Drug response prediction method based on graph neural network and convolutional neural network

Abstract

The invention discloses a drug response prediction method based on a graph neural network and a convolutional neural network, and belongs to the technical field of drug response prediction, and the method comprises the following steps: S1, data preprocessing; s2, constructing a network; s3, performing network training; and S4, drug reaction prediction. According to the method, the convolutional neural network and the graph neural network are combined, the drug molecule graph data are added, and the structural features of drug molecules are introduced, so that the final regression prediction is more accurate; feature extraction is carried out on drug molecule fingerprint global features derived from the sequence by using a stacked one-dimensional convolution module, and neighborhood information is enhanced; through a designed information interaction fusion branch, information interaction between drug characteristics and cell line characteristics is realized, interaction between drug reaction information and drug smiles sequence information is also realized, and the prediction performance of the model is improved through an obtained new characteristic vector.

Description

Technical Field

[0001] The present invention relates to the technical field of drug reaction prediction, and in particular to a drug reaction prediction method based on graph neural network and convolutional neural network. Background Art

[0002] The heterogeneity of cancer greatly affects the treatment outcomes of cancer. Predicting drug responses in vitro is expected to help develop personalized treatment plans. In recent years, several computational models based on machine learning and deep learning have been proposed to predict in vitro drug responses. However, most of these methods capture drug features based on a single drug description (e.g., drug structure) without considering the relationship between drugs and biological entities (e.g., targets, diseases, and side effects). In addition, most of these methods collect the characteristics of drugs and cell lines separately without considering the pairwise interactions between drugs and cell lines.

[0003] Therefore, being able to predict drug responses in vitro is expected to help doctors target specific treatments for different types of tumors while minimizing drug toxicity to patients and saving medical expenses. Recently, high-throughput screening technologies have paved the way for researchers to analyze cancer genomic patterns and measure the sensitivity of candidate drugs in vitro. Therefore, accurately and robustly discovering anticancer drug responses based on patients' clinical molecular characteristics has become a major challenge in the era of precision medicine. Accurate prediction of unknown cell lines or cases based on multi-patient omics data and known drug response data is an effective means to address this challenge. In addition, the introduction of graph neural networks on the basis of traditional convolutional neural networks and the extraction of features of different structures are conducive to improving the accuracy of drug response prediction. Moreover, when predicting drug response, a variety of drug-related or cell line-related data of primers are also conducive to the prediction of drug response.

[0004] In summary, the laboratory method of in vivo validation is very time-consuming and labor-intensive. Using only drug data and cell line data ignores the information interaction between the two. Using a single convolutional neural network or multi-layer perceptron network ignores the structural information of drug molecules. To this end, a drug response prediction method based on graph neural network and convolutional neural network is proposed. Summary of the invention

[0005] The technical problem to be solved by the present invention is: how to extract representative features about drugs and cell lines, and when obtaining representative features, how to make the two representative features interact with each other, realize automatic prediction through machines, and reduce laboratory research costs. A drug response prediction method based on graph neural networks and convolutional neural networks is provided.

[0006] The present invention solves the above technical problems through the following technical solutions, and the present invention comprises the following steps:

[0007] S1: Data preprocessing

[0008] Obtain the benchmark data set of the cell lines required for the smile sequences of the drugs, and preprocess the smile sequences and missing values ​​of the drugs according to the task requirements;

[0009] S2: Network construction

[0010] Construct a drug response prediction network, which includes multiple input branches, splicing modules and regressors. Each input branch obtains a corresponding feature vector, which is spliced ​​and modified by the splicing module and then input into the regressor for prediction. The input branch includes two information interaction fusion branches and one drug molecule graph extraction branch. The information interaction fusion branch uses a multi-layer perceptron and a two-dimensional convolutional neural network to perform interactive fusion of drug information itself and interactive fusion of drug information with cell line information. The drug molecule graph extraction branch uses a graph isomorphism neural network to extract the feature vector of the drug molecule graph.

[0011] S3: Network training

[0012] The data set obtained after preprocessing is divided according to a set ratio, and the drug response prediction network is trained using the training set to obtain a trained drug response prediction model that meets the performance index;

[0013] S4: Prediction of drug response

[0014] The cell line data in the test set and the unknown drug data are input into the drug response prediction model for prediction to obtain the prediction results.

[0015] Furthermore, in step S1, the specific processing process is as follows:

[0016] S11: Encode the smiles sequence of the drug to obtain 6 drug molecular fingerprint matrices, and construct a drug molecular map based on the chemical structure of the smiles sequence itself. At the same time, introduce 6 drug and biological entity related information matrices, combined with 4 cell line matrices, a total of 16 data matrices are obtained, of which 6 drug molecular fingerprint matrices and 6 drug and biological entity related information matrices are all drug matrices, a total of 12, and the drug and biological entity related information matrix is ​​also the drug entity matrix;

[0017] S12: Eliminate the data with missing values ​​for the reaction between the cell line and the drug to obtain the final data set;

[0018] S13: The 16 data matrices are processed using Chebyshev distance to obtain corresponding 16 similarity matrices.

[0019] Furthermore, in the step S11, the six drug molecule fingerprint matrices are respectively an extended connection fingerprint matrix, a Pubchem substructure fingerprint matrix, a daylight fingerprint matrix, an RDKit 2D normalized fingerprint matrix, an interpretable substructure partition fingerprint matrix and an extended simple graph fingerprint matrix; the six drug entity matrices are respectively a joint interaction matrix between drugs, a drug-target interaction matrix, an interaction matrix between drugs and known diseases, an interaction matrix between drugs and miRNA, an interaction matrix between drugs and cancer, and an IC50 value matrix of drug and cell line pairs; the four cell line matrices are respectively a gene expression data matrix, a copy number mutation data matrix, a gene mutation data matrix, and an IC50 value matrix of cell line and drug pairs.

[0020] Furthermore, in step S13, the calculation formula of the similarity matrix is ​​as follows:

[0021]

[0022] Among them, V x and V y represent the x-th and y-th rows of the matrix respectively, Respectively represent V x and V y The zth element of .

[0023] Furthermore, in the step S2, the first information interaction fusion branch is used to realize the information interaction fusion of the drug information itself, including a first inner product unit and a first outer product unit;

[0024] In the first inner product unit:

[0025] Embedding vector d for 6 drug entities i Embedded with 6 drug molecular fingerprint vectors d j Perform inner product operation:

[0026]

[0027] Wherein, ⊙ represents the inner product operation, i and j represent the numbers of the drug entity matrix embedding vector and the drug molecule fingerprint embedding vector during the inner product operation;

[0028] After the inner product operation, 36 different interaction vectors are obtained to represent the inner product output of the drug entity matrix embedding vector and the drug molecule fingerprint embedding vector;

[0029] Using Multi-layer Perceptron MLP inner1 Perform feature extraction on the inner product output obtained above to obtain the first inner product feature vector, where the multi-layer perceptron MLP inner1It includes 4 fully connected layers, and the number of neurons in the 4 fully connected layers are 1024, 1024, 512, and 128 respectively;

[0030] In the first outer product unit:

[0031] Embedding vector d for 6 drug entities i Embedded with 6 drug molecular fingerprint vectors d j Perform outer product operation:

[0032]

[0033] in, represents the outer product operation;

[0034] After the outer product operation, a set of feature interaction graphs with 36 channels are obtained to represent the outer product output of the drug entity matrix embedding vector and the drug molecule fingerprint embedding vector;

[0035] Use the two-dimensional convolutional neural network CNN1 to perform feature extraction operation on the outer product output obtained above to obtain the first outer product feature vector, wherein the two-dimensional convolutional neural network CNN1 includes two groups of residual blocks, and the corresponding feature extraction process is as follows:

[0036] g i+1 =Triplet Attention(γ(g i ))+Triplet Attention(δ(g i ,W i ′))

[0037] Among them, the residual block includes the direct mapping part γ(g i ) and the residual part δ(g i ,W i ′), a triple attention mechanism Triplet Attention is added at the end of the direct mapping part and the residual part.

[0038] Furthermore, in the step S2, the second information interaction fusion branch is used to realize the information interaction fusion between the cell line and the drug, including a second inner product unit and a second outer product unit;

[0039] In the second inner product unit:

[0040] For the 12 drug embedding vectors d′ x Embedded vector c′ with 4 cell lines y Perform inner product operation:

[0041]

[0042] Among them, ⊙ represents the inner product operation, x and y represent the numbers of the drug embedding vector and the cell line embedding vector during the inner product operation;

[0043] After the inner product operation, 48 different interaction vectors are obtained to represent the inner product output of the drug embedding vector and the cell line embedding vector;

[0044] Using Multi-layer Perceptron MLP inner2 Perform feature extraction on the inner product output obtained above to obtain the second inner product feature vector, where the multi-layer perceptron MLP inner2 It includes 4 fully connected layers, and the number of neurons in the 4 fully connected layers are 1024, 1024, 512, and 128 respectively;

[0045] In the second outer product unit:

[0046] For the 12 drug embedding vectors d′ x Embedded vector c′ with 4 cell lines y Perform outer product operation:

[0047]

[0048] in, represents the outer product operation;

[0049] After the outer product operation, a set of feature interaction graphs with 48 channels are obtained to represent the outer product output of the drug embedding vector and the cell line embedding vector;

[0050] Use the two-dimensional convolutional neural network CNN2 to perform feature extraction on the outer product output obtained above to obtain the second outer product feature vector, wherein the two-dimensional convolutional neural network CNN2 includes two groups of residual blocks, and the corresponding feature extraction process is as follows:

[0051] f i+1 =Triplet Attention(α(f i ))+Triplet Attention(β(f i ,W i ))

[0052] Among them, the residual block includes the direct mapping part α(f i ) and the residual part β(f i ,W i ), a triple attention mechanism Triplet Attention is added at the end of the direct mapping part and the residual part.

[0053] Furthermore, the drug embedding vector includes a drug entity embedding vector and a drug molecule fingerprint embedding vector. The drug embedding vector is obtained as follows: a similarity matrix corresponding to a drug matrix is ​​processed according to a set batch size to obtain a preliminary embedding vector, and the preliminary embedding vector is multiplied by a corresponding drug feature transformation matrix to obtain a drug embedding vector;

[0054] The cell line embedding vector is obtained as follows: the similarity matrix corresponding to the cell line matrix is ​​processed according to the set batch size to obtain a preliminary embedding vector, and the preliminary embedding vector is multiplied by the corresponding cell line feature transformation matrix to obtain the cell line embedding vector.

[0055] Furthermore, in the step S2, the drug reaction prediction network further includes a fusion feature extraction branch, a drug entity matrix global feature extraction branch, a drug molecule fingerprint matrix global feature extraction branch and a cell line matrix global feature extraction branch;

[0056] In the fusion feature extraction branch, the SNF algorithm is used to fuse 12 drug matrices to obtain a fusion matrix Fuse feature, and the multi-layer perceptron MLP is used to extract the drug matrix. fuse Perform feature extraction to obtain a fused feature vector, where the multi-layer perceptron MLP fuse It includes 2 fully connected layers, and the number of neurons in the 2 fully connected layers is 128 and 128 respectively;

[0057] In the drug entity matrix global feature extraction branch, for the drug entity total matrix d total , using a multi-layer perceptron MLP drug Extract the overall features and obtain the first drug global feature vector, where the multi-layer perceptron MLP drug It includes two fully connected layers, the number of neurons in the two fully connected layers is 128 and 128 respectively, and the total matrix of drug entities d total The similarity matrices corresponding to the six drug entity matrices were obtained by splicing;

[0058] In the drug molecule fingerprint matrix global feature extraction branch, the similarity matrices corresponding to the 6 drug molecule fingerprint matrices obtained after encoding are spliced ​​to obtain a spliced ​​drug molecule fingerprint total matrix, and feature extraction is performed using a three-stacked one-dimensional convolution module to obtain a second drug global feature vector;

[0059] In the cell line matrix global feature extraction branch, for the cell line total matrix C total , using a multi-layer perceptron MLP cellExtract the overall features and obtain the global feature vector of the cell line, where the multi-layer perceptron MLP cell It includes two fully connected layers, the number of neurons in the two fully connected layers is 128 and 128 respectively, and the total matrix C of the cell line total The similarity matrices corresponding to the four cell line matrices were obtained by splicing.

[0060] Furthermore, in step S2, a three-layer graph isomorphic neural network is used to extract drug molecule graph data, and the output of each layer is spliced, and skip connections and a global maximum pooling layer are added to the network for processing.

[0061] Furthermore, in step S3, the mean square error loss function is used as the loss function of network training:

[0062]

[0063] Among them, f(x) represents the predicted value, y represents the true value, and n represents the number of samples.

[0064] Compared with the prior art, the present invention has the following advantages: the drug response prediction method based on graph neural network and convolutional neural network adopts a method combining convolutional neural network and graph neural network, adds drug molecule graph data, and introduces the structural characteristics of drug molecules, so that the final regression prediction is more accurate; for the global features of drug molecule fingerprints derived from sequences, stacked one-dimensional convolution modules are used for feature extraction to strengthen neighborhood information; through the designed information interaction fusion branch, not only the information interaction between drug features and cell line features is realized, but also the interaction between drug response information and drug smiles sequence information is realized, and the obtained new feature vector improves the prediction performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 is a flow chart of a drug reaction prediction method based on a graph neural network and a convolutional neural network in an embodiment of the present invention;

[0066] Figure 2 is a schematic diagram of the structure of a downsampling module in an embodiment of the present invention;

[0067] Figure 3 It is a schematic diagram of the overall structure of the drug response prediction model in an embodiment of the present invention;

[0068] Figure 4 is a schematic diagram of the structure of a triple attention module in an embodiment of the present invention;

[0069] Figure 5 It is a visualization effect diagram of the prediction result in the embodiment of the present invention. DETAILED DESCRIPTION

[0070] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiment.

[0071] like Figure 1 As shown, this embodiment provides a technical solution: a drug reaction prediction method based on graph neural network and convolutional neural network, comprising the following steps:

[0072] S1: Data preprocessing

[0073] In step S1, the following sub-steps are included:

[0074] S11: Encode the smiles sequence of the drug to obtain 6 drug molecular fingerprint matrices, and construct a drug molecular map based on the chemical structure of the smiles sequence itself. At the same time, introduce 6 drug and biological entity related information matrices, combined with 4 cell line matrices, a total of 16 data matrices are obtained, of which 6 drug molecular fingerprint matrices and 6 drug and biological entity related information matrices (drug entity matrices) are all drug matrices, a total of 12;

[0075] S12: Eliminate the data with missing values ​​for the reaction between the cell line and the drug to obtain the final data set;

[0076] S13: For the above 16 data matrices, the Chebyshev distance is used to process the above matrices to obtain the similarity matrix of the 16 data matrices:

[0077]

[0078] Among them, V x and V y represent the x-th and y-th rows of the matrix respectively, Respectively represent V x and V y The zth element of . The 16 similarity matrices obtained are used as part of the data finally input into the model.

[0079] In the present embodiment, in the step S11, the six drug molecule fingerprint matrices are respectively the extended connection fingerprint matrix (ECFP), the Pubchem substructure fingerprint matrix (PSFP), the daylight fingerprint matrix (DFP), the RDKit 2D normalized fingerprint matrix (RDKFP), the interpretable substructure partition fingerprint matrix (ESPFP) and the extended simple graph fingerprint matrix (ERGFP); the six drug and biological entity related information matrices are respectively the combined interaction matrix (Combine) between drugs, the drug and target interaction matrix (target), the interaction matrix between drugs and known diseases (disease), the interaction matrix between drugs and miRNA (mirna), the interaction matrix between drugs and cancer (adr), and the IC50 value matrix of drugs and cell lines; the benchmark data set used includes 4 cell line matrices, namely the gene expression data matrix, the copy number mutation data matrix, the gene mutation data matrix, and the IC50 value matrix of cell lines and drug pairs. In the present embodiment, the names of the 16 data matrices used are shown in Table 1:

[0080] Table 1 Names of drug response matrix and cell line matrix

[0081]

[0082]

[0083] Among them, M combined : Combined interactions between drugs; M disease : Known associations between drugs and 5181 diseases; M mirna : known drug-microRNA associations; M adr :Drug-cancer association; M target : Interaction between drug and target; M dc : IC50 value of drug and cell line pair; M rdkfp :RDKit 2D normalized fingerprint; M espfp : Interpretable substructure partition fingerprint map; M ergfp : Extended simplified graph fingerprint; M ecfp : Extended connection fingerprint map; M psfp : PubChem substructure fingerprint; M dfp : Sunlight fingerprint; M cd : IC50 value of the reaction between the cell line and the drug pair; M mu : Cell line gene mutation data matrix; M exp : Cell line gene expression data matrix; M cnv : Cell line copy number mutation data matrix.

[0084] S2: Extract drug characteristics and cell line characteristics;

[0085] In step S2, the following sub-steps are included:

[0086] S21: An information interaction module is designed for the drug entity matrix and the drug molecule fingerprint matrix. The drug entity matrix and the drug molecule fingerprint matrix are both R n×170 , and the drug molecule fingerprint matrix is ​​mostly a binary matrix, which can retain the structural information of the drug molecule to the greatest extent. In this paper, the drug entity embedding vector is named {d1, d2, d3, d4, d5, d6}, and the drug molecule fingerprint embedding vector is named {d7, d8, d9, d 10 ,d 11 ,d 12 The present invention designs an inner product unit and an outer product unit for information interaction between the two. In the inner product unit, the present invention embeds each drug entity with a vector d i Embedded vector with drug molecular fingerprint j Perform inner product operation. The operation of inner product operation is as follows:

[0087]

[0088] Wherein, ⊙ represents the inner product operation, i and j represent the numbers of the drug entity matrix embedding vector and the drug molecule fingerprint embedding vector during the inner product operation. Here, the present invention obtains 36 different interaction vectors to represent the inner product output between the two vectors. The present invention uses a multi-layer perceptron MLP inner1 The inner product output obtained above is extracted to obtain the first inner product feature vector (Inner embeddings1). The MLP used in the present invention inner1 It consists of 4 fully connected layers, and the number of neurons in the four fully connected layers are 1024, 1024, 512, and 128 respectively.

[0089] In the outer product unit, the present invention embeds each drug entity with a vector d i Embedded vector with drug molecular fingerprint j Perform outer product operation. The operation of outer product operation is as follows:

[0090]

[0091] in, Represents an outer product operation. Here, the present invention obtains a set of feature interaction graphs with a channel number of 36. Then, the present invention uses a convolutional neural network CNN1 to perform a feature extraction operation on this set of feature graphs to obtain a second inner product feature vector Outer embeddings1. The convolutional neural network CNN1 designed by the present invention is composed of two groups of residual blocks. The feature extraction process is as follows:

[0092] g i+1 =Triplet Attention(γ(g i ))+Triplet Attention(δ(g i ,W i ′))

[0093] The residual block consists of a direct mapping part γ(g i ) and the residual part δ(g i ,W i ′), and similarly, the present invention adds a triple attention mechanism (Triplet Attention module) at the end of the direct mapping part and the residual part to improve the expression ability of the output features.

[0094] S22: Similarly, the present invention also designs an inner product unit and an outer product unit for the information interaction between the cell line and the drug. In the inner product unit, the present invention embeds the 12 drug vectors d′ x Embedded vector c′ with 4 cell lines y Perform inner product operation. The operation of inner product operation is as follows:

[0095]

[0096] Wherein, ⊙ represents the inner product operation, x and y represent the numbers of the drug embedding vector and the cell line embedding vector during the inner product operation. Here, the present invention obtains 48 different interaction vectors to represent the inner product output of the drug embedding vector and the cell line embedding vector. The present invention uses a multi-layer perceptron MLP inner2 The inner product output obtained above is extracted to obtain the second inner product feature vector (Inner embeddings2). The multi-layer perceptron MLP used in the present invention inner2 It consists of 4 fully connected layers, and the number of neurons in the four fully connected layers are 1024, 1024, 512, and 128 respectively.

[0097] In the outer product unit, the present invention embeds each drug vector d′ x and the cell line embedding vector c′ y Perform outer product operation. The operation of outer product operation is as follows:

[0098]

[0099] in, Represents the outer product operation. Here, a set of feature interaction graphs with a channel number of 48 is obtained. Then, the present invention uses the convolutional neural network CNN2 to perform feature extraction operations on this set of feature graphs to obtain the second outer product feature vector (Outerembeddings2). The convolutional neural network CNN2 designed by the present invention is composed of two sets of residual blocks. The feature extraction process is as follows:

[0100] f i+1 =Triplet Attention(α(f i ))+Triplet Attention(β(f i ,W i ))

[0101] The residual block consists of the direct mapping part α(f i ) and the residual part β(f i ,W i ), and the present invention adds a triple attention mechanism at the end of the direct mapping part and the residual part. The triple attention mechanism can effectively capture the interaction information between the three dimensions of C, H, and W, making the output features more expressive. In addition, the attention mechanism consumes less computing resources, which saves computing resources while maximizing the prediction accuracy of the model.

[0102] The triple attention mechanism is as follows Figure 4 As shown in the figure, the first branch is the channel attention calculation branch. The input features are Z-Pooled, followed by 7x 7 convolution, and finally the sigmoid activation function generates the spatial attention weights. The second branch is the channel C and spatial W dimension interaction capture branch. The input features are first permuted to become H×C×W dimensional features, then Z-Pooled on the H dimension, followed by 7x 7 convolution, and finally the sigmoid activation function is used to generate the spatial attention weights; finally, the tensor needs to be permuted to become C×H×W dimensional features for element-wise addition. The third branch: channel C and spatial H dimension interactive capture branch, the input feature first undergoes the dimension operation (permuter) of the rearranged tensor to become a W×H×C dimension feature, then performs Z-Pool on the W dimension, followed by 7x 7 convolution, and finally the Sigmoid activation function generates the spatial attention weight; finally, it needs to undergo the dimension operation (permuter) of the rearranged tensor to become a C×H×W dimension feature for element-wise addition. Finally, the output features of the three branches are added and averaged (Avg).

[0103] Among them, element-wise: between each feature map, the elements at the corresponding positions are calculated, including addition, subtraction, multiplication, and division. The Z-pool layer is responsible for reducing the C-dimensional tensor to 2 dimensions, connecting the average pooled features and the maximum pooled features on this dimension. This enables the layer to retain the rich representation of the actual tensor while reducing its depth to make further calculations lighter.

[0104] It can be seen from Table 2 that the information fusion method of the present invention is more accurate for the final prediction than the traditional information interaction fusion method.

[0105] Table 2 Comparison of different fusion methods

[0106] Interaction mode RMSE MAE R AddFusion 0.944 0.707 0.941 ConcatFusion 0.904 0.678 0.946 MultiplyFusion 0.982 0.733 0.937 OursFusion 0.821 0.615 0.956

[0107] In this embodiment, in steps S21 and S22:

[0108] For the feature extraction of cell lines, the present invention designs four cell line feature transformation matrices {C1, C2, C3, C4}∈R n ×s , where s is set to 128, then the preliminary embedding vector c of the i-th cell line matrix is i Transformed cell line embedding vector c i ′(i.e. c′ y ) can be calculated as:

[0109] c i ′=c i C i

[0110] After the above mapping process, the present invention maps the cell line features to the same 128-dimensional embedding vector as the drug features, preparing for subsequent feature interaction.

[0111] On the other hand, the similarity matrices corresponding to the four cell line matrices are concatenated (rows remain unchanged and columns are increased) to obtain the total cell line matrix C total , MLP is used to extract the overall features of the cell line. At this stage, there are two fully connected layers in the MLP network, and the number of neurons in the fully connected layers is 128 and 128 respectively.

[0112] It should be noted that the cell line embedding vector is obtained as follows: the similarity matrix corresponding to the cell line matrix is ​​processed according to the set batch size to obtain a preliminary embedding vector, and the preliminary embedding vector is multiplied by the corresponding cell line feature transformation matrix to obtain the cell line embedding vector.

[0113] Similarly, 12 drug feature transformation matrices are used, namely {A1, A2, ..., A12}∈R n×s , where s is set to 128, then the initial embedding vector d of the i-th drug matrix is i Transformed drug embedding vector d i ′(i.e. d′ x ) can be calculated as:

[0114] d i ′=d i A i

[0115] After the above mapping process, the present invention can transform the initial embedding vector d of the drug matrix through 12 drug feature transformation matrices. i The dimensions of are mapped into the same vector space.

[0116] On the other hand, the total drug entity matrix d is obtained by concatenating the similarity matrices corresponding to the six drug entity matrices (rows remain unchanged and columns are increased during concatenation). total The present invention uses an MLP network to extract the overall characteristics of drug entity information. In this stage, there are two fully connected layers in the MLP network, and the number of neurons in the fully connected layers is 128 and 128 respectively.

[0117] It should be noted that the drug embedding vector (including drug entity embedding vector and drug molecule fingerprint embedding vector) is obtained as follows: the similarity matrix corresponding to the drug matrix is ​​processed according to the set batchsize to obtain a preliminary embedding vector, and the preliminary embedding vector is multiplied by the corresponding drug feature transformation matrix to obtain the drug embedding vector.

[0118] In this embodiment, in step S2, the drug response prediction network further includes a fusion feature extraction branch, a drug entity matrix global feature extraction branch, a drug molecule fingerprint matrix global feature extraction branch, and a cell line matrix global feature extraction branch;

[0119] In the fusion feature extraction branch, the present invention uses the SNF algorithm to fuse 12 drug matrices (original drug matrices without Cherchev distance processing), and adds the similarity information between the drug matrices to the regression prediction model of the present invention. This process can be described as:

[0120] Fuse feature=SNF([D1,D2,D3,D4,D5,D6,D7,D8,D9,D 10 ,D 11 ,D 12 ]),D i ∈R 170×170

[0121] Among them, D1~D 12are the corresponding 12 original drug matrices;

[0122] Here i∈{1…12}, the present invention will eventually obtain a fusion matrix Fusefeature∈R 170×170 For this fusion matrix, the present invention chooses to use MLP fuse It extracts features and obtains fused feature vectors (Fuseembeddings), using the MLP fuse It consists of two fully connected layers, where the number of neurons in the two fully connected layers is 128 and 128 respectively. This process can be described as:

[0123] Fuse out =Linear(ReLU(BN(Linear(Fuse_feature))))

[0124] Among them, BN represents Batchnorm layer, ReLU represents activation function, and Linear represents fully connected layer

[0125] In the drug entity matrix global feature extraction branch, for the total drug entity matrix d total ∈R n×1020 , the present invention uses a multi-layer perceptron MLP drug The overall features are extracted to obtain the first drug global feature vector (Drug1embeddings). In this stage, there are two fully connected layers in the MLP network, and the number of neurons in the fully connected layers is 128 and 128 respectively. total It can be calculated as:

[0126] d′ total =Linear(ReLU(BN(Linear(d total ))))

[0127] Among them, BN represents Batchnorm layer, ReLU represents activation function, and Linear represents fully connected layer.

[0128] In the global feature extraction branch of the drug molecule fingerprint matrix, the similarity matrices corresponding to the six drug molecule fingerprint matrices obtained after encoding are spliced ​​(rows remain unchanged and columns are increased during splicing) to obtain the spliced ​​drug molecule fingerprint total matrix, and feature extraction is performed using a three-stacked one-dimensional convolution module to obtain the second drug global feature vector (Drug2embeddings). The above encoding process uses the toolkit in the RDKit third-party library in Python to encode the smiles sequence and obtain a set of unique representations containing drug smiles sequences.

[0129] In the cell line matrix global feature extraction branch, for the cell line total matrix C total ∈R n×2698 , using MLP cell The overall features of the cell line are extracted to obtain the global feature vector of the cell line (Cells embeddings). At this stage, there are two fully connected layers in the MLP network, and the number of neurons in the fully connected layers is 128 and 128 respectively. t ' otal It can be calculated as:

[0130] c t ' otal =Linear(ReLU(BN(Linear(c total ))))

[0131] Among them, BN represents Batchnorm layer, ReLU represents activation function, and Linear represents fully connected layer.

[0132] In this embodiment, in step S2, in the two-dimensional convolutional neural network, two layers of convolutional neural network with residual structure (residual block) are used, and a point convolution layer with a convolution kernel size of 1 is adopted, and a triple attention module, a maximum pooling layer, and a batch normalization layer are introduced in the downsampling module for processing.

[0133] In this embodiment, in step S2, a three-layer graph isomorphic neural network is used to extract drug molecule graph data, and the output of each layer is spliced, and skip connections and a global maximum pooling layer are added to the network for processing.

[0134] In this embodiment, after each feature vector is obtained, it is spliced ​​and modified by a splicing module and then input into a regressor for prediction.

[0135] The drug response prediction model uses three evaluation indicators to evaluate the effect, namely RMSE: root mean square error; MAE: maximum average error; R: determination coefficient. As can be seen from Table 2, after adding the drug molecular graph data, the prediction ability of the model has been improved.

[0136] Table 2 Comparison between drug molecule graph data added and drug molecule graph data

[0137] Features RMSE MAE R Without Graph 0.825 0.624 0.955 Graph 0.821 0.615 0.956

[0138] In this embodiment, it can be seen from Table 4 that when the graph neural network is replaced with a graph convolutional neural network (GCN), a graph attention network (GAT), and a graph transformer, the root mean square error and the maximum average error both increase, reflecting the effectiveness of the graph isomorphic neural network (GIN).

[0139] Table 4 Performance of different graph neural networks

[0140] Graph Method RMSE MAE R Graph Transformer 0.831 0.618 0.954 GCN 0.831 0.627 0.954 GAT 0.830 0.624 0.955 GIN 0.821 0.615 0.956

[0141] S3: Model training

[0142] The training set, validation set, and test set are divided into 8:1;1 ratios for model training and prediction, and the mean square error loss function is selected as the final loss function.

[0143]

[0144] Among them, f(x) represents the predicted value, y represents the true value, and n represents the number of samples.

[0145] S4: Prediction of drug response

[0146] In step S4, drug response prediction is considered as a regression task. The value predicted by the model is the IC50 response value of the drug. The difference between the calculated value and the true value is calculated to obtain a more accurate prediction result. Figure 3 As shown. The initial learning rate is set to 0.0001, the weight decay rate is 0.0003, the inactivation rate is 0.5, and a total of 500 rounds of training are performed. However, if the loss function continues to increase for 10 consecutive epochs, the training is stopped to prevent overfitting. The cell line data and drug data of the test set are input into the drug response prediction model for prediction to obtain the prediction results. The visualization of the prediction results is shown in the figure below. Figure 5 shown.

[0147] In this embodiment, the results are shown in Table 5. It can be seen from Table 5 that the performance of the present invention is better than other methods in terms of the three evaluation indicators, reflecting the effectiveness of adding graph data and considering the interaction of multiple information, and also indicating that the one-dimensional convolution module is more suitable for extracting features from sequences. Compared with the second highest method MSDRP, the model improves by 1.4%, 1.1%, and 0.2% in root mean square error (RMSE), maximum average error (MAE), and determination coefficient (R), respectively.

[0148] Table 5 Performance of different methods

[0149]

[0150]

[0151] In summary, the drug reaction prediction method based on graph neural network and convolutional neural network in the above embodiment first performs necessary preprocessing operations on the public drug smiles sequence and 12 data matrices of drugs, 3 omics matrices of cell lines, and IC50 value matrices of drug and cell line pairs; for drug features, a graph isomorphism neural network is used to extract the features of the drug molecule graph, a one-dimensional convolution module is used to extract features of the drug molecule fingerprint from the drug smiles sequence itself, and a multi-layer perceptron is used to extract features of the information matrix related to drugs and biological entities; for cell line features, a multi-layer perceptron with strong mapping ability is selected for feature extraction. Extraction; for the interactive fusion of the two features, the inner product and outer product operations are performed on the drug features and the cell line features to obtain features of different structures, which are finally sent to the multi-layer perceptron and the two-dimensional convolutional neural network for feature extraction. In addition, the present invention takes into account the feature interaction between the drug and biological entity related data and the smiles sequence data of the drug itself, and performs the same outer product and inner product operations as above on the two; finally, the obtained multiple feature vectors are input into the final regressor for regression prediction; various experiments have shown the effectiveness of the method of the present invention and the rationality of the modules adopted. The proposed model has high prediction accuracy and is expected to be deployed on the equipment in the future for production and application.

[0152] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A drug response prediction method based on graph neural network and convolutional neural network, characterized in that: The following steps are involved: S1: Data preprocessing Obtain the benchmark data set of the cell lines required for the smile sequences of the drugs, and preprocess the smile sequences and missing values ​​of the drugs according to the task requirements; S2: Network construction Construct a drug response prediction network, which includes multiple input branches, splicing modules and regressors. Each input branch obtains a corresponding feature vector, which is spliced ​​and modified by the splicing module and then input into the regressor for prediction. The input branch includes two information interaction fusion branches and one drug molecule graph extraction branch. The information interaction fusion branch uses a multi-layer perceptron and a two-dimensional convolutional neural network to perform interactive fusion of drug information itself and interactive fusion of drug information with cell line information. The drug molecule graph extraction branch uses a graph isomorphism neural network to extract the feature vector of the drug molecule graph. S3: Network training The data set obtained after preprocessing is divided according to a set ratio, and the drug response prediction network is trained using the training set to obtain a trained drug response prediction model that meets the performance index; S4: Prediction of drug response The cell line data in the test set and the unknown drug data are input into the drug response prediction model for prediction to obtain the prediction results.

2. A drug reaction prediction method based on graph neural network and convolutional neural network according to claim 1, characterized in that: In step S1, the specific processing process is as follows: S11: Encode the smiles sequence of the drug to obtain 6 drug molecular fingerprint matrices, and construct a drug molecular map based on the chemical structure of the smiles sequence itself. At the same time, introduce 6 drug and biological entity related information matrices, combined with 4 cell line matrices, a total of 16 data matrices are obtained, of which 6 drug molecular fingerprint matrices and 6 drug and biological entity related information matrices are all drug matrices, a total of 12, and the drug and biological entity related information matrix is ​​also the drug entity matrix; S12: Eliminate the data with missing values ​​for the reaction between the cell line and the drug to obtain the final data set; S13: The 16 data matrices are processed using Chebyshev distance to obtain corresponding 16 similarity matrices.

3. A drug reaction prediction method based on graph neural network and convolutional neural network according to claim 2, characterized in that: In the step S11, the six drug molecule fingerprint matrices are respectively an extended connection fingerprint matrix, a Pubchem substructure fingerprint matrix, a daylight fingerprint matrix, an RDKit 2D normalized fingerprint matrix, an interpretable substructure partition fingerprint matrix and an extended simple graph fingerprint matrix; the six drug entity matrices are respectively a joint interaction matrix between drugs, a drug-target interaction matrix, an interaction matrix between drugs and known diseases, an interaction matrix between drugs and miRNA, an interaction matrix between drugs and cancer, and an IC50 value matrix of drug and cell line pairs; the four cell line matrices are respectively a gene expression data matrix, a copy number mutation data matrix, a gene mutation data matrix, and an IC50 value matrix of cell line and drug pairs.

4. A drug reaction prediction method based on graph neural network and convolutional neural network according to claim 2, characterized in that: In step S13, the calculation formula of the similarity matrix is ​​as follows: Among them, V x and V y represent the x-th and y-th rows of the matrix respectively, Respectively represent V x and V y The zth element of .

5. The drug reaction prediction method based on graph neural network and convolutional neural network according to claim 2, characterized in that: In the step S2, the first information interaction fusion branch is used to realize the information interaction fusion of the drug information itself, including a first inner product unit and a first outer product unit; In the first inner product unit: Embedding vector d for 6 drug entities i Embedded with 6 drug molecular fingerprint vectors d j Perform inner product operation: Wherein, ⊙ represents the inner product operation, i and j represent the numbers of the drug entity matrix embedding vector and the drug molecule fingerprint embedding vector during the inner product operation; After the inner product operation, 36 different interaction vectors are obtained to represent the inner product output of the drug entity matrix embedding vector and the drug molecule fingerprint embedding vector; Using Multi-layer Perceptron MLP inner1 Perform feature extraction on the inner product output obtained above to obtain the first inner product feature vector, where the multi-layer perceptron MLP inner1 It includes 4 fully connected layers, and the number of neurons in the 4 fully connected layers are 1024, 1024, 512, and 128 respectively; In the first outer product unit: Embedding vector d for 6 drug entities i Embedded with 6 drug molecular fingerprint vectors d j Perform outer product operation: in, represents the outer product operation; After the outer product operation, a set of feature interaction graphs with 36 channels are obtained to represent the outer product output of the drug entity matrix embedding vector and the drug molecule fingerprint embedding vector; Use the two-dimensional convolutional neural network CNN1 to perform feature extraction operation on the outer product output obtained above to obtain the first outer product feature vector, wherein the two-dimensional convolutional neural network CNN1 includes two groups of residual blocks, and the corresponding feature extraction process is as follows: g i+1 =Triplet Attention(γ(g i ))+Triplet Attention(δ(g i ,W i ′ )) Among them, the residual block includes the direct mapping part γ(g i ) and the residual part δ(g i ,W i ′ ), a triple attention mechanism Triplet Attention is added at the end of the direct mapping part and the residual part.

6. A drug reaction prediction method based on graph neural network and convolutional neural network according to claim 5, characterized in that: In the step S2, the second information interaction fusion branch is used to realize the information interaction fusion between the cell line and the drug, including a second inner product unit and a second outer product unit; In the second inner product unit: For 12 drug embedding vectors d ′ x Embedding vector c with 4 cell lines ′ y Perform inner product operation: Among them, ⊙ represents the inner product operation, x and y represent the numbers of the drug embedding vector and the cell line embedding vector during the inner product operation; After the inner product operation, 48 different interaction vectors are obtained to represent the inner product output of the drug embedding vector and the cell line embedding vector; Using Multi-layer Perceptron MLP inner2 Perform feature extraction on the inner product output obtained above to obtain the second inner product feature vector, where the multi-layer perceptron MLP inner2 It includes 4 fully connected layers, and the number of neurons in the 4 fully connected layers are 1024, 1024, 512, and 128 respectively; In the second outer product unit: For 12 drug embedding vectors d ′ x Embedding vector c with 4 cell lines ′ y Perform outer product operation: in, represents the outer product operation; After the outer product operation, a set of feature interaction graphs with 48 channels are obtained to represent the outer product output of the drug embedding vector and the cell line embedding vector; Use the two-dimensional convolutional neural network CNN2 to perform feature extraction on the outer product output obtained above to obtain the second outer product feature vector, wherein the two-dimensional convolutional neural network CNN2 includes two groups of residual blocks, and the corresponding feature extraction process is as follows: f i+1 =Triplet Attention(α(f i ))+Triplet Attention(β(f i ,W i )) Among them, the residual block includes the direct mapping part α(f i ) and the residual part β(f i ,W i ), a triple attention mechanism Triplet Attention is added at the end of the direct mapping part and the residual part.

7. A drug reaction prediction method based on graph neural network and convolutional neural network according to claim 6, characterized in that: The drug embedding vector includes a drug entity embedding vector and a drug molecule fingerprint embedding vector. The drug embedding vector is obtained as follows: a similarity matrix corresponding to a drug matrix is ​​processed according to a set batch size to obtain a preliminary embedding vector, and the preliminary embedding vector is multiplied by a corresponding drug feature transformation matrix to obtain a drug embedding vector; The cell line embedding vector is obtained as follows: the similarity matrix corresponding to the cell line matrix is ​​processed according to the set batch size to obtain a preliminary embedding vector, and the preliminary embedding vector is multiplied by the corresponding cell line feature transformation matrix to obtain the cell line embedding vector.

8. The drug reaction prediction method based on graph neural network and convolutional neural network according to claim 7, characterized in that: In step S2, the drug response prediction network further includes a fusion feature extraction branch, a drug entity matrix global feature extraction branch, a drug molecule fingerprint matrix global feature extraction branch, and a cell line matrix global feature extraction branch; In the fusion feature extraction branch, the SNF algorithm is used to fuse 12 drug matrices to obtain a fusion matrix Fuse feature, and the multi-layer perceptron MLP is used to extract the drug matrix. fuse Perform feature extraction to obtain a fused feature vector, where the multi-layer perceptron MLP fuse It includes 2 fully connected layers, and the number of neurons in the 2 fully connected layers is 128 and 128 respectively; In the drug entity matrix global feature extraction branch, for the drug entity total matrix d total , using a multi-layer perceptron MLP drug Extract the overall features and obtain the first drug global feature vector, where the multi-layer perceptron MLP drug It includes two fully connected layers, the number of neurons in the two fully connected layers is 128 and 128 respectively, and the total matrix of drug entities d total The similarity matrices corresponding to the six drug entity matrices were obtained by splicing; In the drug molecule fingerprint matrix global feature extraction branch, the similarity matrices corresponding to the 6 drug molecule fingerprint matrices obtained after encoding are spliced ​​to obtain a spliced ​​drug molecule fingerprint total matrix, and feature extraction is performed using a three-stacked one-dimensional convolution module to obtain a second drug global feature vector; In the cell line matrix global feature extraction branch, for the cell line total matrix C total , using a multi-layer perceptron MLP cell Extract the overall features and obtain the global feature vector of the cell line, where the multi-layer perceptron MLP cell It includes two fully connected layers, the number of neurons in the two fully connected layers is 128 and 128 respectively, and the total matrix C of the cell line total The similarity matrices corresponding to the four cell line matrices were obtained by splicing.

9. The drug reaction prediction method based on graph neural network and convolutional neural network according to claim 1, characterized in that: In step S2, a three-layer graph isomorphic neural network is used to extract drug molecule graph data, and the output of each layer is spliced, and skip connections and a global maximum pooling layer are added to the network for processing.

10. The drug reaction prediction method based on graph neural network and convolutional neural network according to claim 1, characterized in that: In step S3, the mean square error loss function is used as the loss function for network training: Among them, f(x) represents the predicted value, y represents the true value, and n represents the number of samples.

Images