Drug target identification method based on deep learning framework

Abstract

The invention relates to the technical field of computer biological information, and discloses a drug target identification method based on a deep learning framework, and the method comprises the following steps: S1, obtaining a Chemogenetic-Model data set of drug-target interaction; s2, obtaining a structure file of the target, and obtaining a structure feature representation and an edge feature representation of the target through the structure file of the target; s3, obtaining sequence tag coding feature representation, graph structure feature representation and physical and chemical feature representation of the drug and the target; s4, the sequence tag coding feature representation of the medicine is sent into a three-layer multi-scale convolutional neural network, the step length of each layer is 1, the filling is 0, and the sizes of convolution kernels are 3 * 3, 5 * 5 and 7 * 7 respectively. The reference data set used by the method is only one third of the reference data set used by a comparative advanced method, but on the same target identification test data set, the MM-IDTarget shows the performance close to or superior to that of the comparative advanced method on all Top-K evaluation indexes.

Description

Technical Field

[0001] The present invention relates to the field of computer bioinformatics, and in particular to a drug target identification method based on a deep learning framework. Background Art

[0002] Accurately identifying drug targets plays a key role in research fields such as drug design, drug repositioning, and treatment of diseases with unclear pathogenic mechanisms. Computer-assisted drug target identification methods have developed over decades, from structure- and ligand-based (drug) methods to current data-driven methods based on machine learning. Data-driven methods based on machine learning can identify targets with biological activity for a given drug from a large number of drug-target pairs at a relatively fast speed and with high accuracy. Data-driven methods based on machine learning can be divided into two categories for identifying drug targets: methods based on traditional machine learning and methods based on deep learning.

[0003] In the prediction of drug-target binding affinity based on deep learning, drug-target binding affinity prediction is generally regarded as a regression task to predict the strength of the interaction between the drug and the target. Classic methods include sequence methods such as DeepDTA and DeepDTAF based on drug and target sequences, combination methods such as GraphDTA, DataDTA and S2DTA based on drug and target sequences and structures, and structural methods such as GTAMP-DTA, MSGNN-DTA and PLANET based on drug and target structures. The study of target prediction probability ranking based on deep learning is to further sort the identified possible targets into Top-K based on drug-target interaction and binding affinity, for example, CODD-Pred method and GuiltyTargets-COVID-19 method.

[0004] Although most of the deep learning-based methods introduce target features, and compared with traditional machine learning methods, the performance indicators of the models, such as accuracy, root mean squared error (RMSE), and Pearson Correlation Coefficient (PCC), have also been greatly improved, only a small number of results use graph neural networks to extract structural features that affect drug target recognition from drugs and targets. Graph neural networks are still in the initial exploration stage in drug target recognition research. In addition, for existing deep learning-based drug-target interaction or affinity recognition methods, only the classification or regression performance indicators of the model are considered, while the key evaluation indicator Top-K for evaluating the quality of model target recognition is rarely calculated. Further use of graph neural networks, a cutting-edge deep learning technology, to build a drug target recognition model that meets the user's expected Top-K threshold still faces great challenges. Therefore, we propose a drug target recognition method based on a deep learning framework. Summary of the invention

[0005] The purpose of the present invention is to provide a drug target identification method based on a deep learning framework, which solves the problems raised in the background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a drug target identification method based on a deep learning framework, comprising the following steps:

[0007] The specific steps are as follows:

[0008] S1: Get the drug-target interaction dataset. The Chemogenomic-Model dataset comes from the BindingDB database (Version 2020) and the ChEMBL27 database. The dataset contains 153,281 drug (compound or small molecule)-target (protein) pairs, which are related to 859 targets and 93,281 drug small molecules. After preprocessing the dataset, we obtained 47,247 drug-target pairs, including 23,481 positive samples and 23,766 negative samples. These drug-target pairs are used as the benchmark datasets for this study. The preprocessing steps of the dataset are as follows:

[0009] (1) Based on 859 targets, we divided the drug-target pairs into 859 sub-datasets;

[0010] (2) We clustered the compounds in each sub-dataset based on similarity, with the similarity threshold set to 0.3, thereby dividing each sub-dataset into multiple compound clusters;

[0011] (3) In each compound cluster, if the number of positive (negative) samples is 1, we will retain the sample; if the number of positive (negative) samples is greater than 1, we will extract 25% of the positive (negative) samples (round up the result) and retain them;

[0012] (4) Finally, in order to evaluate the performance of the model, we further divide the samples extracted from each cluster into a training set and a test set: if the number of positive (negative) samples in a cluster is less than or equal to 2, then all positive (negative) samples of the cluster are used as the training set; if the number of positive (negative) samples of the cluster is greater than 2 and less than or equal to 10, then 10% of the test set samples are randomly selected (rounded up), and the remaining samples are used as the training set; if the number of positive (negative) samples of the cluster is greater than 10, then 8% of the test set samples are randomly selected (rounded up), and the remaining positive (negative) samples are used as the training set.

[0013] S2: Obtain the target's structure file (PDB), and obtain the target's structural feature representation and edge feature representation through the PDB file.

[0014] S3: Obtain sequence tag encoding feature representation, graph structure feature representation and physicochemical feature representation of drugs and targets.

[0015] S4: The sequence label encoding feature representation of the drug (target) is sent to a three-layer multi-scale convolutional neural network. The step size of each layer is 1, the padding is 0, and the sizes of the convolution kernels are 3x3, 5x5 and 7x7 respectively. Each convolution layer is followed by a normalization operation to obtain the three-layer features of the drug (target) in the multi-scale convolutional neural network, concatenate them and generate the sequence label encoding features through linear transformation.

[0016] S5: The structural feature representation of the drug graph is sent to a three-layer GraphTransformer framework to extract the structural modal features of the drug. The number of heads of each layer of GraphTransformer is 8, the number of neurons in each hidden layer is 128, each hidden layer is followed by a normalization operation, and finally connected to a maximum pooling operation.

[0017] S6: The target graph structural feature representation is sent to the Residual EW-GCN graph convolutional network. The Residual EW-GCN architecture is derived from GPCNDTA. The Residual EW-GCN contains three GCN layers and a fully connected network. First, the target graph feature representation is input into the first layer of GCN and the fully connected layer respectively. Secondly, the output features of the first layer of GCN and the output features of the fully connected network are concatenated. Then, the concatenated features are passed through the second and third layers of GCN in turn to obtain features with residuals, and finally L2 normalization is performed on them.

[0018] S7: Cross-attention is used to fuse the sequence label encoding feature representation and graph structure feature representation of the drug (target), and then the fusion feature representation of the drug and the target is fused to obtain the drug-target cross-attention fusion feature representation.

[0019] S8: After the obtained drug-target cross-attention fusion feature representation is concatenated with the drug (target) physicochemical feature representation, it is sent to the fully connected layer as a comprehensive feature of drug-target interaction to predict drug-target interaction.

[0020] S9: Compare the true label with the predicted result, and update the network parameters to minimize the cross entropy loss function until the maximum number of iterations is reached. The training ends and the trained MM-IDTarget model is obtained.

[0021] S10: For the trained MM-IDTarget model, we use the recall based on Top-K to evaluate the performance of the model in drug target recognition. The specific calculation process is as follows:

[0022] (1) In the drug small molecule dataset used for testing, each drug small molecule will obtain a ranked list of scores (rounded to two decimal places) of interactions with 859 targets from the model.

[0023] (2) If a drug small molecule has the same interaction score with multiple targets, we will further use the advanced method DeepPurpose to predict the interaction scores between the drug small molecule and the target and rank the obtained scores.

[0024] (3) If the true target of a drug small molecule is within the range of the Top-K targets identified by the model, the target identification of the drug small molecule is defined as correct. The recall evaluation index based on Top-K is used to measure the proportion of drug small molecule corresponding targets correctly identified in a given test drug small molecule dataset.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] (1) Use cutting-edge deep learning technologies such as Graph Transformer, MCNN (Multiple Convolutional Neural Network) and Residual EW-GCN to extract deep sequence and structural modal features from drugs and targets.

[0027] (2) Utilizing a cross-attention based multimodal fusion strategy, we can effectively achieve information complementarity within drugs and targets and feature fusion between drugs and targets.

[0028] (3) The benchmark dataset used is only one-third of the benchmark dataset used by the compared advanced methods. However, on the same target recognition test dataset, MM-IDTarget shows performance close to or better than the compared advanced methods in all Top-K evaluation indicators. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0030] Figure 1 This is a flow chart of a drug target identification method based on a deep learning framework of the present invention. DETAILED DESCRIPTION

[0031] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0032] See also Figure 1 The present invention provides a technical solution: a drug target identification method based on a deep learning framework, comprising the following steps:

[0033] Step 1: The Chemogenomic-Model dataset comes from the BindingDB database (Version 2020) and the ChEMBL27 database. The dataset contains 153,281 drug (compound or small molecule)-target (protein) pairs, which are related to 859 targets and 93,281 drug small molecules. After preprocessing the dataset, we obtained 47,247 drug-target pairs, including 23,481 positive samples and 23,766 negative samples. These drug-target pairs are used as the benchmark dataset for this study. The dataset is preprocessed as follows:

[0034] (1) Based on 859 targets, 153,281 drug-target pairs were divided into 859 sub-datasets;

[0035] (2) Clustering the compounds in each sub-dataset based on similarity, with the similarity threshold set to 0.3, thereby dividing each sub-dataset into multiple compound clusters;

[0036] (3) In each compound cluster, if the number of positive (negative) samples is 1, we will retain the sample; if the number of positive (negative) samples is greater than 1, we will extract 25% of the positive (negative) samples (round up the result) and retain them;

[0037] (4) Finally, in order to evaluate the performance of the model, the samples extracted from each cluster are further divided into a training set and a test set: if the number of positive (negative) samples in the cluster is less than or equal to 2, then all positive (negative) samples of the cluster are used as the training set; if the number of positive (negative) samples of the cluster is greater than 2 and less than or equal to 10, then 10% of the test set samples are randomly selected (rounded up), and the remaining samples are used as the training set; if the number of positive (negative) samples of the cluster is greater than 10, then 8% of the test set samples are randomly selected (rounded up), and the remaining positive (negative) samples are used as the training set.

[0038] After these division operations, the training set contains 42,394 samples (positive samples: 21,243, negative samples: 21,151), and the test set contains 4,853 samples (positive samples: 2,523, negative samples: 2,330), ensuring the balance of positive and negative samples in the dataset (Table 1).

[0039] Table 1 Statistical analysis of the dataset

[0040]

[0041] The drug-target interaction data also showed diversity (Table 2). For example, 292 targets had 10 or more known data points (the relationship between the target and the relevant drug (interaction or non-interaction) has been verified), 102 targets had more than 100 data points, and 19 targets had more than 500 data points. In addition, 183 targets participated in the formation of 10 or more positive samples, 69 targets participated in the formation of more than 100 positive samples, and 5 targets participated in the formation of more than 500 positive samples; 276 targets participated in the formation of 10 or more negative samples, 48 ​​targets participated in the formation of more than 100 negative samples, and 4 targets participated in the formation of more than 500 negative samples. On average, each target interacted with 110.3 drugs and participated in the formation of 86.7 positive samples and 67.2 negative samples.

[0042] Table 2 Distribution of total samples, positive samples and negative samples related to the target

[0043]

[0044] Finally, to test the performance of the model in identifying targets, we used the same comparison dataset as Chemogenomic-Model. The comparison dataset comes from the validation set of the target identification method SwissTargetPrediction (fromChEMBL24), which contains 67 different targets and 500 different drugs, of which 445 drugs contain 2 different human targets, 51 drugs contain 3 different human targets, 3 drugs contain 4 different human targets, and 1 drug contains 6 different human targets. At the same time, all data are annotated as direct binders with high activity (Ki, KD, IC50 or EC50) <1nM, and are associated with 1061 ligand-target interactions.

[0045] Step 2: In the label encoding feature representation of drug and target data, the label encoding of the drug is represented by its SMILES representation, mapping each character in the SMILES to a unique integer, thereby forming the label encoding representation of the drug. The label encoding of the target is mapped to a corresponding integer through each amino acid in the target sequence. For example, alanine (A) is mapped to 1, cysteine ​​(C) is mapped to 2, glutamic acid (E) is mapped to 4, and so on.

[0046] MCNN is used to extract multi-scale features of target sequences and drug SMILES. The three branches in MCNN consist of convolutional layers with different receptive fields. We stack three convolutional layers and pass the outputs of different convolutional layers through the maximum pooling layer respectively. The results of all outputs are added together as the multi-scale features of the sequence. First, we apply an embedding layer to map the label encoding representation of the sequence into a feature matrix s. Then, we input s into a feature extraction network composed of convolutional layers with different receptive fields. Finally, we add the results of the three different branches to obtain the multi-scale features Y of the sequence (Formula 1).

[0047]

[0048] Among them, R represents a tensor reshaped to the same size, C represents a standard convolutional layer, f represents the number of convolutional branches, and k represents the number of standard convolutional layers in each branch f.

[0049] Step 3: The physicochemical property characteristics of drug data are extracted from the SMILES expression of the drug through the molecular fingerprint tool (ECFP4). Specifically, the extended connection fingerprint (ECFP4) with a bond diameter of 2 is used, with each atom as the center, adjacent atoms are captured within a radius of 2, and connected in a certain order to form a feature matrix for substructure construction.

[0050] The physicochemical property characteristics of target data are composed of 7 types of descriptors, including amino acid composition descriptor, autocorrelation coefficient descriptor, CTD (composition, transition and distribution) descriptor, joint ternary descriptor, quasi-sequence order descriptor, pseudo amino acid composition descriptor and target chemometrics descriptor, etc. These descriptors are subjected to principal component analysis (PCA) to obtain the key features of the target.

[0051] Step 4: The graph structure feature representation of drug data is extracted from the SMILES expression of the drug through the RDKit tool. The vertices of the graph are composed of atoms in the drug molecules, while the edges represent the bond relationships between atoms. The feature of each vertex is a 44-dimensional vector, including information such as atomic symbol, degree, charge, number of free radical electrons, atomic hybridization type, whether it is aromatic, total number of hydrogen atoms and atomic chirality. The features of the edges of the drug graph structure are 10 dimensions in total, consisting of four categories: bond type between atoms, whether it is a conjugated bond, whether the bond is in a ring, and stereo configuration.

[0052] The Graph Transformer framework is used as an encoder to extract the graph structure modal features of drugs. Each Graph Transformer layer includes a multi-head attention mechanism and a feedforward neural network to transform and update node features.

[0053] In drug graph G D In the example, the feature of node i is expressed as (d n is the node feature dimension), the edge feature between nodes i and j is represented by (d e is the edge feature dimension), and the features of nodes and edges are linearly projected to obtain their hidden representations (Formula 2) and (Formula 3):

[0054]

[0055] in, are learnable parameters, a0, b0∈R d is the deviation (d represents the dimension). In order to further encode the position λ i (Formula 4) added to the node feature In (Formula 5), ​​we use the Laplacian eigenvector (Formula 6) as the position encoding in the graph conversion network. The Laplacian eigenvector Δ of each graph can be pre-calculated by factorizing the graph Laplacian matrix:

[0056]

[0057] In formulas 4 and 5, the k smallest non-trivial feature vectors of node i are used as their corresponding positional encoding representation λ i (λ i ∈R k×1 ), is a learnable parameter, Indicates deviation, Represents the position encoding feature matrix. In Formula 6, I is the identity matrix, A is the graph adjacency matrix, D is the degree matrix of the graph, and F and U represent eigenvalues ​​and eigenvectors, respectively. The graph converter updates node and edge features, mainly based on the multi-head attention mechanism. The update process of the l+1 layer node and edge features is shown in Formulas 7 to 9:

[0058]

[0059]

[0060] Among them, Q k，l ,K k，l ,V k，l , is the parameter of the linear layer. k∈[1,gt] represents the number of attention heads. k Represents the dimension of each attention head. represents the self-attention weight matrix after adding edge features. Next, the node features of the l+1 layer and edge features are input into the feedforward network respectively, and through normalization (Formula 12) and ReLU activation function, the final node features of the l+1 layer are obtained. (Formula 10) and edge features (Formula 11):

[0061]

[0062] in, and represents a learnable parameter.

[0063]

[0064] Among them, X d The original characteristics of the drug, and are the maximum and minimum values, respectively. is the feature matrix after minimum-maximum normalization. In addition, we also perform a maximum pooling operation on the obtained drug feature matrix (Formula 13).

[0065]

[0066] in, represents the j-th eigenvector in the input matrix, Indicates the number of feature matrix columns.

[0067] Step 5: The graph structure feature representation of the target data is constructed based on the tertiary structure of the target. First, the carbon alpha atom of each amino acid residue is used as the vertex of the graph. Then, whether there is an edge is determined based on whether the Euclidean distance between any two carbon alpha atoms is less than or equal to 8.0 angstroms. In the adjacency matrix of the graph, the position where the edge exists is set to 1, otherwise it is set to 0. The vertex feature matrix of the target is a 33-dimensional vector, including features such as amino acid type, molecular weight, acidity and alkalinity, hydrophobicity, whether it is fat, whether it is aromatic, etc. The graph structure features of the target use not only vertex features but also edge features. The edge features of the target graph structure have a total of 3 dimensions, which are the similarity, distance and cosine value of the angle between carbon a atoms.

[0068] The Residual EW-GCN architecture is used to extract the structural modal features from the target graph. The Residual EW-GCN architecture is derived from GPCNDTA and consists of three EW-GCN layers. Unlike GCN, EW-GCN can update the features of a node using the weights of all edges connected to the node. The input of EW-GCN is the graph, node features, and the weight matrix of the edges connected to the node. In order to accelerate the convergence of the model, we connect a normalization layer after each EW-GCN layer.

[0069] In the process of node feature update, we select two C α The maximum value of the cosine similarity of the atomic feature vector, the inverse of the Euclidean distance, and the cosine value of the angle is used to construct the edge weight matrix

[0070] Thus, we can obtain the adjacency matrix of Laplacian edge weighted node x (Formula 14):

[0071]

[0072] in, for The diagonal matrix of and are the adjacency matrix and identity matrix of the node respectively, is the dimension of the feature matrix, ew represents the weight of the edge, and T represents the target. The adjacency matrix of the edge-weighted node and the node features of the lth layer So as to obtain the characteristics of the l+1th layer nodes (Formula 15).

[0073]

[0074] in, represents a learnable parameter. For the target graph, we use the node features as the residual of EW-GCN. First, the residual of EW-GCN layer 1 is concatenated with the vertex features after linear transformation and used as the input of EW-GCN layer 2, while the output of layer 2 is used as the input of layer 3. In the residual EW-GCN, we use element-by-element summation to aggregate the features of all nodes in the target graph.

[0075] Step 6: MM-IDTarget uses the sequence modal features and physicochemical modal features of drugs and targets. In order to test the impact of these modal features on the performance of MM-IDTarget, we conducted 7 variant experiments of MM-IDTarget based on the physicochemical features and sequence modal features of drugs and targets. In the process of conducting these variant experiments, the structural modal feature extraction process of drugs and targets in the variants remains unchanged. From the experimental results (Table 3), it is not difficult to find that when there are only sequence modal features of drugs and targets, the performance indicators of the variant models do not exceed 50%, while when there are only drug molecular fingerprints or physicochemical properties of targets, the variant models all show good performance, especially when based only on the physicochemical properties of the target, the performance of the variant model is only slightly different from that of the MM-IDTarget model, indicating that the physicochemical characteristics of the target play a key role in the target recognition process. However, based on the physicochemical modal features of the target, the sequential modal features of the drug and the target are integrated, and the constructed MM-IDTarget method achieves better performance than all variants. The Top-K indicators (K = 1, 3, 5, 7, 10) are 0.83%, 1.41%, 2.56%, 1.41% and 0.84% ​​higher than the optimal performance of the variants, respectively. These experimental results show that integrating the sequential modal features and physicochemical modal features of the drug and the target is conducive to the complementary fusion of key information affecting the drug-target interaction, thereby obtaining comprehensive information on the drug-target interaction and improving the performance of the target recognition model.

[0076] Table 3 Comparison of target recognition performance of variant models generated by sequence modality and physicochemical modality

[0077]

[0078] A: represents the molecular fingerprint ECPF4 B: represents the physicochemical properties of the target C: sequence modal features extracted from drug SMILES and target sequences

[0079] Step 7: In addition to the sequence modal features and physicochemical modal features of drugs and targets, the structural modal features of drugs and targets are also important components of the multimodal features of MM-IDTarget. Therefore, we also conducted a comprehensive evaluation of the variant models for extracting the structural modal features of drugs and targets. We conducted experiments on MM-IDTarget and three variant models while retaining the physicochemical modal features and sequence modal features of drugs and targets. The experimental results (Table 4) show that in the absence of target structural modal features, the two variant models (Variant 2 and Variant 3) achieved similar performance under all Top-K indicators. After only introducing the target structural modal features, the variant model (Variant 1) has a slightly lower Top-1 index value than the variant model (Variant 3), and all other Top-K index values ​​are improved compared with the optimal performance achieved by the variant model without target structural modal features. However, when the structural modal features of drugs and targets are integrated at the same time, MM-IDTarget shows better performance than all variant models, for example, Top-1, Top-3, Top-5, Top-7 and Top-10 are improved by 0.95%, 3.19%, 3.59%, 3.3% and 3.49% respectively. This shows that the structural modal features of drugs and targets are an indispensable part of the multimodal features used by the target recognition model, and can effectively improve the performance of the target recognition model.

[0080] Table 4 Comparison of performance of variant models generated by structural modalities in identifying targets

[0081]

[0082] D: Drug structure mode E: Target structure mode

[0083] Step 8: In this study, we used both inner cross and outer cross to fuse the features of the sequence and structural modalities of the drug and the target. The inner cross part completes the feature complementation of the sequence and structural modalities of the drug (or target), while the outer cross part completes the feature fusion of the drug and the target. In order to evaluate the impact of inner cross and outer cross on the performance of the target recognition model, we conducted experiments on the corresponding variant models. The experimental results show (Table 5) that the performance of the variant models without cross attention and only using inner cross or outer cross is not much different, while the variant model using only inner cross achieved the best performance. Based on the variant model using only inner cross, we further introduced the outer cross to form the MM-IDTarget method, which showed better performance than all variant models. The experimental results of these variant models can illustrate that the feature complementation of the same object completed by the inner cross and the feature fusion of different objects completed by the outer cross are helpful to improve the performance of the target recognition model.

[0084] Table 5 Comparison of target recognition performance of cross-attention variant models

[0085]

[0086] F: Modal fusion based on inner cross attention G: Modal fusion based on outer cross attention

[0087] Step 9: To further evaluate the generalization ability of MM-IDTarget, we compared the generalization performance of MM-IDTarget with that of Chemogenomic-Model, an advanced method with the best target identification performance, on two additional test sets: PDSP Ki (778) and NPASS (181). The test sets PDSP Ki (778) and NPASS (181) are consistent with the test sets used to verify the generalization ability of the model in Chemogenomic-Model. The PDSP Ki (778) dataset consists of 778 compound-target interaction pairs consisting of 442 compounds, while the NPASS (181) dataset consists of 181 compound-target interaction pairs consisting of 122 compounds. The experimental results (Table 6) show that, except for the Top-1 indicator on the PDSP Ki (778) dataset, MM-IDTarget is slightly weaker than Chemogenomic-Model, but all other indicators on the PDSP Ki (778) dataset and all indicators on the NPASS (181) dataset are better than Chemogenomic-Model, and the value of each indicator is increased by more than 10%. In addition, on two different additional test sets, most evaluation indicators of MM-IDTarget exceeded 50%, which shows that MM-IDTarget has strong practical application capabilities.

[0088] Table 6 Performance comparison of MM-IDTarget and advanced methods

[0089]

[0090] Step 10: In order to fully evaluate the practical application capabilities of MM-IDTarget, in addition to evaluating its performance on the generalized dataset, we also use MM-IDTarget to identify the targets of approved drugs. In this study, the number of targets in the benchmark dataset we used is 859, among which some targets have corresponding approved therapeutic drugs. In order to obtain samples for the case study, we searched for drugs corresponding to 859 targets in the TTD database (Therapeutic Target Database), and obtained corresponding samples of approved drugs and targets. In order to avoid the obtained drug-target pairs already appearing in the benchmark dataset, we deleted the samples existing in the benchmark dataset and randomly selected 40 samples from the remaining samples to construct the case study dataset TTD_40. The 40 drug-target pairs in TTD_40 consist of 38 targets and 38 drugs. The experimental results in Table 7 show that MM-IDTarget also showed good performance in the application of identifying targets of approved drugs, with the evaluation indicators Top-5 and Top-7 exceeding 50%, and Top-10 exceeding 70%. From the results of the case study, MM-IDTarget can well integrate the multimodal features of drugs and targets to obtain comprehensive features that affect the performance of the target identification model. It is an effective and trustworthy target identification tool.

[0091] Table 7 Case performance analysis of MM-IDTarget method

[0092]

[0093] Step 11: Compare the true label with the predicted result and minimize the cross entropy loss function. The calculation process of the loss function is shown in Formula 16.

[0094]

[0095] Where n represents the number of samples, y true and pred Represents the true label and predicted label.

[0096] Compared with advanced target recognition methods, these advanced methods include HitPickV2, PPB2, PPB, TargetNet, SwissTargetPrediction and Chemogenomic-model. Among these methods, PPB2 applies a variety of combined methods to predict targets. In this study, we only compare with the results with the best target recognition performance. In addition, the Top-K index values ​​of these methods for target recognition are all derived from Chemogenomic-model. Compared with the best performance of advanced methods, MM-IDTarget is slightly lower than HitPickV2 by 0.86% in the Top-3 index, but it is improved by 6.68%, 1.39%, 1.24% and 1.32% in the Top-1, Top-5, Top-7 and Top-10 indicators respectively. It is worth noting that although these advanced methods and MM-IDTarget use the same dataset to evaluate the model target identification performance, MM-IDTarget has only 47,247 samples, which is only about one-third of these advanced methods. However, when the target recognition model uses a much smaller dataset sample for training and testing than other advanced methods, MM-IDTarget achieves performance similar to or higher than the compared advanced methods in all evaluation indicators, which shows that MM-IDTarget is an efficient and reliable target recognition method. The indicators are shown in Table 8.

[0097] Table 8 Performance comparison of model MM-IDTarget and advanced methods

[0098]

[0099] MM-IDTarget is a novel deep learning framework for drug target identification. Various cutting-edge deep learning techniques, such as Graph Transformer, MCNN, and Residual EW-GCN, are used to extract deep sequence and structural modal features from drugs and targets. Then, the cross-attention mechanism is used to extract multimodal features for information complementation within drugs and targets and feature fusion between drugs and targets. At the same time, the molecular fingerprint of drugs and the physicochemical characteristics of targets are integrated to obtain key features that affect drug-target interactions. Finally, although the benchmark dataset we use is only one-third of the benchmark dataset used by the compared advanced methods, on the same target identification test dataset, MM-IDTarget shows performance close to or better than the compared advanced methods in all Top-K evaluation indicators. In addition, we further comprehensively tested the application capabilities of MM-IDTarget under the generalization ability test dataset and actual application cases, and found that MM-IDTarget can still show good target identification performance, which shows that MM-IDTarget is an effective and trustworthy target identification tool.

[0100] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the attached claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any figure mark in the claims should not be regarded as limiting the claims involved.

[0101] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A drug target identification method based on a deep learning framework, characterized in that: The following steps are involved: S1: Obtain the drug-target interaction Chemogenomic-Model dataset; S2: Obtain the target's structure file, and obtain the target's structure feature representation and edge feature representation through the target's structure file; S3: Obtain sequence tag encoding feature representation, graph structure feature representation and physicochemical feature representation of drugs and targets; S4: The sequence label encoding feature representation of the drug is sent to a three-layer multi-scale convolutional neural network. The step size of each layer is 1, the padding is 0, and the sizes of the convolution kernels are 3x3, 5x5 and 7x7 respectively. Each convolution layer is followed by a normalization operation to obtain the three-layer features of the drug in the multi-scale convolutional neural network, concatenate them and generate the sequence label encoding features through linear transformation; S5: The structural feature representation of the drug graph is sent to a three-layer graph converter framework to extract the structural modal features of the drug. The number of heads in each layer of the graph converter is 8, the number of neurons in each hidden layer is 128, each hidden layer is followed by a normalization operation, and the maximum pooling operation is connected at the end; S6: Send the target graph structure feature representation to the Residual EW-GCN graph convolutional network. The Residual EW-GCN contains three GCN layers and a fully connected network. First, the target graph feature representation is input into the first GCN layer and the fully connected layer respectively. Then, the output features of the first GCN layer and the output features of the fully connected network are concatenated. Then, the concatenated features are passed through the second and third GCN layers in turn to obtain features with residuals. Finally, L2 normalization is performed on them. S7: Cross-attention is used to fuse the sequence label encoding feature representation and graph structure feature representation of the drug, and then the fusion feature representation of the drug and the target is fused to obtain the drug-target cross-attention fusion feature representation; S8: After the obtained drug-target cross-attention fusion feature representation is concatenated with the drug physicochemical feature representation, it is sent to the fully connected layer as a comprehensive feature of drug-target interaction to predict drug-target interaction; S9: Compare the true label with the predicted result, and update the network parameters to minimize the cross entropy loss function until the maximum number of iterations is reached. The training is completed and the trained MM-IDTarget model is obtained. S10: For the trained MM-IDTarget model, the recall based on Top-K is used to evaluate the performance of the model in drug target recognition. S11: Compare the true label with the predicted result and minimize the cross entropy loss function.

2. The method for drug target identification based on a deep learning framework according to claim 1, characterized in that: The Chemogenomic-Model dataset in S1 is derived from the BindingDB database (Version 2020) and the ChEMBL27 database.

3. A drug target identification method based on a deep learning framework according to claim 2, characterized in that: The dataset contains 153,281 drug-target pairs, which are associated with 859 targets and 93,281 drug small molecules. After preprocessing the dataset, 47,247 drug-target pairs were obtained, including 23,481 positive samples and 23,766 negative samples.

4. The method for drug target identification based on a deep learning framework according to claim 3, characterized in that: The preprocessing steps of the dataset are as follows: (1) Based on 859 targets, drug-target pairs were divided into 859 sub-datasets; (2) Clustering the compounds in each sub-dataset based on similarity, with the similarity threshold set to 0.3, thereby dividing each sub-dataset into multiple compound clusters; (3) In each compound cluster, if the number of positive or negative samples is 1, retain the sample; if the number of positive or negative samples is greater than 1, extract 25% of the positive or negative samples and retain them; (4) The samples extracted from each cluster are further divided into a training set and a test set: if the number of positive or negative samples in the cluster is less than or equal to 2, then all positive or negative samples of the cluster are used as the training set; if the number of positive or negative samples of the cluster is greater than 2 and less than or equal to 10, then 10% of the test set samples are randomly selected, and the remaining samples are used as the training set; if the number of positive or negative samples of the cluster is greater than 10, then 8% of the test set samples are randomly selected, and the remaining positive or negative samples are used as the training set.

5. The method for drug target identification based on a deep learning framework according to claim 1, characterized in that: The specific calculation process in S10 is as follows: (1) In the drug small molecule dataset used for testing, each drug small molecule will obtain a score-ranked list of interactions with 859 targets from the model; (2) If a drug small molecule has the same interaction score with multiple targets, DeepPurpose is further used to predict the interaction scores between the drug small molecule and the target, and the obtained scores are ranked; (3) If the true target of a drug small molecule is within the range of the Top-K targets identified by the model, the target identification of the drug small molecule is defined as correct. The recall evaluation index based on Top-K is used to measure the proportion of drug small molecule corresponding targets correctly identified in a given test drug small molecule dataset.

6. The method for drug target identification based on a deep learning framework according to claim 1, characterized in that: The calculation process of the loss function in S11 is shown in the formula: Where n represents the number of samples, y true and pred Represents the true label and predicted label.

Images