Method for analyzing and predicting drug-target interaction by using deep learning

Abstract

The invention discloses a method for analyzing and predicting drug-target interaction by using deep learning, which comprises the following steps: encoding a drug and a target protein, and extracting the characteristics of the drug and the target protein for interaction; inputting the interaction characteristics of the drug and the target protein into a multi-head cross attention mechanism, and extracting the interaction characteristics of the drug and the target protein; obtaining independent characteristics of the drug and the target protein; fusing and re-classifying the independent characteristics of the drug, the independent characteristics of the target protein and the interaction characteristics of the drug and the target protein, and predicting drug-target interaction; the invention provides a network based on feature extraction between global molecules and local feature molecules of a drug-target, and the DTIs of the drug-target is predicted; according to the method, the interaction features of the drug target are extracted by adopting a cross attention mechanism, and the independent features of the drug target are extracted by using GAT and CNN, so that rich drug-target information is obtained, and the prediction performance is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of analyzing and predicting drug-target interactions, and in particular to a method for analyzing and predicting drug-target interactions using deep learning. Background Art

[0002] Drug discovery is a time-consuming and economically costly task, and there is no guarantee that a drug can be brought to market. In order to reduce the risk of drug discovery, some drug positioning strategies have been proposed to find new uses for existing drugs.

[0003] Several drugs, such as sildenafil, thalidomide, zidovudine, minoxidil, and celecoxib, have been identified and brought to the market through drug repositioning approaches, and the use of reliable predictions of drug-target interactions (DTIs) can help facilitate drug repositioning and drug discovery.

[0004] At present, the data size of drugs and target proteins is growing rapidly. In a large number of data sets, potential DTIs can be predicted. The current methods mainly include similarity-based methods and network-based methods. These methods are mostly based on known information, and calculate the similarity of drug-drug, target-target or drug-target associations to predict potential DTIs; however, existing methods for calculating similarity may ignore important information.

[0005] In recent years, deep learning technology has been applied to many research fields. Applying deep learning technology in the field of DTIs will help promote drug repositioning and drug discovery. Summary of the invention

[0006] In view of the above-mentioned problems, the purpose of the present invention is to provide a method for predicting drug-target interactions using deep learning analysis, apply deep learning technology in the field of drug-target interactions (DTIs), speed up the screening of new drugs among approved drugs, and help shorten clinical trial time.

[0007] The embodiment of the present invention provides a method for predicting drug-target interaction using deep learning analysis, comprising:

[0008] S1. Encode drugs and target proteins, and extract the features of drugs and target proteins for interaction;

[0009] S2, input the interactive features of drugs and target proteins into the multi-head cross attention mechanism to extract the interactive features of drugs and target proteins;

[0010] S3. Obtain independent features of drugs and target proteins;

[0011] S4. Fusion and reclassification of drug-independent features, target protein-independent features, and drug-target protein interaction features to predict drug-target interactions.

[0012] Furthermore, the encoding of the drug and the target protein includes:

[0013] The target protein sequence and the drug SMILES sequence are regarded as documents, the amino acid bases of the target protein sequence and each character of the drug SMILES sequence are regarded as words, and the target protein sequence and the drug SMILES sequence are encoded using the Word2vec module;

[0014] For the target protein sequence, the k-mer context window is used to segment the target protein sequence into amino acid sequence fragments;

[0015] Each character in the drug SMILES sequence is segmented into a word to represent it.

[0016] Furthermore, the step of obtaining the characteristics of the drugs for interaction in S1 includes the following steps:

[0017] S11a, embed the word obtained after segmenting the SMILES sequence into the pre-trained dictionary Smi2Vec;

[0018] S12a, based on the dictionary Smi2Vec, obtain the features of drugs used for interaction, and the drug mutual feature matrix is ​​expressed as l d ×100, where l d is the length of the SMILES sequence and 100 is the dimension of the drug interaction features.

[0019] Furthermore, the step of obtaining the characteristics of the target protein for interaction in S1 includes the following steps:

[0020] S11b, embedding the amino acid sequence after segmentation of the target protein sequence into the pre-trained dictionary Prot2Vec;

[0021] S12b, according to the dictionary Prot2VecS1, the target protein interaction features are obtained, and the target protein mutual features are represented as l p ×100 where l p is the length of the amino acid sequence, and 100 is the characteristic dimension of the mutual characteristics of the target proteins.

[0022] Furthermore, the multi-head cross-attention module consists of a drug attention component and a protein attention component, both of which have the same weight;

[0023] The step of extracting the interactive features of the drug and the target protein is:

[0024] S31, passing drug feature map through linear layer based on drug attention component Calculate drug query vector Use another linear layer to calculate the target protein feature map Mapping to get drug key vector and drug value vector

[0025] The query vector, key vector, and value vector of a drug are represented as:

[0026]

[0027] in, are the matrix weights of the three linear layers, d head =d cnn / heads is the channel dimension and heads is the number of attention heads.

[0028] S32, based on the protein attention component, the target protein feature map is mapped into the linear layer to calculate the target protein query vector Generate target protein bond vectors using drug feature map mapping and target protein value vector

[0029] The query vector, key vector and value vector of the target protein are expressed as:

[0030]

[0031] is the weight matrix shared by the target protein attention module and the drug attention module.

[0032] S33. Use the Softmax function to calculate the drug attention matrix and the target protein attention matrix. The formula is:

[0033]

[0034] are the dimensions of the query vector and key vector;.

[0035] The drug attention matrix of each attention head Multiply the drug value vector matrix The participating drug feature map of each attention head is obtained; then the participating drug feature maps of all attention heads are concatenated along the channel dimension and sent to the linear layer to obtain the final participating drug feature map

[0036]

[0037] in, is the initial weight matrix;

[0038] Attention matrix for target proteins and target protein value matrix Perform the same operation to obtain the final target protein feature map

[0039]

[0040] Finally, the participating drug and target protein feature maps are combined with the original feature matrix to obtain the final drug and target protein interaction feature matrix:

[0041]

[0042] here is the drug interaction feature matrix and is the target protein interaction feature matrix.

[0043] Furthermore, the step of obtaining independent characteristics of drugs includes the following steps:

[0044] S21a, use RDKit to convert the SMILES sequence into the corresponding molecular graph. The drug molecular graph is represented as G(V,E), where V is the set of all atoms in a molecule and E is the set of covalent chemical bonds;

[0045] S22a, the drug molecule topological structure information is extracted through T-layer GAT, and convolution is performed at the T(X)th layer to calculate the attention weight of the i-th atom in the molecular graph and its neighbor node j∈N(i);

[0046] S23a, based on the attention weight, update the i-th atomic vector information;

[0047] S24a, after T times of convolution, the drug molecular polymerization formula is obtained as the drug independent feature represented by G drug .

[0048]

[0049] Among them, Z i It is the feature of the i-th vertex in the graph after T convolutions.

[0050] Furthermore, the step of obtaining independent characteristics of the target protein comprises the following steps:

[0051] S21b, use the corpus of all human target proteins in the UniProt database to train the Word2vec model; S22b, use the trained Word2vec model to obtain the target protein sequence feature matrix

[0052] S23b, use multiple convolution kernels of multi-scale one-dimensional convolution CNN to extract representative features of the target protein sequence feature matrix, the formula is expressed as:

[0053] C i =σ(P cnn *W i cnn +b i cnn ) (8)

[0054] Where W i cnn belong is a learnable parameter, m is the feature dimension of the hidden layer, n is the number of different convolution kernels, and σ() is a nonlinear activation function.

[0055] S24b, extract the most significant features through the maximum pooling layer c i =max{C i}, during the training process, BatchNorm is used to make the model converge efficiently. Obtain the independent feature G of the target protein prot ;

[0056] G prot =MLP(Concat(c1,...,c n )) (9)

[0057] Here MLP() is a multi-layer fully connected layer.

[0058] Furthermore, the method of fusing and reclassifying the drug-independent features, the target protein-independent features, and the interactive features of the drug and the target protein comprises the steps of:

[0059] The independent features and interactive features are input into the fully connected layer MLP for fusion to obtain the predicted fusion features. The formula is expressed as

[0060]

[0061] Among them, δ is the linear activation function, W out and b out is a learnable parameter, is the predicted label.

[0062] Furthermore, when training the fully connected layer MLP, the cross entropy loss function is the back propagation loss function, and the formula is:

[0063]

[0064] Among them, y i is the actual label, is the predicted label, Θ is the set of all parameters in the end-to-end model, N is the number of training samples, and λ is the L2 regularization coefficient.

[0065] Beneficial effects of the present invention:

[0066] The present invention proposes a network based on feature extraction between global molecules and local feature molecules of drug-target to predict drug-target DTIs. The cross-attention mechanism is used to extract the interactive features of drug targets, and GAT and CNN are used to extract independent features of drug targets, so as to obtain rich information about drug-target and improve the prediction performance.

[0067] The entire model of the present invention is lightweight and has fewer parameters than other models. The extracted interactions are bidirectional, extracting not only the interaction between the target protein and the drug, but also the interaction between the drug and the target protein, which is beneficial to improving the characterization capability of the features. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a schematic diagram of the process of the present invention;

[0069] Figure 2 It is a schematic diagram of the principle flow of the method of the present invention;

[0070] Figure 3 It is a schematic diagram of the module structure of the method of the present invention;

[0071] Figure 4 A schematic diagram of the interaction characteristics between the drug and the target protein is obtained for the present invention;

[0072] Figure 5 It is a schematic diagram of the structure of the electronic device of the present invention. DETAILED DESCRIPTION

[0073] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar symbols throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0074] In the current exploration of drug-target interaction DTIs, there is a lack of practical and effective methods to apply deep learning technology in the field of DTIs.

[0075] In view of the above problems, the present invention provides a method for predicting drug-target interaction using deep learning analysis. Figure 1-2 A schematic flow chart of a method for predicting drug-target interactions by analyzing intramolecular properties through deep learning provided by an embodiment of the present invention, the method comprising the steps of:

[0076] S1. Encode drugs and target proteins, and extract the features of drugs and target proteins for interaction;

[0077] like Figure 3 , 4 As shown, the model of the present invention mainly includes five modules: a data preprocessing module (Word2vec), a graph attention module (GAT, GraphAttention), a convolutional module (CNN, ConvolutionalNeuralNetwork), a multi-head cross attention module (MCA, Multhead-CrossAttention) and a (Prediction) prediction module.

[0078] Among them, the GAT module and the CNN module are mainly used to extract the independent features of drugs and target proteins separately; the data preprocessing module (Word2vec) is mainly used to construct another representation of drugs and target proteins; the multi-head cross attention module (MCA) mainly interacts with the mutual features of drugs and the mutual features of target proteins to extract the interactive features of the interaction between drugs and target proteins. The prediction module inputs the extracted independent features and interactive features into the prediction module to obtain the prediction results.

[0079] First, obtain the mutual characteristics between the drug and the target protein.

[0080] Drugs and target proteins can be encoded using the Word2vec module. Word2vec is an unsupervised method for creating word vectors. The Word2vec neural network model is shallow and dual-layered and is used to reconstruct language word text after training.

[0081] When using the Word2vec module, the target protein sequence and the drug sequence are regarded as documents, and each character of the separated amino acid base and the drug is regarded as a word. Then, the Word2vec module is used to encode the target protein sequence and the drug SMILES sequence.

[0082] For the target protein sequence, the k-mer context window is first used to segment the target protein sequence into amino acid sequence fragments. For a target protein sequence, it can be segmented into 3 amino acid sequences.

[0083] For drugs, each character in the SMILES sequence is segmented into a word to represent;

[0084] The features of drug interactions are obtained by segmenting the drug SMILES sequence and embedding the words into the pre-trained dictionary Smi2Vec. The features of drug interactions are obtained based on the dictionary Smi2Vec. The drug mutual feature matrix is ​​represented as l d ×100, where ld is the length of the SMILES sequence and 100 is the dimension of the drug interaction feature.

[0085] Obtaining the features of target proteins for interaction includes: embedding the amino acid sequence of the target protein sequence into the pre-trained dictionary Prot2Vec, and obtaining the features of target proteins for interaction based on the dictionary Prot2VecS1. The mutual features of target proteins are represented as l p ×100 where l p is the length of the amino acid sequence, and 100 is the characteristic dimension of the mutual characteristics of the target proteins.

[0086] S2. Input the interactive features of drugs and target proteins into the multi-head cross attention mechanism to extract the interactive features of drugs and target proteins.

[0087] The interaction features between drugs and target proteins are obtained based on the multi-head cross-attention module, which is mainly composed of a drug attention component and a protein attention component, both of which have the same weight.

[0088] First, the query vector, key vector, and value vector of the drug are extracted, and the drug feature map is passed through the linear layer based on the drug attention component. Calculate drug query vector Use two more linear layers to calculate the target protein feature map Mapping to get drug key vector and drug value vector

[0089] The query vector, key vector, and value vector of a drug are represented as:

[0090]

[0091] in, are the matrix weights of the three linear layers, d head =d cnn / heads is the channel dimension and heads is the number of attention heads.

[0092] Then the query vector, key vector, and value vector of the target protein are obtained: The protein attention component follows a similar process as the drug attention component, and the two attention components share these three linear layers.

[0093] Based on the protein attention component, the protein attention component maps the target protein feature map into the linear layer to calculate the target protein query vector Generate target protein bond vectors using drug feature map mapping and target protein value vector

[0094] The query vector, key vector and value vector of the target protein are expressed as:

[0095]

[0096] in, is the weight matrix shared by the target protein attention module and the drug attention module.

[0097] Then use the Softmax function to calculate the attention matrix of drugs and target proteins. The formula is:

[0098]

[0099] in, are the dimensions of the query and key vectors.

[0100] The drug attention matrix of each attention head is multiplied by the drug value vector matrix of each attention head to obtain the participating drug feature map of each attention head. Then the participating drug feature maps of all attention heads are concatenated along the channel dimension and sent to the linear layer to obtain the final participating drug feature map.

[0101]

[0102] here is the initial weight matrix.

[0103] Attention matrix for target proteins and target protein value matrix Perform the same operation to obtain the final target protein feature map

[0104]

[0105] Finally, the participating drug feature maps and target protein feature maps are combined with various independent feature matrices to obtain the final drug and target protein feature maps:

[0106]

[0107] here is the drug interaction feature matrix and is the target protein interaction feature matrix.

[0108] S3. Obtain independent features of drugs and target proteins.

[0109] Obtaining independent characteristics of drugs includes: representing the drugs as molecular graphs, and the nodes of the molecular graphs include: the number of atoms, the degree of hybridization, the number of hydrogen atoms, whether it is an aromatic hydrocarbon, whether it is in a ring, etc.

[0110] Most datasets contain only simplified molecular input forms, so RDKit can be used to convert SMILES sequences into corresponding molecular graphs and extract atomic features.

[0111] The drug molecule graph can be represented as G(V,E), where V is the set of all atoms in a molecule and E is the set of covalent chemical bonds;

[0112] Then, the topological structure information of the drug molecule is extracted through T layers of GAT, where T is the total number of layers of GAT. During the T(X)th convolution, the attention weights of the i-th atom of the molecular graph and its neighbor nodes j∈N(i) are calculated. Then, based on the attention weights, the i-th atom vector is updated by aggregating neighbor information. Finally, after T convolutions, each atom contains the subgraph topological information of T radius. The aggregated representation of the drug molecule is represented as G drug ;

[0113]

[0114] Among them, Z i It is the feature of the i-th vertex in the graph after T convolutions.

[0115] Obtaining independent features of the target protein includes: In order to extract the independent features of the target protein, a multi-scale one-dimensional convolutional CNN can be used to obtain the features of the target protein; the advantage of the multi-scale one-dimensional convolutional CNN is that it can not only extract spatially related information, but also extract common features in images or data sequences.

[0116] Different from Prot2Vec, the Word2vec model is trained here using a corpus containing all human target proteins in the UniProt database to obtain a training vocabulary of target proteins with a dimension of 100, which is used to initialize the embedding of CNN.

[0117] Obtain target protein sequence through embedding The characteristic matrix of p is the dimension of the embedding vector.

[0118] Multiple filters with multiple different convolution kernels are used to extract representative features of the target protein sequence, and the sliding window width of the i-th convolution kernel is Ki. The formula is expressed as:

[0119] C i =σ(P cnn *W i cnn +b i cnn ) (8)

[0120] Where Wi cnn belong is a learnable parameter, m is the feature dimension of the hidden layer, and n is the number of different convolution kernels. σ() is a nonlinear activation function. Then the most significant feature c is extracted through the maximum pooling layer. i =max{C i}, during the training process, BatchNorm is used to make the model converge efficiently. Finally, the independent feature G of the target can be obtained prot ;

[0121] G prot =MLP(Concat(c1,...,c n )) (9)

[0122] Here MLP() is a multi-layer fully connected layer.

[0123] S4. Fusion and reclassification of drug-independent features, target protein-independent features, and drug-target protein interaction features to predict drug-target interactions.

[0124] Drug-target prediction is modeled as a binary task. All independent and interactive features learned above (6)(7)(9) are concatenated and fused in the downstream fully connected layers.

[0125]

[0126] δ is a linear activation function, W out and b out is a learnable parameter, is the predicted label.

[0127] For the drug-target pairs in the training dataset and their true label sets, the cross entropy loss function is defined as the back-propagation loss function.

[0128]

[0129] Among them, y i is the actual label, is the predicted label, Θ is the set of all parameters in the end-to-end model (e.g., the combined network of CNN, GAT, Multhead-CrossAttention, and classifier), N is the number of training samples, and λ is the L2 regularization coefficient.

[0130] The present invention can also use the AdamW optimizer to optimize the training process, including alternating training between a single drug and a drug combination, promoting the model to improve drug feature learning, and allowing the drug combination to be evaluated in a validation set after training.

[0131] The performance of the drug combination synergy screening prediction method of the present invention can be evaluated using the following formula:

[0132] Recall (Rec) = TP / (TP+FN);

[0133] Precision(Pre)=TP / (TP+FP);

[0134] F1=2·(Rec·Pre) / (Rec+Pre);

[0135] Accuracy(Acc)=(TP+TN) / (TP+FN+FP+TN);

[0136] Where TP represents the number of true positive samples, TN represents the number of true negative samples, FP represents the number of false positive samples, and FN represents the number of false negative samples. AUC is synonymous with AUROC and represents the area under the ROC curve. AUPR is the area under the precision-recall (PR) curve.

[0137] Experimental results: The present invention compares the drug target prediction method with the five leading deep learning methods. In order to comprehensively evaluate the performance, all the comparison methods are run 5 times for cross-validation on all drug target data sets, and the average experimental results are obtained, as shown in Table 1:

[0138] Table 1: Comparison of advanced models on different datasets

[0139]

[0140]

[0141] Ablation experiment: Ablation experiments were performed from interactive extraction and independent extraction. In terms of independent ablation, the parts of interactive extraction of drug target information were removed to confirm their importance in improving performance. ours(independent) means that ours only uses independent features. ours(Interactive_A) means that ours only uses interactive features. ours(out_A) means that ours's interactive features use ordinary attention mechanisms, and the key values ​​of the drugs will not be input into the targets, and the key values ​​of the targets will not be input into the drugs. From the results in the table, the performance of ours is better than ours(out_A), which confirms that the use of the cross-attention structure has important significance for improving the performance of the model. Ours also performs better than ours(independent) and ours(Interactive_A) on the DrugBank dataset, which indicates that the combination of the interactive features of multi-head cross attention with the independent features of GAT and CNN can improve the performance of the proposed model. From the perspective of model design, ablation experiments show that combining independent features and interactive features is effective. In addition, the performance of ours is also better than ours(Interactive_A), indicating that the combination of GAT, CNN and cross attention structure improves the performance of ours, which may be due to the complementarity of the features extracted by different feature extractors, as shown in Table 2:

[0142] Table 2: Experiments of different modules of the model in drugbank

[0143] Methods AUC AUPR Precision Recall ours(independent) 0.843 0.842 0.735 0.807 ours(Interactive_a) 0.873 0.871 0.735 0.880 ours(out_A) 0.921 0.924 0.846 0.837 ours 0.937 0.946 0.870 0.866

[0144] Experiments on balanced and unbalanced datasets: In order to test the performance of the model on unbalanced datasets, an unbalanced benchmark dataset Davis was also constructed. Davis is a dataset containing kinase-related proteins and drug pairs. The Davis dataset includes 72 kinase inhibitors and interactions between 442 kinases, of which 68 drugs and 379 target proteins were selected, resulting in 7,320 positive examples and 18,452 negative examples. The ratio of positive samples to negative samples is approximately 1:2.5. As shown in Table 3, the model outperforms other methods. The experimental results show that the model is still reliable when the dataset training is unbalanced.

[0145] In order to better evaluate the generalization ability of the model, the real experimental environment is simulated by collecting training sets and test sets on time series. As shown in Table 4, the data approved by DrugBank release version 5.0.3 (before October 2016) is used as the training set. Then the newly approved data after October 2016 is used as an independent test set and filtered to ensure that the drugs and targets in the test set do not appear in the training set. As shown in Table 5, the experimental results obtained are significantly improved compared with other methods, and the model can handle the problem of independent test sets in a real data environment.

[0146] Table 3. Comparison results of different algorithms on imbalanced datasets

[0147]

[0148]

[0149] Table 4. Overview of training and test datasets

[0150] Datasets Target Drugs Positive Negative Trainingset 1606 1471 6312 6312 Testset 1326 1144 949 3675

[0151] Table 5. Experimental results on independent test set

[0152] Methods AUC AUPR Precision Recall DeepDTA 0.784 0.519 0.450 0.635 DeepConv-DTI 0.782 0.472 0.459 0.626 TransformerCPI 0.782 0.500 0.427 0.660 Moltrans 0.501 0.203 0.183 0.417 IIFDTI 0.797 0.527 0.436 0.679 GIFDTI 0.794 0.553 0.479 0.678 Ours 0.805 0.561 0.543 0.604

[0153] Case Study: The purpose of the case study is to evaluate the actual ability of the model in identifying unknown DTIs. In the case study, all known DTIs in the benchmark dataset were first used as positive samples to form the training dataset, which were collected from DrugBank5.0.3. For negative samples, the drugs and targets that were not found to interact in the positive samples were randomly paired. In addition, the number of negative samples and positive samples in the training dataset was the same. The newly approved data after DrugBank5.0.3 was used as an independent test set and filtered to remove the drugs and targets in the training set, and an equal number of negative samples were also constructed. The training set was divided into a training set and a validation set with a 1:1 ratio. The cutoff value was set to 0.5 to determine the predicted DTI. In other words, if the prediction score of a drug-target pair was greater than 0.5, it was predicted to have an interaction. In terms of prediction score, the top 20 pairs in the test set were selected for further validation and the latest version of DrugBank was used for validation. Following the same steps of the model, the top 20 drug-target pairs predicted by each comparison model were selected for further study in the case study. The top 20 pairs of drugs and targets with the highest prediction scores are shown in Table 6. It is worth noting that the top 20 pairs of DTIs can be verified by the latest version of the DrugBank database, which means that when the model was trained, the drugs and targets were not connected, and they could be predicted by the model as candidate DTIs and verified in the DrugBank database. Therefore, compared with other compared algorithms, the model has better performance in discovering unknown DTIs. Taking IIFDTI as an example, only 10 of the top 20 pairs passed the verification of the DrugBank dataset.

[0154] Table 6. Model predictions for the top 20 results

[0155]

[0156]

[0157] The present invention also provides an electronic device, Figure 5 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention, such as Figure 5 As shown, the electronic device may include: a processor, a communications interface, a memory, and a communications bus, wherein the processor, the communications interface, and the memory communicate with each other via the communications bus. The processor may call the logic instructions in the memory, for example, to execute the following method:

[0158] S1. Encode drugs and target proteins, and extract the features of drugs and target proteins for interaction;

[0159] S2, input the interactive features of drugs and target proteins into the multi-head cross attention mechanism to extract the interactive features of drugs and target proteins;

[0160] S3. Obtain independent features of drugs and target proteins;

[0161] S4. Fusion and reclassification of drug-independent features, target protein-independent features, and drug-target protein interaction features to predict drug-target interactions.

[0162] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0163] An embodiment of the present invention further provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method provided in each of the above embodiments is implemented, for example, including:

[0164] S1. Encode drugs and target proteins, and extract the features of drugs and target proteins for interaction;

[0165] S2, input the interactive features of drugs and target proteins into the multi-head cross attention mechanism to extract the interactive features of drugs and target proteins;

[0166] S3. Obtain independent features of drugs and target proteins;

[0167] S4. Fusion and reclassification of drug-independent features, target protein-independent features, and drug-target protein interaction features to predict drug-target interactions.

[0168] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0169] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting drug-target interactions using deep learning analysis, characterized in that: include: S1. Encode drugs and target proteins, and extract the features of drugs and target proteins for interaction; S2, input the interactive features of drugs and target proteins into the multi-head cross attention mechanism to extract the interactive features of drugs and target proteins; S3. Obtain independent features of drugs and target proteins; S4. Fusion and reclassification of drug-independent features, target protein-independent features, and drug-target protein interaction features to predict drug-target interactions.

2. The method for predicting drug-target interaction using deep learning analysis according to claim 1, characterized in that: The encoding of drugs and target proteins includes: The target protein sequence and the drug SMILES sequence are regarded as documents, the amino acid bases of the target protein sequence and each character of the drug SMILES sequence are regarded as words, and the target protein sequence and the drug SMILES sequence are encoded using the Word2vec module; For the target protein sequence, the k-mer context window is used to segment the target protein sequence into amino acid sequence fragments; Each character in the drug SMILES sequence is segmented into a word to represent it.

3. The method for predicting drug-target interaction using deep learning analysis according to claim 1, characterized in that: Acquiring the features of drugs for interaction in S1 includes the following steps: S11a, embed the word obtained after segmenting the SMILES sequence into the pre-trained dictionary Smi2Vec; S12a, based on the dictionary Smi2Vec, obtain the features of drugs used for interaction, and the drug mutual feature matrix is ​​expressed as l d ×100, where l d is the length of the SMILES sequence and 100 is the dimension of the drug interaction features.

4. The method for predicting drug-target interaction using deep learning analysis according to claim 1, characterized in that: Acquiring the characteristics of the target protein for interaction in S1 includes the following steps: S11b, embedding the amino acid sequence after segmentation of the target protein sequence into the pre-trained dictionary Prot2Vec; S12b, according to the dictionary Prot2VecS1, the target protein interaction features are obtained, and the target protein mutual features are represented as l p ×100 where l p is the length of the amino acid sequence, and 100 is the characteristic dimension of the mutual characteristics of the target proteins.

5. The method for predicting drug-target interaction using deep learning analysis according to claim 1, characterized in that: The multi-head cross-attention module consists of a drug attention component and a protein attention component, both of which have the same weight; The step of extracting the interactive features of the drug and the target protein is: S31, passing drug feature map through linear layer based on drug attention component Calculate drug query vector Use another linear layer to calculate the target protein feature map Mapping to get drug key vector and drug value vector The query vector, key vector, and value vector of a drug are represented as: in, are the matrix weights of the three linear layers, d head =d cnn / heads is the channel dimension and heads is the number of attention heads; S32, based on the protein attention component, the target protein feature map is mapped into the linear layer to calculate the target protein query vector Generate target protein bond vectors using drug feature map mapping and target protein value vector The query vector, key vector and value vector of the target protein are expressed as: is the weight matrix shared by the target protein attention module and the drug attention module; S33. Use the Softmax function to calculate the drug attention matrix and the target protein attention matrix. The formula is: is the dimension of query vector and key vector; The drug attention matrix of each attention head Multiply the drug value vector matrix The participating drug feature map of each attention head is obtained; then the participating drug feature maps of all attention heads are concatenated along the channel dimension and sent to the linear layer to obtain the final participating drug feature map in, is the initial weight matrix; Attention matrix for target proteins and target protein value matrix Perform the same operation to obtain the final target protein feature map Finally, the participating drug and target protein feature maps are combined with the original feature matrix to obtain the final drug and target protein interaction feature matrix: here is the drug interaction feature matrix and is the target protein interaction feature matrix.

6. The method of predicting drug-target interaction using deep learning analysis according to claim 1, characterized in that: The step of obtaining independent characteristics of drugs comprises the following steps: S21a, use RDKit to convert the SMILES sequence into the corresponding molecular graph. The drug molecular graph is represented as G(V,E), where V is the set of all atoms in a molecule and E is the set of covalent chemical bonds; S22a, the drug molecule topological structure information is extracted through T-layer GAT, and convolution is performed at the T(X)th layer to calculate the attention weight of the i-th atom in the molecular graph and its neighbor node j∈N(i); S23a, based on the attention weight, update the i-th atomic vector information; S24a, after T times of convolution, the drug molecular polymerization formula is obtained as the drug independent feature represented by G drug ; Among them, Z i It is the feature of the i-th vertex in the graph after T convolutions.

7. The method of predicting drug-target interaction using deep learning analysis according to claim 1, characterized in that: The method of obtaining independent characteristics of the target protein comprises the steps of: S21b, train the Word2vec model using the corpus of all human target proteins in the UniProt database; S22b. Use the trained Word2vec model to obtain the target protein sequence feature matrix S23b, use multiple convolution kernels of multi-scale one-dimensional convolution CNN to extract representative features of the target protein sequence feature matrix, the formula is expressed as: C i =σ(P cnn *W i cnn +b i cnn ) (8) Where W i cnn belong is a learnable parameter, m is the feature dimension of the hidden layer, n is the number of different convolution kernels; σ() is a nonlinear activation function; S24b, extract the most significant features through the maximum pooling layer c i =max{C i }, During the training process, BatchNorm is used to make the model converge efficiently; Obtain independent features of target protein G prot ; G prot =MLP(Concat(c1,...,c n )) (9) Here MLP() is a multi-layer fully connected layer.

8. The method of predicting drug-target interaction using deep learning analysis according to claim 1, characterized in that: The drug-independent features, target protein-independent features and drug-target protein interaction features are fused and reclassified; including step: The independent features and interactive features are input into the fully connected layer MLP for fusion to obtain the predicted fusion features. The formula is expressed as Among them, δ is the linear activation function, W out and b out is a learnable parameter, is the predicted label.

9. The method of predicting drug-target interaction using deep learning analysis according to claim 1, characterized in that: When training the fully connected layer MLP, the cross entropy loss function is the back propagation loss function, and the formula is: Among them, y i is the actual label, is the predicted label, Θ is the set of all parameters in the end-to-end model, N is the number of training samples, and λ is the L2 regularization coefficient.

Images