fusion on drug representations and cancer cell line representations to obtain drug-cancer cell line instance pairs;

[0092] The drug tumor suppressor sensitivity prediction unit predicts the IC50 value of the candidate drug in the cancer cell line environment by inputting the drug-cancer cell line instance pair into the drug tumor suppressor sensitivity prediction model.

[0093] It should be noted that the drug cancer inhibition sensitivity prediction device based on graph neural network is also provided with an input control interface at the input end of the processing module for controlling the input data; the output end of the processing module is also provided with an output control interface for controlling the output data; the output control interface is connected to the external interface, and its output expression form, in addition to the IC50 value, can also be used for drug feature visualization output, cancer cell line feature visualization output, etc.

[0094] Regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0095] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can be referred to the partial description of the method embodiment. The system embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present application scheme. Ordinary technicians in this field can understand and implement it without paying any creative work.

[0096] Accordingly, the present application also provides an electronic device comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned graph neural network-based drug suppression sensitivity prediction method. As shown in Figure 5, a hardware structure diagram of any device with data processing capabilities in which the graph neural network-based drug suppression sensitivity prediction method provided in an embodiment of the present invention is located. In addition to the processor, memory, and network interface shown in Figure 5, any device with data processing capabilities in which the device in the embodiment is located may also include other hardware based on the actual function of the device with data processing capabilities, which will not be described in detail.

[0097] Accordingly, the present application also provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the above-mentioned method for predicting drug cancer suppression sensitivity based on a graph neural network. The computer-readable storage medium can be an internal storage unit of any device with data processing capabilities as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), an SD card, a flash card (Flash Card), etc. equipped on the device. Furthermore, the computer-readable storage medium can also include both an internal storage unit and an external storage device of any device with data processing capabilities. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and can also be used to temporarily store data that has been output or is to be output.

[0098] The above embodiments are intended only to illustrate the design concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. The scope of protection of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design concepts disclosed in the present invention are within the scope of protection of the present invention.

Claims

1. A method for predicting drug anticancer sensitivity based on graph neural network, characterized in that: The method comprises: Obtain raw data on drug candidates and cancer cell lines; Extract drug information features of candidate drugs and expand them; construct drug molecular feature vectors based on drug information features, and construct drug adjacency matrix, node feature matrix and edge feature matrix with drug atoms as nodes and chemical bonds as edges; A graph neural network model is constructed based on the drug adjacency matrix, node feature matrix and edge feature matrix, and the drug graph features are obtained by iterating the graph neural network model. The drug graph features are aggregated with the drug molecule feature vector to obtain the drug representation. Obtain the expression of gene sequence, gene expression, gene mutation, gene methylation, and gene copy number in cancer cell lines, and extract cancer cell line characterization; Perform feature fusion on drug representation and cancer cell line representation to obtain drug-cancer cell line instance pairs; The drug-cancer cell line example pairs are input into the drug tumor suppressor sensitivity prediction model to predict the IC50 value of the candidate drug in the cancer cell line environment.

2. The method for predicting drug tumor suppressor sensitivity based on graph neural network according to claim 1, characterized in that: Drug information features include drug atomic features, drug chemical bond features, and drug molecular features; The drug adjacency matrix is ​​a matrix of dimension (shape, shape), where "shape" represents the number of nodes of the drug, i.e., the number of atoms of the drug; The node feature matrix is ​​a matrix of dimension (shape, dim1), where "dim1" represents the number of features of each node of the drug, that is, the number of features of each atom of the drug; The edge feature matrix is ​​a matrix of dimension (shape, shape, dim2), where "dim2" represents the number of features of each edge of the drug, that is, the number of features of the drug's chemical bonds; The drug molecule feature vector is a feature vector of dimension (1, dim3), where "dim3" represents the molecular feature dimension of the drug.

3. The method for predicting drug tumor suppressor sensitivity based on graph neural network according to claim 1 or 2, characterized in that: The node feature matrix is ​​constructed in the same way as the edge feature matrix; The construction process of the node feature matrix includes: Get data source A and data source B, and the first node feature matrix V corresponding to data source A and data source B A , the second node feature matrix V B The expression is as follows: When the feature F B1 does not appear in data source A, feature F B1 Add directly to the first node feature matrix V A At the end, the node feature matrix V is obtained, and the expression is as follows: When the feature F B1 F in data source A A1 is a representation of the same property, and the characteristic F A1 With F B1 When the data is discretized, take F A1 With F B1 In the expression, more features are used as the representation of this property; When the feature F A1 With F B1 When the data is non-discretized, F A1 With F B1 Direct connection is used as a representation of this property. The expression of the node feature matrix V is:

4. The method for predicting drug tumor suppressor sensitivity based on graph neural network according to claim 1, characterized in that: The drug graph features obtained by iterative graph neural network model include: The graph neural network model is updated in the form of node updates, where the node features are updated in the lth layer of the graph neural network model. The expression is as follows: In the formula, the edge feature matrix is ​​a matrix of dimension (shape, shape, dim2), where "dim2" represents the number of features of each edge of the drug, that is, the number of features of the drug chemical bonds; during the node update process, the edge feature matrix is ​​split into "dim2" (shape, shape, 1) edge feature matrices; V (l+1) represents the node feature matrix of the l+1th layer, Updata represents the update function, V (l) represents the node feature submatrix of the lth layer, σ is the activation function, represents the ith edge feature submatrix after the edge feature matrix on the lth layer is split, W i (l+1) and Represents the graph neural network model parameters.

5. The method for predicting drug tumor suppressor sensitivity based on graph neural network according to claim 1, characterized in that: The drug graph features obtained by iterative graph neural network model include: The graph neural network model is updated in an edge-update manner, where edge features are updated at the lth layer of the graph neural network model. The expression is as follows: In the formula, Represents the edge e between node i and node j in the l+1th layer ij The update variable, σ is the activation function, V i (l+1) represents the feature vector of node i at layer l+1, V j (l+1) represents the feature vector of node j at layer l+1, and represents the graph neural network model parameters, represents the edge feature matrix between node i and node j in the l+1th layer, Updata represents the update function, Represents the edge feature matrix between node i and node j in layer l.

6. The method for predicting drug tumor suppressor sensitivity based on graph neural network according to claim 1, characterized in that: Obtain the expression of gene expression, gene mutation, gene methylation, and gene copy number of stored gene sequences in cancer cell lines, and extract cancer cell line characteristics including: The expressed sequence X of cancer cell line C1 c1 The expression is as follows: In the formula, n represents the number of pathogenic genes in cancer cell lines, X c1,gi Indicates the expression level of the i-th pathogenic gene in the cancer cell line, express c1,gi represents the expression level of the ith pathogenic gene in the cancer cell line, mutat c1,gi represents the expression level of gene mutation of the i-th pathogenic gene in the cancer cell line, methy c1,gi represents the gene methylation expression of the i-th pathogenic gene in the cancer cell line, copy c1,gi represents the expression level of the gene copy number of the i-th pathogenic gene in the cancer cell line; The expression sequence X of cancer cell line C1 c1 Input into the Transformer network and output the representation vector Z c1 ; The representation vector Z c1 Input into a feed-forward neural network to obtain cancer cell line representation.

7. The method for predicting drug tumor suppressor sensitivity based on graph neural network according to claim 1, characterized in that: The drug-cancer cell line example pairs are input into the drug tumor suppressor sensitivity prediction model, and the IC50 values ​​of the candidate drugs in the cancer cell line environment are predicted to include: The drug tumor suppressor sensitivity prediction model is a classifier. The drug-cancer cell line instance pair is input into the classifier to predict the drug tumor suppressor sensitivity, and the IC50 value of the drug in the cancer cell line environment is predicted. The expression is as follows: Pred=Softmax(Mol_Cell_Feature*W+b) In the formula, Mol_Cell_Feature represents the drug-cancer cell line instance pair, W and b are the parameters of the drug tumor inhibitor sensitivity prediction model. number.

8. A drug anticancer sensitivity prediction device based on graph neural network, characterized in that: The device comprises: a multi-source data integration module, a storage module and a processing module; The multi-source data integration module is used to obtain the original data of candidate drugs and cancer cell lines; extract the drug information features of candidate drugs and expand them; construct the drug molecule feature vector based on the drug information features, and construct the drug adjacency matrix, node feature matrix and edge feature matrix with the drug atoms as nodes and chemical bonds as edges; The storage module includes: a drug information database and a cancer cell line information database; Drug information database, used to store drug molecule feature vectors, drug adjacency matrix, node feature matrix and edge feature matrix; Cancer cell line information database, used to store gene expression, gene mutation, gene methylation, and gene copy number expression of gene sequences in cancer cell lines; The processing modules include: a drug characterization extraction unit, a cancer cell line characterization extraction unit, a drug-cancer cell line instance pair acquisition unit, and a drug tumor suppressor sensitivity prediction unit; Drug characterization extraction unit: construct a graph neural network model based on the drug adjacency matrix, node feature matrix and edge feature matrix, and iterate the graph neural network model to obtain drug graph features; after the drug graph features are aggregated with the drug molecule feature vector, the drug characterization is obtained; A cancer cell line characterization extraction unit extracts cancer cell line characterization by obtaining the expression of gene expression, gene mutation, gene methylation, and gene copy number of gene sequences in cancer cell lines; A drug-cancer cell line instance pair acquisition unit, used for performing feature fusion on drug representation and cancer cell line representation to obtain a drug-cancer cell line instance pair; The drug tumor suppressor sensitivity prediction unit predicts the IC50 value of the candidate drug in the cancer cell line environment by inputting the drug-cancer cell line instance pair into the drug tumor suppressor sensitivity prediction model.

9. An electronic device, characterized in that: It comprises one or more processors for implementing the drug cancer inhibition sensitivity prediction method based on graph neural network according to any one of claims 1 to 7.

10. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, it is used to implement the drug cancer inhibition sensitivity prediction method based on graph neural network described in any one of claims 1 to 7.

Images