A drug reaction prediction method based on a graph neural network and a convolutional neural network

Abstract

The application discloses a drug reaction prediction method based on a graph neural network and a convolutional neural network, and belongs to the technical field of drug reaction prediction, and comprises the following steps: S1, data preprocessing; S2, network construction; S3, network training; and S4, drug reaction prediction.The method combines the convolutional neural network and the graph neural network, adds drug molecule graph data, introduces the structural features of drug molecules, and makes the final regression prediction more accurate; a stacked one-dimensional convolution module is used for feature extraction on the global features of the drug molecule fingerprints derived from sequences, and the neighborhood information is strengthened; through the designed information interaction fusion branch, not only the information interaction between the drug features and the cell line features is realized, but also the interaction between the drug reaction information and the drug smiles sequence information is realized, and the obtained new feature vector improves the prediction performance of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of drug response prediction, and specifically relates to a drug response prediction method based on a graph neural network and a convolutional neural network. BACKGROUND

[0002] The heterogeneity of cancer greatly affects the treatment outcome of cancer. In vitro prediction of drug response 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 response. However, most of these methods capture drug features based on single drug descriptions (such as drug structure) without considering the relationship between drugs and biological entities (such as targets, diseases, and side effects). In addition, these methods mostly collect features of drugs and cell lines separately without considering the paired interaction between drugs and cell lines.

[0003] Therefore, the ability to predict drug response in vitro is expected to help doctors treat different types of tumors specifically while minimizing drug toxicity to patients and saving medical costs. Recently, high-throughput screening technology has paved the way for researchers to analyze cancer genomic patterns and measure candidate drug sensitivity in vitro. Therefore, accurately and robustly discovering anti-cancer drug responses based on patient clinical molecular characteristics has become a major challenge in the era of precision medicine. Based on the omics data of multiple patients and known drug response data, accurate and robust prediction of unknown cell lines or cases is an effective means to solve this challenge. In addition, introducing graph neural networks on the basis of traditional convolutional neural networks to extract features of different structures is conducive to improving the accuracy of drug response prediction. Moreover, using a variety of drug-related or cell line-related data when predicting drug response is also conducive to predicting drug response.

[0004] In summary, laboratory methods verified in vivo are time-consuming and labor-intensive, using only drug data and cell line data ignores the information interaction between the two, and using a single convolutional neural network or multilayer perceptron network ignores the structural information of drug molecules. Therefore, a drug response prediction method based on a graph neural network and a convolutional neural network is proposed. SUMMARY

[0005] The technical problem to be solved by the present application is how to extract representative features related to drugs and cell lines, how to enable the two representative features to interact when obtaining the representative features, how to automatically predict by machine, and how to reduce laboratory research costs. A drug response prediction method based on a graph neural network and a convolutional neural network is provided.

[0006] The present application solves the above technical problems by the following technical solutions. The present application comprises the following steps:

[0007] S1: data preprocessing

[0008] A reference dataset of cell lines required for obtaining the smiles sequence of the drug is obtained, and the smiles sequence of the drug and missing values are preprocessed according to the task requirements;

[0009] S2: network construction

[0010] A drug reaction prediction network is constructed, which includes multiple input branches, a splicing module and a regressor. Each input branch obtains a corresponding feature vector, which is input into the regressor through the splicing module for prediction work. The input branch includes two information interaction fusion branches and one drug molecule graph extraction branch. The information interaction fusion branch performs interaction fusion work of drug information itself and interaction fusion work of drug information and cell line information through a multilayer perceptron and a two-dimensional convolutional neural network. The drug molecule graph extraction branch extracts a drug molecule graph feature vector through a graph isomorphism neural network.

[0011] S3: network training

[0012] The data set obtained after preprocessing is divided according to the set proportion, and the drug reaction prediction network is trained using the training set to obtain a drug reaction prediction model that meets the performance indicators after training.

[0013] S4: drug reaction prediction

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

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

[0016] S11: The smiles sequence of the drug is encoded to obtain 6 drug molecule fingerprint matrices. A drug molecule graph is constructed based on the chemical structure of the smiles sequence itself, and 6 drug and biological entity related information matrices are introduced. Combined with 4 cell line matrices, a total of 16 data matrices are obtained, of which 6 drug molecule fingerprint matrices and 6 drug and biological entity related information matrices are drug matrices, a total of 12, and the drug and biological entity related information matrix is a drug entity matrix.

[0017] S12: The reaction data between the cell lines and the drugs with missing values are removed to obtain the final data set.

[0018] S13: For the above 16 data matrices, the Chebyshev distance is used for processing to obtain the corresponding 16 similarity matrices.

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

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

[0021]

[0022] wherein, V x and Vy respectively represent the xth row and yth row of the matrix, respectively represent the zth element of V x and Vy.

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

[0024] In the first inner product unit:

[0025] The inner product operation is performed on the six drug entity embedding vectors d i and the six drug molecule fingerprint embedding vectors d j

[0026]

[0027] wherein, represents the inner product operation, i and j represent the numbers of the drug entity matrix embedding vectors and the drug molecule fingerprint embedding vectors in 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 vectors and the drug molecule fingerprint embedding vectors;

[0029] The multi-layer perception MLP inner1 is used to perform feature extraction operation on the above obtained inner product output to obtain a first inner product feature vector, wherein the multi-layer perception MLP inner1 includes four fully connected layers, and the neuron numbers of the four fully connected layers are respectively 1024, 1024, 512 and 128.​

[0030] In the first outer product unit:

[0031] The outer product operation is performed on the 6 drug entity embedding vectors d i and the 6 drug molecule fingerprint embedding vectors d j

[0032]

[0033] wherein, 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 embedding vectors and the drug molecule fingerprint embedding vectors;

[0035] A two-dimensional convolutional neural network CNN1 is used to perform feature extraction on the above obtained outer product output, and a first outer product feature vector is obtained, wherein the two-dimensional convolutional neural network CNN1 includes two sets 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] wherein, the residual block includes a direct mapping part γ(g i ) and a residual part δ(g i ,W i ′), and a triple attention mechanism Triplet Attention is added at the end of the direct mapping part and the residual part.

[0038] Further, in the step S2, the second information interaction fusion branch is used to realize the information interaction fusion between the cell lines and the drugs, including a second inner product unit and a second outer product unit.

[0039] In the second inner product unit:

[0040] The inner product operation is performed on the 12 drug embedding vectors d′ x and the 4 cell line embedding vectors c′ y

[0041]

[0042] wherein, ⊙ represents the inner product operation, and x and y represent the numbers of the drug embedding vectors and the cell line embedding vectors when the inner product operation is performed.

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

[0044] using a multi-layer perception MLP inner2 The inner product output obtained above is subjected to feature extraction operation to obtain a second inner product feature vector, wherein the multi-layer perception MLP inner2 includes 4 fully connected layers, and the number of neurons of the 4 fully connected layers is 1024, 1024, 512 and 128 respectively;

[0045] In the second outer product unit:

[0046] The 12 drug embedding vectors d' x and the 4 cell line embedding vectors c' y are subjected to outer product operation:

[0047]

[0048] wherein, represents outer product operation;

[0049] A set of feature interaction maps with a channel number of 48 are obtained by outer product operation to represent the outer product output of the drug embedding vectors and the cell line embedding vectors;

[0050] The outer product output obtained above is subjected to feature extraction operation using a two-dimensional convolutional neural network CNN2 to obtain a second outer product feature vector, wherein the two-dimensional convolutional neural network CNN2 includes two sets 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] wherein, the residual block includes a direct mapping part α(f i ) and a residual part β(f i ,W i ), and a triplet attention mechanism Triplet Attention is added at the end of the direct mapping part and the residual part.

[0053] Further, the drug embedding vectors include drug entity embedding vectors and drug molecule fingerprint embedding vectors, and the drug embedding vectors are obtained in the following manner: a similarity matrix corresponding to a drug matrix is processed according to a set batchsize to obtain preliminary embedding vectors, and the preliminary embedding vectors are multiplied by a corresponding drug feature transformation matrix to obtain the drug embedding vectors.

[0054] The cell line embedding vector is obtained in the following manner: the similarity matrix corresponding to the cell line matrix is processed according to a set batchsize, that is, a preliminary embedding vector is obtained, and the preliminary embedding vector is multiplied by the corresponding cell line feature transformation matrix, so that the cell line embedding vector is obtained.

[0055] Further, in the step S2, the drug reaction prediction network further comprises a fusion feature extraction branch, a drug entity matrix global feature extraction branch, a drug molecular 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 the 12 drug matrices to obtain a fusion matrix Fuse feature, and the multi-layer perception MLP fuse is used for feature extraction to obtain a fusion feature vector, wherein the multi-layer perception MLP fuse includes two fully connected layers, and the number of neurons of the two 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 , the multi-layer perception MLP drug is used for overall feature extraction to obtain a first drug global feature vector, wherein the multi-layer perception MLP drug includes two fully connected layers, and the number of neurons of the two fully connected layers is 128 and 128, respectively. total The drug entity total matrix d total is obtained by splicing the similarity matrices corresponding to the six drug entity matrices.

[0058] In the drug molecular fingerprint matrix global feature extraction branch, the similarity matrices corresponding to the six drug molecular fingerprint matrices obtained after encoding are spliced to obtain a spliced drug molecular fingerprint total matrix, and a three-stack one-dimensional convolution module is used for feature extraction 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 , the multi-layer perception MLP cell is used for overall feature extraction to obtain a cell line global feature vector, wherein the multi-layer perception MLP cell includes two fully connected layers, and the number of neurons of the two fully connected layers is 128 and 128, respectively. total The cell line total matrix C total is obtained by splicing the similarity matrices corresponding to the four cell line matrices.

[0060] Further, in the step S2, a three-layer structure graph isomorphism neural network is used to extract the drug molecule graph data, and the outputs of each layer are spliced, a skip connection and a global maximum pooling layer are added to the network for processing.

[0061] Further, in the step S3, a mean square error loss function is used as the loss function for network training.

[0062]

[0063] Wherein, 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 drug reaction prediction method based on the graph neural network and the convolutional neural network has the following advantages: the drug reaction prediction method based on the graph neural network and the convolutional neural network adopts the method of combining the convolutional neural network and the graph neural network, adds the drug molecule graph data, introduces the structural characteristics of the drug molecules, and makes the final regression prediction more accurate; the one-dimensional convolution module is used for feature extraction on the global characteristics of the drug molecule fingerprint derived from the sequence, and the neighborhood information is strengthened; through the designed information interaction and fusion branch, not only the information interaction between the drug characteristics and the cell line characteristics is realized, but also the interaction between the drug reaction information and the drug smiles sequence information is realized, and the obtained new feature vector improves the prediction performance of the model. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 is a flowchart of the drug reaction prediction method based on the graph neural network and the convolutional neural network in the embodiment of the application;

[0066] Figure 2 is a structural diagram of a down-sampling module in the embodiment of the application;

[0067] Figure 3 is a whole structure diagram of a drug reaction prediction model in the embodiment of the application;

[0068] Figure 4 is a structural diagram of a triple attention module in the embodiment of the application;

[0069] Figure 5 is a visual effect diagram of a prediction result in the embodiment of the application. DETAILED DESCRIPTION

[0070] The embodiment of the application will be described in detail below, and the embodiment is implemented on the premise of the technical scheme of the application, and detailed implementation modes and specific operation processes are given, but the protection scope of the application is not limited to the following embodiment.

[0071] As Figure 1As shown, the embodiment provides a technical scheme: a drug reaction prediction method based on a graph neural network and a 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 molecule fingerprint matrices, and construct a drug molecule graph based on the chemical structure of the smiles sequence itself, while introducing 6 drug and biological entity related information matrices, and combining 4 cell line matrices, to obtain a total of 16 data matrices, wherein the 6 drug molecule fingerprint matrices and the 6 drug and biological entity related information matrices (drug entity matrices) are drug matrices, a total of 12;

[0075] S12: remove the reaction data between the cell lines and the drugs with missing values to obtain the final data set;

[0076] S13: for the above 16 data matrices, use Chebyshev distance to process the above matrices to obtain a similarity matrix of the 16 data matrices:

[0077]

[0078] wherein V x and Vy represent the xth and yth rows of the matrix, respectively, and V x and Vy represent the zth element of V

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

[0080] Table 1 Drug reaction matrix and cell line matrix name

[0081]

[0082]

[0083] M combined : combined interaction between drugs; M disease : known association between drug and 5181 diseases; M mirna : known drug-microrna association; M adr : drug and 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 : explainable substructure partition fingerprint; M ergfp : extended reduced graph fingerprint; M ecfp : extended connection fingerprint; M psfp : PubChem substructure fingerprint; M dfp : daylight fingerprint; M cd : IC50 value of reaction between cell line and 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: extracting features of the drug and features of the cell line;

[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, and 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 maximize the retention of the structural information of the drug molecule. The drug entity embedding vector is named as {d1, d2, d3, d4, d5, d6}, and the drug molecule fingerprint embedding vector is named as {d7, d8, d9, d 10 , 11 , 12}. The present application designs an inner product unit and an outer product unit for the information interaction of the two. In the inner product unit, the present application performs an inner product operation on each drug entity embedding vector d i and the drug molecule fingerprint embedding vector d j , and the operation of the 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 when the inner product operation is performed, and here, the present application obtains 36 different interaction vectors to represent the inner product output between the two vectors. The present application uses a multi-layer perception MLP inner1 to extract the inner product output obtained above to obtain a first inner product feature vector (Inner embeddings1), and the MLP inner1 used by the present application is composed of 4 fully connected layers, and the numbers of neurons of the four fully connected layers are 1024, 1024, 512 and 128 respectively.

[0089] In the outer product unit, the present application performs an outer product operation on each drug entity embedding vector d i and the drug molecule fingerprint embedding vector d j , and the operation of the outer product operation is as follows:

[0090]

[0091] , wherein represents an outer product operation, where the application obtains a set of 36 channel feature interaction maps, then the application uses a convolutional neural network CNN1 to perform feature extraction operations on the set of feature maps to obtain a second inner product feature vector Outer embeddings1, the convolutional neural network CNN1 designed by the application is composed of two groups of residual blocks, and 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 is composed of a direct mapping part γ(g i ) and a residual part δ(g i ,W i ′), and the same as the application, the triplet attention mechanism (Triplet Attention module) is added at the end of the direct mapping part and the residual part, to improve the expression ability of the output feature.

[0094] S22: The same application for information interaction between cell lines and drugs also designs inner product units and outer product units, in the inner product unit, the application performs inner product operations on 12 drug embedding vectors d′ x and 4 cell line embedding vectors c′ y , the operation of the inner product operation is as follows:

[0095]

[0096] Where, ⊙ represents an inner product operation, x and y represent the numbers of drug embedding vectors and cell line embedding vectors when performing inner product operations, here the application obtains 48 different interaction vectors to represent the inner product output of drug embedding vectors and cell line embedding vectors. The application uses a multi-layer perceptron MLP inner2 to extract the inner product output obtained above to obtain a second inner product feature vector (Inner embeddings2), the multi-layer perceptron MLP inner2 used by the application is composed of 4 fully connected layers, and the number of neurons of the four fully connected layers is 1024, 1024, 512 and 128 respectively.

[0097] In the outer product unit, the application performs outer product operations on each drug embedding vector d′ x and cell line embedding vector c′ y , the operation of the outer product operation is as follows:

[0098]

[0099] wherein, represents an outer product operation, here, a set of 48-channel feature maps is obtained, then the application uses a convolutional neural network CNN2 to perform feature extraction on the set of feature maps to obtain a second outer product feature vector (Outerembeddings2), the convolutional neural network CNN2 designed by the application is composed of two groups of residual blocks, and the feature extraction process is as follows:

[0100] f i+1 = Triplet Attention (a (f i ))+ Triplet Attention (b (f i ,W i )

[0101] The residual block is composed of a direct mapping part a (f i ) and a residual part b (f i ,W i ), and the application adds a triple attention mechanism at the end of the direct mapping part and the residual part, which can effectively capture the interaction information between each dimension in the C, H and W three dimensions, so that the output features are more expressive, and the attention mechanism consumes less computing resources, so that the model prediction accuracy is improved to the maximum while saving computing resources.

[0102] The triple attention mechanism is as shown in Figure 4 The first branch: channel attention calculation branch, the input feature is subjected to Z-Pool, then 7x7 convolution, and finally Sigmoid activation function to generate spatial attention weight. The second branch: channel C and spatial W dimension interaction capture branch, the input feature is first subjected to dimension operation of rearranging tensor (permuter) to become HxCxW dimension feature, then subjected to Z-Pool in H dimension, then subjected to 7x7 convolution, and finally Sigmoid activation function is used to generate spatial attention weight; finally, the dimension operation of rearranging tensor (permuter) is needed to become CxHxW dimension feature, which is convenient for element-wise addition. The third branch: channel C and spatial H dimension interaction capture branch, the input feature is first subjected to dimension operation of rearranging tensor (permuter) to become WxHxC dimension feature, then subjected to Z-Pool in W dimension, then subjected to 7x7 convolution, and finally Sigmoid activation function is used to generate spatial attention weight; finally, the dimension operation of rearranging tensor (permuter) is needed to become CxHxW dimension feature, which is convenient for element-wise addition. Finally, the output features of the three branches are added and averaged (Avg).

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

[0104] From Table 2, it can be seen that the information fusion method of the application is more accurate for the final prediction compared with 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 the steps S21 and S22 in the embodiment, the following steps are performed:

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

[0109] c i

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

[0111] On the other hand, for the cell line total matrix C total obtained by splicing (when splicing, the row is invariant and the column is increased) the similarity matrices corresponding to the four cell line matrices, the overall features of the cell lines are extracted using an MLP. There are two fully connected layers in the MLP network at this stage, and the number of neurons in the fully connected layers is 128, 128 respectively.

[0112] It should be noted that the cell line embedding vector can be obtained as follows: the similarity matrix corresponding to the cell line matrix is processed according to the set batchsize, i.e. the preliminary embedding vector is obtained, and the preliminary embedding vector is multiplied by the corresponding cell line feature transformation matrix, so that the cell line embedding vector is obtained.

[0113] Similarly, 12 drug feature transformation matrices, i.e. {A1, A2, …, A​12}∈R n×s where s is set to 128, then the preliminary embedding vector d i of the i-th drug matrix can be mapped to the same vector space by 12 drug feature transformation matrices. i (i.e. d' x ) can be calculated as:

[0114] d' i = d i A i

[0115] After the above mapping process, the present application can map the dimension of the preliminary embedding vector d i of the drug matrix to the same vector space by 12 drug feature transformation matrices.

[0116] On the other hand, for the similarity matrix corresponding to the 6 drug entity matrices, the drug entity total matrix d total obtained by splicing (when splicing, the row is invariant and the column is increased) is used to extract the overall features of the drug entity information by using the MLP network. In the MLP network at this stage, there are two fully connected layers, and the number of neurons in the fully connected layers is 128, 128 respectively.

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

[0118] In the step S2 in the embodiment, 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.

[0119] In the fusion feature extraction branch, the present application uses the SNF algorithm to fuse 12 drug matrices (original drug matrices, without Chelsher distance processing), and adds the similarity information between the drug matrices in the regression prediction model of the present application. 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] where D1~D 12for the corresponding 12 original drug matrixes;

[0122] Here i∈{1…12}, finally the invention will get a fusion matrix Fuse feature∈R 170×170 For this fusion matrix, the invention chooses to use MLP fuse Feature extraction is performed to obtain a fusion feature vector (Fuseembeddings), and the MLP used is fuse It is composed of two fully connected layers, and the number of neurons of the two fully connected layers is 128, 128 respectively. This process can be described as:

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

[0124] Where BN represents the Batchnorm layer, ReLU represents the activation function, and Linear represents the fully connected layer

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

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

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

[0128] In the drug molecule fingerprint matrix global feature extraction branch, the similarity matrix corresponding to the six drug molecule fingerprint matrices obtained after encoding is spliced (the row is unchanged and the column is increased when splicing), and the spliced drug molecule fingerprint total matrix is obtained. Feature extraction is performed using a one-dimensional convolution module with three stacks 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 to obtain a unique representation containing the drug smiles sequence.

[0129] In the cell line matrix global feature extraction branch, for the cell line total matrix C total ∈R n×2698 , the overall features of the cell line are extracted using an MLP cell to obtain cell line global feature vectors (Cells embeddings). In the MLP network at this stage, there are two fully connected layers, and the number of neurons in the fully connected layers is 128 and 128, respectively. The c′ total after MLP mapping can be calculated as:

[0130] c′ total =Linear(ReLU(BN(Linear(c total )))

[0131] where BN represents a Batchnorm layer, ReLU represents an activation function, and Linear represents a fully connected layer.

[0132] In this embodiment, in the step S2, in the two-dimensional convolutional neural network, two layers of convolutional neural networks with residual structure (residual blocks) 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 down-sampling module for processing.

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

[0134] In this embodiment, after obtaining each feature vector, the splicing module is used for splicing, and then the regressor is input for prediction work.

[0135] The drug response prediction model uses three evaluation indexes to evaluate the effect, which are RMSE: root mean square error; MAE: maximum average error; and R: determination coefficient. As can be seen from Table 3, after adding the drug molecule graph data, the prediction ability of the model is improved.

[0136] Table 3 Comparison of drug molecule graph data 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, as can be seen from Table 4, when the graph neural network is replaced by a graph convolutional neural network (GCN) and a graph attention network (GAT) and a graph transformer, the root mean square error and the maximum average error are increased, which reflects the effectiveness of the graph isomorphism 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, the validation set and the test set are divided in the ratio of 8:1:1, the model is trained and predicted, and the mean square error loss function is selected as the final loss function.

[0143]

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

[0145] S4: Drug response prediction

[0146] In the step S4, the drug response prediction is regarded as a regression task, the value predicted by the model is the IC50 response value of the drug, and the difference between the true value is calculated to obtain a more accurate prediction result. The model design diagram is as shown in Figure 3 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, but if the loss function is continuously increased for 10 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 result. The visualization effect diagram of the prediction result is as shown in Figure 5 .

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

[0148] Table 5 Performance of different methods

[0149]

[0150]

[0151] In summary, the drug reaction prediction method based on the graph neural network and the convolutional neural network of the above-mentioned embodiments first carries out necessary preprocessing operations on the published drug smiles sequence and 12 kinds of data matrices of the drug, 3 kinds of omics matrices of the cell line and the IC50 value matrix of the drug and cell line pair;For drug characteristics, a graph isomorphism neural network is used to extract the characteristics of the drug molecule graph, a one-dimensional convolution module is used to extract the characteristics of the drug molecule fingerprint from the drug smiles sequence itself, and a multi-layer perceptron is used to extract the characteristics of the drug and biological entity related information matrix;For cell line characteristics, a multi-layer perceptron with strong mapping ability is selected for feature extraction;For the interaction and fusion of the two kinds of characteristics, the inner product and outer product operations are carried out on the drug characteristics and the cell line characteristics, different structure characteristics are obtained, and finally sent to the multi-layer perceptron and the two-dimensional convolutional neural network for feature extraction, and the application considers the feature interaction of the drug and biological entity related data and the drug itself smiles sequence data, and the same outer product and inner product operations are performed on the two;Finally, the obtained various characteristic vectors are input into the last regressor for regression prediction;A variety of experiments show that the method of the application is effective and the modules used are reasonable, the proposed model has high prediction accuracy, and is expected to be deployed on equipment in the future to be put into production and application.

[0152] Although the embodiments of the application have been shown and described above, it should be understood that the above-mentioned embodiments are exemplary and cannot be understood as limiting the application, and those skilled in the art can make changes, modifications, replacements and variations to the above-mentioned embodiments within the scope of the application.

Claims

1. A drug response prediction method based on a graph neural network and a convolutional neural network, characterized in that, The method comprises the following steps: S1: data preprocessing Obtain the baseline data set of the cell line required for the smiles sequence of the drug, and preprocess the smiles sequence of the drug and the missing values according to the task requirements; S2: network construction A drug reaction prediction network is constructed, which comprises multiple input branches, a splicing module and a regressor. Each input branch obtains a corresponding feature vector, which is input into the regressor through the splicing module for prediction. The input branch comprises two information interaction fusion branches and a drug molecular graph extraction branch. The information interaction fusion branches perform interaction and fusion of drug information itself and interaction and fusion of drug information and cell line information through a multi-layer perception and a two-dimensional convolutional neural network. The drug molecular graph extraction branch extracts a drug molecular graph feature vector through a graph isomorphism neural network. S3: network training The data set obtained after preprocessing is divided according to a set proportion, and the drug reaction prediction network is trained using the training set to obtain a drug reaction prediction model that meets the performance indicators after training; S4: drug reaction prediction The cell line data and unknown drug data in the test set are input into the drug reaction prediction model for prediction to obtain a prediction result. In the step S2, the first information interaction fusion branch is used to realize information interaction and fusion of drug information itself, comprising a first inner product unit and a first outer product unit. In the first inner product unit: For 6 drug entity embedding vectors d i With 6 drug molecule fingerprint embedding vectors d j Performing inner product operation: Wherein, represents the inner product operation, i and j represent the numbering of the drug entity matrix embedding vector and the drug molecular fingerprint embedding vector when the inner product operation is performed; 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 molecular fingerprint embedding vector. Using a multi-layer perceptron, MLP inner1 A feature extraction operation is performed on the inner product output obtained above to obtain a first inner product feature vector, wherein the multi-layer perceptron, MLP inner1 includes 4 fully connected layers, and the number of neurons of the 4 fully connected layers are 1024, 1024, 512, and 128, respectively. In the first outer product unit: For 6 drug entity embedding vectors d i With 6 drug molecule fingerprint embedding vectors d j Performing an outer product operation: wherein denotes the outer product operation; After the outer product operation, a group 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 molecular fingerprint embedding vector; A two-dimensional convolutional neural network CNN1 is used to perform feature extraction on the above obtained outer product output to obtain a first outer product feature vector. The two-dimensional convolutional neural network CNN1 comprises 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 ′)) Wherein, the residual block includes a direct mapping part γ(g i ) and a residual part δ(g i ,W′ i ), and a triplet attention mechanism is added at the end of the direct mapping part and the residual part. In the step S2, the second information interaction fusion branch is used to realize information interaction and fusion between the cell line and the drug, comprising a second inner product unit and a second outer product unit. In the second inner product unit: For 12 drug embedding vectors d' x With 4 cell line embedding vectors c' y Performing inner product operations: Wherein, represents the inner product operation, x and y represent the numbering of the drug embedding vector and the cell line embedding vector when the inner product operation is performed; 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 a multi-layer perceptron, MLP inner2 A feature extraction operation is performed on the obtained inner product output to obtain a second inner product feature vector, wherein the multi-layer perceptron, MLP inner2 includes 4 fully connected layers, and the number of neurons of 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 With 4 cell line embedding vectors c' y Performing an outer product: wherein denotes the outer product operation; After the outer product operation, a group 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; A two-dimensional convolutional neural network CNN2 is used to perform feature extraction on the above obtained outer product output to obtain a second outer product feature vector. The two-dimensional convolutional neural network CNN2 comprises two groups of residual blocks, and the corresponding feature extraction process is as follows: f i+1 = Triplet Attention(a(f i )) + Triplet Attention(b(f i ), W i )) Wherein, the residual block includes a direct mapping part a(f i ) and a residual part β(f i ,W i ), and a triplet attention mechanism is added at the end of the direct mapping part and the residual part.

2. The drug response prediction method based on a graph neural network and a convolutional neural network according to claim 1, characterized in that, In the 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 graph based on the chemical structure of the smiles sequence itself, while introducing 6 drug and biological entity related information matrices, and combining 4 cell line matrices to obtain a total of 16 data matrices, wherein the 6 drug molecular fingerprint matrices and the 6 drug and biological entity related information matrices are drug matrices, a total of 12, and the drug and biological entity related information matrix is also a drug entity matrix; S12: remove the data with missing values between the cell lines and the drugs to obtain the final data set; S13: for the above 16 data matrices, use Chebyshev distance to process to obtain the corresponding 16 similarity matrices.

3. The drug response prediction method based on a graph neural network and a convolutional neural network according to claim 2, characterized in that, In the step S11, the 6 drug molecular fingerprint matrices are respectively an extended connection fingerprint matrix, a Pubchem substructure fingerprint matrix, a daylight fingerprint matrix, a RDKit 2D normalized fingerprint matrix, an interpretable substructure partition fingerprint matrix and an extended simple graph fingerprint matrix; the 6 drug entity matrices are respectively a joint interaction matrix between drugs, a drug and target interaction matrix, a drug and known disease interaction matrix, a drug and miRNA interaction matrix, a drug and cancer interaction matrix, and an IC50 value matrix of the drug and cell line pair; and the 4 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 the cell line and drug pair.

4. The drug response prediction method based on a graph neural network and a convolutional neural network according to claim 2, characterized in that, In the step S13, the calculation formula of the similarity matrix is as follows: where V x and Vydenote the xthand ythrow of the matrix, respectively, denote the zthelement of V x and Vy, respectively.

5. The drug response prediction method based on a graph neural network and a convolutional neural network according to claim 1, characterized in that, The drug embedding vector includes a drug entity embedding vector and a drug molecular fingerprint embedding vector, and the drug embedding vector is obtained in the following manner: the similarity matrix corresponding to the drug matrix is processed according to a set batchsize to obtain a preliminary embedding vector, and the preliminary embedding vector is multiplied by a corresponding drug feature transformation matrix to obtain the drug embedding vector; The cell line embedding vector is obtained in the following manner: the similarity matrix corresponding to the cell line matrix is processed according to a set batchsize to obtain a preliminary embedding vector, and the preliminary embedding vector is multiplied by a corresponding cell line feature transformation matrix to obtain the cell line embedding vector.

6. The drug response prediction method based on a graph neural network and a convolutional neural network according to claim 5, characterized in that, 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 molecular 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 for fusion of 12 drug matrices to obtain a fusion matrix Fuse feature, and a multi-layer perception MLP fuse Feature extraction is performed to obtain a fusion feature vector, wherein the multi-layer perception MLP fuse includes 2 fully connected layers, and the number of neurons of 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 , the extraction of overall features is performed using a multi-layer perception MLP drug to obtain a first drug global feature vector, wherein the multi-layer perception MLP drug includes 2 fully connected layers, and the number of neurons of the 2 fully connected layers is 128 and 128 respectively, and the drug entity total matrix d total is obtained by splicing 6 similarity matrices corresponding to the drug entity matrices; In the drug molecular fingerprint matrix global feature extraction branch, the similarity matrices corresponding to the 6 drug molecular fingerprint matrices obtained after encoding are spliced to obtain a spliced drug molecular fingerprint total matrix, and a three-layer stacked one-dimensional convolution module is used for feature extraction 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 , the extraction of overall features is performed using a multi-layer perceptron MLP cell , to obtain a cell line global feature vector, wherein the multi-layer perceptron MLP cell includes 2 fully connected layers, and the number of neurons of the 2 fully connected layers is 128 and 128 respectively, and the cell line total matrix C total is obtained by splicing 4 similarity matrices corresponding to the cell line matrices.

7. The drug response prediction method based on a graph neural network and a convolutional neural network according to claim 1, characterized in that, In the step S2, a graph isomorphism neural network with a three-layer structure is used to extract drug molecular graph data, and the outputs of each layer are spliced, and a skip connection and a global maximum pooling layer are added to the network for processing.

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

Images