Anticancer drug reaction prediction method based on isomorphic multiple branches

Abstract

The invention relates to an anti-cancer drug response prediction method based on isomorphic multiple branches, and is applied to the field of bioinformatics. The method comprises the following steps: S1, preprocessing data to obtain cell line multi-omics data, drug substructure fingerprint data, drug molecular map data and cell line-drug reaction data; s2, constructing a similarity global feature extraction module (SGFE) for feature extraction to obtain cell line similarity features, drug similarity features and drug global features; s3, constructing an isomorphic multi-branch feature extraction module (HHFE) to obtain multi-branch cell line features and multi-branch drug features; s4, constructing a multi-branch multi-layer fusion module (MFML) to obtain a cell line fusion feature and a drug fusion feature, and calculating the correlation between the cell line fusion feature and the drug fusion feature; and S5, predicting the reaction condition between the cell line and the drug by using the correlation, and outputting a prediction result. According to the method, known cell line data and drug data are fully utilized, and the generalization ability of the model is improved.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics, and specifically is a method for predicting anticancer drug reactions based on heterogeneous and homomorphic multi-branching. Background Art

[0002] Cancer is a disease caused by gene mutations. The heterogeneity of cancer means that patients with the same type of cancer may have different responses to the same drug. In the context of precision medicine, anti-cancer drug response prediction (CDR) can help screen drugs and provide patients with more accurate and targeted drug treatment plans.

[0003] With the development of gene sequencing technology, more and more genome databases have been established, and more gene-related omics data and drug reaction data have been gradually expanded. The rich and valuable data provide data support for the research on the prediction of anticancer drug response, making it possible to provide personalized treatment for cancer patients.

[0004] Previous studies have greatly promoted the development of CDR, but there are still problems such as insufficient utilization of cell line and drug data and poor model generalization ability. To solve this problem, GraphDRP uses graph convolutional network (GCN) to extract the graph representation features of drugs, and uses somatic mutation and copy number variation data to represent cell lines, and increases the abundance of cell line features by increasing the type of cell line omics data; GraphTransDRP uses Transformer technology to mine the structural features of drugs based on the molecular graph of drugs, extracts more comprehensive drug features by using more complex network structures, and uses gene expression data to represent cell lines. However, these models do not fully consider the characteristics of various forms of data when selecting data, and there will be feature redundancy or feature missing. In addition, when extracting features, they cannot select appropriate network structures to extract appropriate features for different data, resulting in incomplete feature extraction and affecting the generalization ability and prediction performance of the model. Therefore, analyzing the characteristics of various data forms, combining different forms of data, and developing a network model that can fully extract the features of various forms of data for CDR prediction is of practical significance for shortening drug discovery time and providing guidance for drug screening. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes an anticancer drug response prediction method based on heterogeneous and homomorphic multi-branching to provide guidance for the drug discovery process.

[0006] Specifically, the method comprises the following steps:

[0007] S1: Preprocess the data to obtain cell line multi-omics data, drug substructure fingerprint data, drug molecular graph data, and cell line-drug reaction data;

[0008] S2: Construct a similarity global feature extraction module (SGFE) to extract features and obtain cell line similarity features, drug similarity features, and drug global features;

[0009] S3: Construct a heterogeneous multi-branch feature extraction module (HHFE) to extract the cell line similarity features and drug similarity features to obtain multi-branch cell line features and multi-branch drug features;

[0010] S4: Construct a multi-branch multi-layer fusion module (MFML), fuse the multi-branch cell line features to obtain the cell line fusion features, fuse the drug global features and multi-branch drug features to obtain the drug fusion features, and calculate the correlation between the two;

[0011] S5: Use the correlation to predict the reaction between cell lines and drugs and output the prediction results.

[0012] Preferably, S1 comprises the following steps:

[0013] S1.1: Cell line-drug response tables collected in GDSC and CCLE databases;

[0014] S1.2: Based on the response table, gene expression, copy number variation, and somatic mutation data of cell lines were collected in the GDSC and CCLE databases, respectively;

[0015] S1.3: Cell lines with all three modalities were screened, and 962 and 436 cell line data were obtained in GDSC and CCLE, respectively;

[0016] S1.4: Obtain the PubChem ID of each drug based on the cell line-drug reaction table;

[0017] S1.5: Based on the PubChem ID, the SMILES sequence of the drug was obtained from the PubChem library;

[0018] S1.6: Use the RDKit toolkit to obtain the substructure fingerprint and molecular graph data corresponding to the drug based on the SMILES sequence of the drug.

[0019] Preferably, S2 comprises the following steps:

[0020] S2.1: Perform similarity fusion on multi-omics data of cell lines to obtain similarity features of cell lines;

[0021] S2.2: Calculate drug similarity characteristics based on drug substructure fingerprints;

[0022] S2.3: Calculate the global characteristics of the drug based on the molecular graph data of the drug;

[0023] Preferably, S2.1 comprises the following steps:

[0024] The gene expression data is Gaussian regularized, which can be described as:

[0025]

[0026] exp i represents the expression value of the ith gene in the cell line, represents the regularized expression value of the ith gene, μ i and δ i represent the mean and standard deviation of the expression of the ith gene in all cell lines, respectively;

[0027] Calculate the similarity of each omics data of the cell line separately. This operation can be described as:

[0028]

[0029] x i and x j denote the i-th and j-th specific omics data, 1≤i,j≤962, ||||2 denotes the second normal form, ε is a hyperparameter, E ij Indicates the similarity of specific omics data in cell lines;

[0030] The full kernel matrix F and sparse kernel matrix S are calculated for each omics data. This operation can be described as:

[0031]

[0032] N i represents the top N cell lines most similar to the i-th cell line;

[0033] According to the calculated full nuclear matrix GF and sparse nuclear matrix GS of gene expression, full nuclear matrix CF and sparse nuclear matrix CS of copy number variation, full nuclear matrix MF and sparse nuclear matrix MS of somatic mutation, let GF0=GF, CF0=CF, MF0=MF, and iteratively update the full nuclear matrix. This operation can be described as:

[0034]

[0035] t represents the number of iterations, * T Indicates transposing the matrix*

[0036] Finally, the cell line similarity feature RC is obtained, and the calculation formula is as follows:

[0037]

[0038] Preferably, S2.2 comprises the following steps:

[0039] According to the substructure fingerprint of the drug, the Tanimoto similarity is calculated to obtain the drug similarity feature RD. This operation can be described as:

[0040]

[0041] d i and d j Represent the substructure fingerprints of the i-th and j-th drugs, respectively, 1≤i,j≤228;

[0042] Preferably, S2.3 comprises the following steps:

[0043] Extract features from the drug’s molecular graph to obtain the drug’s global features f molD , the operation can be described as:

[0044]

[0045] is the feature of node i at the k+1th layer, represents the set of neighbor nodes of node i, W (k) is the weight matrix of the kth layer, σ is the activation function, is the attention coefficient between node i and its neighbor node j, and is calculated as follows:

[0046]

[0047] α (k) is the learned attention weight vector, || represents the concatenation operation, and LeakyRelu is the activation function;

[0048] Preferably, S3 comprises the following steps:

[0049] S3.1: Extract features of cell line similarity features and drug similarity features, and obtain branch features of cell line subgraphs and branch features of drug subgraphs by fusing features of subgraphs of different scales;

[0050] S3.2: extracting cell line similarity features and drug similarity features through message passing to obtain cell line heterogeneous branch features and drug heterogeneous branch features;

[0051] S3.3: Extract features in the cell line isomorphism graph to obtain cell line isomorphism branch features;

[0052] S3.4: Extract features in the drug isomorphism graph to obtain drug isomorphism branch features;

[0053] Preferably, S3.1 comprises the following steps:

[0054] Subgraphs of different sizes are constructed by downsampling, and features are extracted from the subgraphs. The features of subgraphs of different sizes are fused to obtain the branch features of the cell line subgraphs f UC and drug subgraph branch feature f UD , the operation can be described as:

[0055] y=X l p l / ||p l ||

[0056] idx=rank(y,k)

[0057]

[0058] A l+1 =A l (idx,idx)

[0059]

[0060] X l represents the node feature of the lth layer, p l is a trainable vector, ||p l || represents vector p l The length of y is X l In the vector p l The projection scalar value on , rank(y,k) means selecting the largest k values ​​from y, and idx records the subscript of the largest k values. is the feature corresponding to k nodes, A l+1 is the adjacency matrix consisting of the corresponding k nodes, W l is the weight matrix, H (l+1) is the updated node feature, and σ is the activation function.

[0061] Preferably, S3.2 comprises the following steps:

[0062] In the cell line-drug heterogeneous graph, feature extraction is performed using message passing, and the cell line heterogeneous branch features f are obtained respectively. HeC and drug isomeric branching characteristics f HeD , the operation can be described as:

[0063]

[0064] represents the normalized adjacency matrix.

[0065] Preferably, S3.3 comprises the following steps:

[0066] The cell line isomorphism branch features f are extracted using message passing in the cell line isomorphism graph HoC , the operation can be described as:

[0067] A c =A*A T

[0068]

[0069] A is the cell line-drug response matrix, D c Yes A c The degree matrix of is the normalized adjacency matrix in the cell line isomorphism graph, σ is the activation function;

[0070] Preferably, S3.4 comprises the following steps:

[0071] The drug isomorphism branch features f are extracted using message passing in the drug isomorphism graph HoD , the operation can be described as:

[0072] A d =A T *A

[0073]

[0074] A is the cell line-drug response matrix, D d Yes A d The degree matrix of is the normalized adjacency matrix in the drug isomorphism graph, σ is the activation function;

[0075] Preferably, S4 comprises the following steps:

[0076] S4.1: Fusion of each branch feature of the cell line yields the cell line fusion feature; fusion of each branch feature of the drug and the global feature of the drug yields the drug fusion feature;

[0077] S4.2: Linearly transform the cell line fusion signature and drug fusion signature to adjust them to the same dimension;

[0078] S4.3: Calculate the correlation between the adjusted cell line fusion signature and the drug fusion signature;

[0079] Preferably, S4.1 comprises the following steps:

[0080] The cell lineage subgraph branch features f UC , cell line heterogeneous branching characteristicsfHeC , cell line isomorphic branching characteristicsf HoC Fusion to obtain cell line fusion characteristics H 1-c , the drug subgraph branch feature f UD , drug isomerization branch characteristics f HeD , drug isomorphic branch characteristics f HoD , drug global characteristics f molD Fusion obtains drug fusion feature H 1-d , the operation can be described as:

[0081] H 1-c =(f UC +f HeC +f HoC ) / 3

[0082] H 1-d =((f UD +f HeD ) / 2+f HoD +f molD ) / 3

[0083] Preferably, S4.2 comprises the following steps:

[0084] The cell line fusion features and drug fusion features are adjusted to the same dimension through linear transformation, and then the correlation between the two after the transformation is calculated. This operation can be described as:

[0085] H (2-c) =H (1-c) W (2-c)

[0086] H (2-d) =H (1-d) W (2-d)

[0087] Here, W (2-c) ∈R c*k and W (2-d) ∈R d*k are the weight matrices for linear transformation of cell lines and drugs, respectively. c is the dimension of cell line fusion features, d is the dimension of drug fusion features, k represents the feature dimension after linear transformation, and H (2-c) and H (2-d) They are the k-dimensional cell line fusion features and k-dimensional drug fusion features after linear transformation;

[0088] Preferably, S4.3 comprises the following steps:

[0089] Calculate the correlation between the k-dimensional cell line fusion features and the k-dimensional drug fusion features. This operation can be described as:

[0090]

[0091] h i is the i-th cell line feature in the k-dimensional cell line fusion feature, h j is the jth drug feature in the k-dimensional drug fusion feature, μ i and μ j They are h i and h j The mean of , 1≤i≤962, 1≤j≤228;

[0092] Preferably, S5 comprises the following steps:

[0093] S5.1: Perform sigmoid activation output on the correlation result and output the prediction result;

[0094] Preferably, S5.1 comprises the following steps:

[0095] The correlation coefficient calculated by S4 is in the range of [-1, 1]. The sigmoid function is used for activation output to ensure that the predicted probability value range of each pair of cell line-drug sensitivity is [0, 1], and the cell line-drug response matrix is ​​obtained. This operation can be described as:

[0096]

[0097] h is the correlation coefficient, γ is the predicted cell line-drug response matrix, and α is a hyperparameter used to accelerate parameter updates.

[0098] The beneficial effects of the method of the present invention are as follows: the present invention uses three kinds of omics data of gene expression, copy number variation and somatic mutation of cell lines to obtain similarity characteristics between cell lines; uses the substructure fingerprint of drugs to calculate the Tanimoto similarity characteristics between drugs, and also calculates the global characteristics of drugs according to the molecular graph of drugs. Based on the similarity of similar cell lines to the same drugs, we use the similarity between cell lines and drugs to predict the reaction of cell lines and drugs. At the same time, by analyzing the characteristics of different drug data, after calculating the similarity according to the substructure fingerprint, the molecular graph data is used to enrich the characteristic information of drug data; on the other hand, a heterogeneous multi-branch network is used to extract the intrinsic connection between cell lines and drugs in the cell line isomorphic graph and the drug isomorphic graph respectively using similarity information, and in the cell line-drug heterogeneous graph, a dual-branch network structure is used to extract comprehensive information flow, capture the relationship characteristics of different types of nodes, and finally perform feature fusion to enhance the generalization of the model and improve the prediction ability of the model. Finally, experiments were conducted in two data sets, GDSC and CCLE, to verify the predictive ability of the present invention for drug response. The present invention can fully mine the characteristic information of cell lines and drugs and improve the ability to predict cell line-drug reactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the prior art and the drawings required for use in the embodiments. The following drawings 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.

[0100] Figure 1 It is a schematic diagram of the process of the present invention;

[0101] Figure 2 It is a system structure diagram of the present invention;

[0102] Figure 3 Schematic diagram of similarity global feature extraction module of the present invention;

[0103] Figure 4 Schematic diagram of a heterogeneous multi-branch feature extraction module of the present invention;

[0104] Figure 5 It is a schematic diagram of a multi-branch and multi-layer fusion module of the present invention; Specific implementation plan

[0105] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all other embodiments obtained by ordinary technicians in this field without creative work based on the embodiments of the present invention are within the scope of protection of the present invention.

[0106] The embodiment of the present invention provides an anticancer drug reaction prediction method based on heterogeneous and homogeneous multi-branching, which is used to predict the reaction between drugs and cancer cell lines and provide guidance for drug screening.

[0107] Reference Figure 1 , the method comprises the following steps:

[0108] S1: Preprocess the data to obtain cell line multi-omics data, drug substructure fingerprint data, drug molecular graph data, and cell line-drug reaction data;

[0109] S2: Construct a similarity global feature extraction module (SGFE) to extract features and obtain cell line similarity features, drug similarity features, and drug global features;

[0110] S3: Construct a heterogeneous multi-branch feature extraction module (HHFE) to extract the cell line similarity features and drug similarity features to obtain multi-branch cell line features and multi-branch drug features;

[0111] S4: Construct a multi-branch multi-layer fusion module (MFML), fuse the multi-branch cell line features to obtain the cell line fusion features, fuse the drug global features and multi-branch drug features to obtain the drug fusion features, and calculate the correlation between the two;

[0112] S5: Use the correlation to predict the reaction between cell lines and drugs and output the prediction results.

[0113] Further, S1 comprises the following steps:

[0114] S1.1: Cell line-drug response tables collected in GDSC and CCLE databases;

[0115] S1.2: Based on the response table, gene expression, copy number variation, and somatic mutation data of cell lines were collected in the GDSC and CCLE databases, respectively;

[0116] S1.3: Screen out cell lines that have all three modalities;

[0117] S1.4: 962 cell lines and 228 drug response datasets were obtained from GDSC; 436 cell lines and 24 drug response datasets were obtained from CCLE;

[0118] S1.5: Obtain the PubChem ID of each drug based on the cell line-drug reaction table;

[0119] S1.6: Based on the PubChem ID, the SMILES sequence of the drug was obtained from the PubChem library;

[0120] S1.7: Use the RDKit toolkit to obtain the substructure fingerprint and molecular graph data corresponding to the drug based on the SMILES sequence of the drug.

[0121] Further, refer to Figure 3 , S2 includes the following steps:

[0122] S2.1: Perform similarity fusion on multi-omics data of cell lines to obtain similarity features of cell lines;

[0123] S2.2: Calculate drug similarity characteristics based on drug substructure fingerprints;

[0124] S2.3: Calculate the global characteristics of the drug based on the molecular graph data of the drug;

[0125] Further, S2.1 includes the following steps:

[0126] The gene expression data is Gaussian regularized, which can be described as:

[0127]

[0128] exp i represents the expression value of the ith gene in the cell line, represents the regularized expression value of the ith gene, μ i and δ i represent the mean and standard deviation of the expression of the ith gene in all cell lines, respectively;

[0129] Calculate the similarity of each omics data of the cell line separately. This operation can be described as:

[0130]

[0131] x i and x j Represent the i-th and j-th specific omics data, 1≤i,j≤962, ||||2 represents the second normal form, E ij represents the similarity of specific omics data in cell lines, and ε is set to 2;

[0132] The full kernel matrix F and sparse kernel matrix S are calculated for each omics data. This operation can be described as:

[0133]

[0134] N i represents the top N cell lines most similar to the i-th cell line;

[0135] According to the calculated full nuclear matrix GF and sparse nuclear matrix GS of gene expression, full nuclear matrix CF and sparse nuclear matrix CS of copy number variation, full nuclear matrix MF and sparse nuclear matrix MS of somatic mutation, let GF0=GF, CF0=CF, MF0=MF, and iterate the full nuclear matrix 3 times to update. This operation can be described as:

[0136]

[0137] t represents the number of iterations, here it is 3, * T Indicates transposing the matrix*

[0138] Finally, the cell line similarity feature RC is obtained, and the calculation formula is as follows:

[0139]

[0140] Among them, RC∈R 962*962, R represents a real number.

[0141] Further, S2.2 includes the following steps:

[0142] According to the substructure fingerprint of the drug, the Tanimoto similarity is calculated to obtain the drug similarity feature RD. This operation can be described as:

[0143]

[0144] d i and d j Represent the substructure fingerprints of the i-th and j-th drugs, 1≤i,j≤228, RD∈R 228*228 , R represents a real number.

[0145] Further, S2.3 includes the following steps:

[0146] Extract features from the drug’s molecular graph to obtain the drug’s global features f molD , the operation can be described as:

[0147]

[0148] is the feature of node i at the k+1th layer, where k=2, and 3 layers of feature extraction are performed. represents the set of neighbor nodes of node i, W (k) is the weight matrix of the kth layer, σ is the activation function, and the relu activation function is used here. is the attention coefficient between node i and its neighbor node j, and is calculated as follows:

[0149]

[0150] α (K) is the learned attention weight vector, || represents the concatenation operation;

[0151] Further, refer to Figure 4 , S3 includes the following steps:

[0152] S3.1: Extract features of cell line similarity features and drug similarity features, and obtain branch features of cell line subgraphs and branch features of drug subgraphs by fusing features of subgraphs of different scales;

[0153] S3.2: extracting cell line similarity features and drug similarity features through message passing to obtain cell line heterogeneous branch features and drug heterogeneous branch features;

[0154] S3.3: Extract features in the cell line isomorphism graph to obtain cell line isomorphism branch features;

[0155] S3.4: Extract features in the drug isomorphism graph to obtain drug isomorphism branch features;

[0156] Further, S3.1 includes the following steps:

[0157] According to the extracted cell line similarity features RC and drug similarity features and RD, we concatenate the node features to obtain the node feature matrix X of the heterogeneous graph: The cell line-drug response table is used as the adjacency matrix A, and the original adjacency matrix is ​​expanded to obtain the adjacency matrix on the heterogeneous graph where A∈R 962*228 , I c is the identity matrix of 962*962, I d It is a 228*228 unit matrix. Subgraphs of different sizes are constructed by downsampling, and features are extracted from the subgraphs. The features of subgraphs of different sizes are fused to obtain the branch features of the cell line subgraphs f UC and drug subgraph branch feature f UD , the operation can be described as:

[0158] y=X l p l / ||p l ||

[0159] idx=rank(y,k)

[0160]

[0161] X l represents the node features of the lth layer, p l is a trainable vector, ||p l || represents vector p l The length of y is X l In the vector p l The projection scalar value on , rank(y,k) means selecting the largest k values ​​from y, and idx records the subscript of the largest k values. is the node feature of the subgraph corresponding to k nodes, A l+1 is the adjacency matrix of the l+1th layer, W l is the weight matrix, H (l +1) is the updated node feature, l is 3, and a total of 3 layers of downsampling feature extraction are performed.

[0162] Further, S3.2 includes the following steps:

[0163] In the cell line-drug heterogeneous graph, feature extraction is performed using message passing, and RC and RD are concatenated to obtain the node feature matrix of the heterogeneous graph. The cell line heterogeneous branch characteristics f HeC and drug isomeric branching characteristics f HeD , the operation can be described as:

[0164]

[0165] in represents the normalized adjacency matrix, l is 2, and a total of 3 layers of feature extraction are performed;

[0166] Further, S3.3 includes the following steps:

[0167] The cell line isomorphism branch features f are extracted using message passing in the cell line isomorphism graph HoC , the operation can be described as:

[0168] A c =A*A T

[0169]

[0170] A is the cell line-drug response matrix, A c is the adjacency matrix of the cell lineage isomorphism graph, is the normalized adjacency matrix in the cell line isomorphism graph, D c Yes A c The degree matrix, σ is the activation function, here we use the relu activation function, l is 2, and a total of 3 layers of feature extraction are performed;

[0171] Further, S3.4 includes the following steps:

[0172] The drug isomorphism branch features f are extracted using message passing in the drug isomorphism graph HoD , the operation can be described as:

[0173] A d =A T *A

[0174]

[0175] A is the cell line-drug response matrix, A c is the adjacency matrix of the drug isomorphism graph, is the normalized adjacency matrix in the drug isomorphism graph, D d Yes A d The degree matrix, σ is the activation function, here we use the relu activation function, l is 2, and a total of 3 layers of feature extraction are performed.

[0176] Further, refer to Figure 5 , S4 comprises the following steps:

[0177] S4.1: Fusion of each branch feature of the cell line yields the cell line fusion feature; fusion of each branch feature of the drug and the global feature of the drug yields the drug fusion feature;

[0178] S4.2: Linearly transform the cell line fusion signature and drug fusion signature to adjust them to the same dimension;

[0179] S4.3: Calculate the correlation between the adjusted cell line fusion signature and the drug fusion signature;

[0180] Further, S4.1 includes the following steps:

[0181] The cell lineage subgraph branch features f UC , cell line heterogeneous branching characteristicsf HeC , cell line isomorphic branching characteristicsf Hoc Fusion to obtain cell line fusion characteristics H 1-C , the drug subgraph branch feature f UD , drug isomerization branch characteristics f HeD , drug isomorphic branch characteristics f HoD , drug global characteristics f molD Fusion obtains drug fusion feature H 1-d , the operation can be described as:

[0182] H 1-c =(f UC +f HeC +f HoC ) / 3

[0183] H 1-d =((f UD +f HeD ) / 2+f HoD +f molD ) / 3

[0184] Further, S4.2 includes the following steps:

[0185] The cell line fusion features and drug fusion features are adjusted to the same dimension through linear transformation, and then the correlation between the two after the transformation is calculated. This operation can be described as:

[0186] H (2-c) =H (1-c) W (2-c)

[0187] H (2-d) =H (1-d) W (2-d)

[0188] W (2-c) ∈R 962*128 and W (2-d)∈R 228*128 are the weight matrices for linear transformation of cell lines and drugs, 962 is the dimension of cell line fusion features, 228 is the dimension of drug fusion features, 128 represents the feature dimension after linear transformation, and H (2-c) and H (2-d) They are the 128-dimensional cell line fusion features and 128-dimensional drug fusion features after linear transformation.

[0189] Further, S4.3 includes the following steps:

[0190] Calculate the correlation between the 128-dimensional cell line fusion features and the 128-dimensional drug fusion features. This operation can be described as:

[0191]

[0192] h i is the i-th cell line feature in the 128-dimensional cell line fusion feature, h j is the jth drug feature in the 128-dimensional drug fusion feature, μ i and μ j They are h i and h j The mean of , 1≤i≤962, 1≤j≤228;

[0193] Further, S5 comprises the following steps:

[0194] S5.1: Perform sigmoid activation output on the correlation result and output the prediction result;

[0195] Further, S5.1 includes the following steps:

[0196] The correlation coefficient calculated by S4 is in the range of [-1, 1]. The sigmoid function is used for activation output to ensure that the predicted probability value range of each pair of cell line-drug sensitivity is [0, 1], and the cell line-drug response matrix is ​​obtained. This operation can be described as:

[0197]

[0198] h is the correlation coefficient, γ is the predicted cell line-drug response matrix, and α is a hyperparameter, which is 5 here to accelerate parameter updates.

[0199] In order to evaluate the performance of the model, some indicators are used to evaluate the prediction results of the model on the two data sets. The calculation formula of the specific evaluation indicators is as follows:

[0200] The full name of AUC is the area under the ROC curve. The horizontal axis is FPR (False Positive Rate) and the vertical axis is TRP (True Positive Rate). It is an indicator used to measure the performance of a binary classification model. The AUC range is [0,1]. The closer the value is to 1, the better the model effect.

[0201] FPR is calculated as follows, indicating the proportion of samples that are actually negative that are incorrectly predicted as positive:

[0202]

[0203] TPR is calculated as follows, indicating the proportion of samples that are actually positive that are correctly predicted as positive:

[0204]

[0205] The full name of AUPRC is the area under the PR curve. The horizontal axis is the recall rate (Recall), from 0 to 1, and the vertical axis represents the precision (precision), from 0 to 1. The AUPR range is [0,1]. The closer the value is to 1, the better the model effect.

[0206] The other indicators are calculated as follows:

[0207]

[0208] Among them: TP means that the actual class is positive and the model predicts it is also positive; FP means that the actual class is negative and the model predicts it is positive; TN means that the actual class is negative and the model predicts it is also negative; FN means that the actual class is positive and the model predicts it is negative.

[0209] In order to evaluate the superiority of this application, in addition to evaluating this model, comparative experiments were conducted with other advanced CDR methods. The results are shown in Table 1 and Table 2. In addition, Table 3 shows the 20 cell line-drug pairs with the highest probability in the prediction results, of which 19 pairs have been experimentally confirmed and 1 pair has not yet been experimentally confirmed. The results show the excellent prediction ability of this application:

[0210] Table 1. Comparison results of independent test performance on the GDSC dataset

[0211]

[0212] Table 2. Comparison results of independent test performance on CCLE dataset

[0213]

[0214] Table 3 Prediction of the top 20 cell line-drug pairs with the highest correlation probability

[0215]

[0216]

Claims

1. A method for predicting anticancer drug response based on heterogeneous and homogeneous multi-branching, characterized in that: The method comprises the following steps: S1: Preprocess the data to obtain cell line multi-omics data, drug substructure fingerprint data, drug molecular graph data, and cell line-drug reaction data; S2: Construct a similarity global feature extraction module (SGFE) to extract features and obtain cell line similarity features, drug similarity features, and drug global features; S3: Construct a heterogeneous multi-branch feature extraction module (HHFE) to extract the cell line similarity features and drug similarity features to obtain multi-branch cell line features and multi-branch drug features; S4: Construct a multi-branch multi-layer fusion module (MFML), fuse the multi-branch cell line features to obtain the cell line fusion features, fuse the drug global features and multi-branch drug features to obtain the drug fusion features, and calculate the correlation between the two; S5: Use the correlation to predict the reaction between cell lines and drugs and output the prediction results.

2. The method for predicting anticancer drug response based on heterogeneous and homogeneous multi-branching according to claim 1, characterized in that: S1 includes the following steps: S1.1: Cell line-drug response tables collected in GDSC and CCLE databases; S1.2: Based on the response table, gene expression, copy number variation, and somatic mutation data of cell lines were collected in the GDSC and CCLE databases, respectively; S1.3: Cell lines with all three modalities were screened, and 962 and 436 cell line data were obtained in GDSC and CCLE, respectively; S1.4: Obtain the PubChemID of each drug based on the cell line-drug reaction table; S1.5: Based on the PubChem ID, the SMILES sequence of the drug was obtained from the PubChem library; S1.6: Use the RDKit toolkit to obtain the substructure fingerprint and molecular graph data corresponding to the drug based on the SMILES sequence of the drug.

3. The method for predicting anticancer drug response based on heterogeneous and homogeneous multi-branching according to claim 1, characterized in that: S2 includes the following steps: S2.1: Perform similarity fusion on multi-omics data of cell lines to obtain similarity features of cell lines; S2.2: Calculate drug similarity characteristics based on drug substructure fingerprints; S2.3: Calculate the global characteristics of the drug based on the molecular graph data of the drug; Among them, S2.1 includes the following steps: The gene expression data is Gaussian regularized, which can be described as: exp i represents the expression value of the ith gene in the cell line, represents the regularized expression value of the ith gene, μ i and δ i represent the mean and standard deviation of the expression of the ith gene in all cell lines, respectively; Calculate the similarity of each omics data of the cell line separately. This operation can be described as: x i and x j denote the i-th and j-th specific omics data, 1≤i,j≤962, ||||2 denotes the second normal form, ε is a hyperparameter, E ij Indicates the similarity of specific omics data in cell lines; The full kernel matrix F and sparse kernel matrix S are calculated for each omics data. This operation can be described as: N i represents the top N cell lines most similar to the i-th cell line; According to the calculated full nuclear matrix GF and sparse nuclear matrix GS of gene expression, full nuclear matrix CF and sparse nuclear matrix CS of copy number variation, full nuclear matrix MF and sparse nuclear matrix MS of somatic mutation, let GF0=GF, CF0=CF, MF0=MF, and iteratively update the full nuclear matrix. This operation can be described as: t represents the number of iterations, * T Indicates transposing the matrix* Finally, the cell line similarity feature RC is obtained, and the calculation formula is as follows: Among them, S2.2 includes the following steps: According to the substructure fingerprint of the drug, the Tanimoto similarity is calculated to obtain the drug similarity feature RD. This operation can be described as: d i and d j Represent the substructure fingerprints of the i-th and j-th drugs, respectively, 1≤i,j≤228; Among them, S2.3 includes the following steps: Extract features from the drug’s molecular graph to obtain the drug’s global features f molD , the operation can be described as: is the feature of node i at the k+1th layer, represents the set of neighbor nodes of node i, W (k) is the weight matrix of the kth layer, σ is the activation function, is the attention coefficient between node i and its neighbor node j, and is calculated as follows: α (k) is the learned attention weight vector, || represents the concatenation operation, and LeakyRelu is the activation function.

4. The method for predicting anticancer drug response based on heterogeneous and homogeneous multi-branching according to claim 1, characterized in that: S3 includes the following steps: S3.1: Extract features of cell line similarity features and drug similarity features, and obtain branch features of cell line subgraphs and branch features of drug subgraphs by fusing features of subgraphs of different scales; S3.2: extracting cell line similarity features and drug similarity features through message passing to obtain cell line heterogeneous branch features and drug heterogeneous branch features respectively; S3.3: Extract features in the cell line isomorphism graph to obtain cell line isomorphism branch features; S3.4: Extract features in the drug isomorphism graph to obtain drug isomorphism branch features; Among them, S3.1 includes the following steps: Subgraphs of different sizes are constructed by downsampling, and features are extracted from the subgraphs. The features of subgraphs of different sizes are fused to obtain the branch features of the cell line subgraphs f UC and drug subgraph branch feature f UD , the operation can be described as: y=X l p l / ||p l || idx=rank(y,k) A l+1 =A l (idx,idx) X l represents the node features of the lth layer, p l is a trainable vector, ||p l || represents vector p l The length of y is X l In the vector p l The projection scalar value on , rank(y,k) means selecting the largest k values ​​from y, and idx records the subscript of the largest k values. is the node feature corresponding to k nodes, A l+1 is the adjacency matrix of the corresponding nodes, W l is the weight matrix of layer l, H (l+1) is the updated node feature, σ is the activation function; Among them, S3.2 includes the following steps: In the cell line-drug heterogeneous graph, feature extraction is performed using message passing, and the cell line heterogeneous branch features f are obtained respectively. HeC and drug isomeric branching characteristics f HeD , the operation can be described as: here represents the normalized adjacency matrix, σ is the activation function; Among them, S3.3 includes the following steps: The cell line isomorphism branch features f are extracted using message passing in the cell line isomorphism graph HoC , the operation can be described as: A c =A*A T A is the cell line-drug response matrix, D c Yes A c The degree matrix of is the normalized adjacency matrix in the cell line isomorphism graph, σ is the activation function; Among them, S3.4 includes the following steps: The drug isomorphism branch features f are extracted using message passing in the drug isomorphism graph HoD , the operation can be described as: A d =A T *A A is the cell line-drug response matrix, D d Yes A d The degree matrix of is the normalized adjacency matrix in the drug isomorphism graph, and σ is the activation function.

5. The method for predicting anticancer drug response based on heterogeneous and homogeneous multi-branching according to claim 1, characterized in that: S4 includes the following steps: S4.1: Fusion of each branch feature of the cell line to obtain the cell line fusion feature; The drug fusion characteristics are obtained by integrating the characteristics of each branch of the drug and the global characteristics of the drug; S4.2: Linearly transform the cell line fusion signature and drug fusion signature to adjust them to the same dimension; S4.3: Calculate the correlation between the adjusted cell line fusion signature and the drug fusion signature; Among them, S4.1 includes the following steps: The cell lineage subgraph branch features f UC , cell line heterogeneous branching characteristicsf HeC , cell line isomorphic branching characteristicsf Hoc Fusion to obtain cell line fusion characteristics H 1-C , the drug subgraph branch feature f UD , drug isomerization branch characteristics f HeD , drug isomorphic branch characteristics f HoD , drug global characteristics f molD Fusion obtains drug fusion feature H 1-d , the operation can be described as: H 1-c =(f UC +f HeC +f HoC ) / 3 H 1-d =((f UD +f HeD ) / 2+f HoD +f molD ) / 3 Among them, S4.2 includes the following steps: The cell line fusion features and drug fusion features are adjusted to the same dimension through linear transformation, and then the correlation between the two after the transformation is calculated. This operation can be described as: H (2-c) =H (1-c) W (2-c) H (2-d) =H (1-d) W (2-d) Here, W (2-c) ∈R c*k and W (2-d) ∈R d*k are the weight matrices for linear transformation of cell lines and drugs, respectively. c is the dimension of cell line fusion features, d is the dimension of drug fusion features, k represents the feature dimension after linear transformation, and H (2-c) and H (2-d) They are the k-dimensional cell line fusion features and k-dimensional drug fusion features after linear transformation; Among them, S4.3 includes the following steps: Calculate the correlation between the k-dimensional cell line fusion features and the k-dimensional drug fusion features. This operation can be described as: h i is the i-th cell line feature in the k-dimensional cell line fusion feature, h j is the jth drug feature in the k-dimensional drug fusion feature, μ i and μ j They are h i and h j The mean of , 1≤i≤962, 1≤j≤228.

6. The method for predicting anticancer drug response based on heterogeneous and homogeneous multi-branching according to claim 1, characterized in that: S5 includes the following steps: S5.1: Perform sigmoid activation output on the correlation result and output the prediction result; Among them, S5.1 includes the following steps: The correlation coefficient calculated by S4 is in the range of [-1, 1]. The sigmoid function is used for activation output to ensure that the predicted probability value range of each pair of cell line-drug sensitivity is [0, 1], and the cell line-drug response matrix is ​​obtained. This operation can be described as: h is the correlation coefficient, γ is the predicted cell line-drug response matrix, and α is a hyperparameter used to accelerate parameter updates.

Images