A Drug Target Identification Method Based on Deep Learning Framework

Abstract

This invention relates to the field of computer bioinformatics technology and discloses a drug target identification method based on a deep learning framework, comprising the following steps: S1: acquiring the drug-target interaction Chemogenomic-Model dataset; S2: acquiring the target structure file, and obtaining the target's structural feature representation and edge feature representation through the target structure file; S3: acquiring the sequence label encoding feature representation, graph structure feature representation, and physicochemical feature representation of the drug and target; S4: feeding the drug's sequence label encoding feature representation into a three-layer multi-scale convolutional neural network, with a stride of 1 and padding of 0 for each layer, and convolutional kernel sizes of 3x3, 5x5, and 7x7, respectively. This invention uses only one-third the benchmark dataset used by the compared state-of-the-art methods, but on the same target identification test dataset, MM-IDTarget demonstrates performance close to or better than the compared state-of-the-art methods on all Top-K evaluation metrics.

Description

Technical Field

[0001] This invention relates to the field of computer bioinformatics technology, specifically to a method for drug target identification based on a deep learning framework. Background Technology

[0002] Accurate identification of drug targets plays a crucial role in research fields such as drug design, drug retargeting, and treatment of diseases with unclear pathogenic mechanisms. Over the past few decades, computer-aided drug target identification methods have evolved from structure- and ligand (drug)-based methods to the current data-driven methods based on machine learning. Data-driven methods based on machine learning can identify targets with biological activity related to a given drug from a massive number of drug-target pairs with relatively fast speed and high accuracy. Data-driven methods based on machine learning for identifying drug targets can be divided into two categories: methods based on traditional machine learning and methods based on deep learning.

[0003] In deep learning-based drug-target binding affinity prediction, the prediction of drug-target binding affinity is generally treated as a regression task, predicting the strength of the interaction between the drug and the target. Classical methods include sequence-based methods such as DeepDTA and DeepDTAF, combined 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. Research on target probability ranking based on deep learning further ranks the identified potential targets using a Top-K ranking method, building upon drug-target interactions and binding affinity predictions. Examples include the CODD-Pred method and the GuiltyTargets-COVID-19 method.

[0004] Although most deep learning-based methods incorporate target features and significantly improve model performance metrics such as accuracy, root mean squared error (RMSE), and Pearson correlation coefficient (PCC) compared to traditional machine learning methods, only a small portion of these studies utilize graph neural networks to extract structural features influencing drug target identification from drugs and targets. Graph neural networks are still in the initial exploratory stage in drug target identification research. Furthermore, existing deep learning-based methods for identifying drug targets based on drug-target interactions or affinity only consider the model's classification or regression performance metrics, rarely calculating the Top-K metric, a crucial evaluation indicator for assessing the model's target identification quality. Further utilizing graph neural networks, a cutting-edge deep learning technology, to construct a drug target identification model that meets the user's desired Top-K threshold remains a significant challenge. Therefore, we propose a drug target identification method based on a deep learning framework. Summary of the Invention

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

[0006] To achieve the above objectives, 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: Obtaining the drug-target interaction dataset. The Chemogenomic-Model dataset is derived from the BindingDB database (Version 2020) and the ChEMBL27 database. This dataset contains 153,281 drug (compound or small molecule)-target (protein) pairs, which are associated with 859 targets and 93,281 drug small molecules. After preprocessing this dataset, we obtained 47,247 drug-target pairs, including 23,481 positive samples and 23,766 negative samples. These drug-target pairs serve as the baseline dataset for this study. The preprocessing steps for the dataset are as follows:

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

[0010] (2) We perform similarity clustering on the compounds in each subset, with a similarity threshold of 0.3, thereby dividing each subset into multiple compound clusters;

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

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

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

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

[0015] S4: The sequence label encoding features of the drug (target) are fed into a three-layer multi-scale convolutional neural network with a stride of 1 and padding of 0 for each layer. The kernel sizes are 3x3, 5x5 and 7x7 respectively. A normalization operation is followed by each convolutional layer to obtain the three-layer features of the drug (target) in the multi-scale convolutional neural network. These features are then concatenated and a linear transformation is performed to generate the sequence label encoding features.

[0016] S5: The structural feature representation of the drug graph is fed into a three-layer graph transformer framework to extract the structural modality features of the drug. Each layer of the graph transformer has 8 heads, each hidden layer has 128 neurons, each hidden layer is followed by a normalization operation, and finally a max pooling operation is connected.

[0017] S6: The target image structural feature representation is fed into 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 one fully connected network. First, the target image 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.

[0018] S7: The sequence label encoding feature representation and graph structure feature representation of the drug (target) are fused using cross-attention, and then the fusion feature representations of the drug and the target are fused to obtain the drug-target cross-attention fusion feature representation.

[0019] S8: The drug-target cross-attention fusion feature representation and the drug (target) physicochemical feature representation are concatenated and fed into the fully connected layer as a comprehensive feature of drug-target interaction to predict drug-target interaction.

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

[0021] S10: For the trained MM-IDTarget model, we use Top-K based recall to evaluate the model's drug target recognition performance. 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 sorted list of scores (rounded to two decimal places) of interaction with 859 targets from the model.

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

[0024] (3) If the true target of a small drug molecule falls within the range of the Top-K targets identified by the model, then the target identification of that small drug molecule is defined as correct. The Top-K-based recall metric measures the proportion of correctly identified targets for a given test small drug molecule dataset.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

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

[0027] (2) By utilizing a multimodal fusion strategy based on cross-attention, information complementarity within drugs and targets and feature fusion between drugs and targets can be effectively achieved.

[0028] (3) The benchmark dataset used is only one-third of the benchmark dataset used by the state-of-the-art methods in comparison. However, on the same target recognition test dataset, MM-IDTarget shows performance close to or better than the state-of-the-art methods in all Top-K evaluation metrics. Attached Figure Description

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

[0030] Figure 1 This is a flowchart of a drug target identification method based on a deep learning framework according to the present invention. Detailed Implementation

[0031] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0032] Please see Figure 1 This 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 is derived from the BindingDB database (Version 2020) and the ChEMBL27 database. This dataset contains 153,281 drug (compound or small molecule)-target (protein) pairs, which are associated with 859 targets and 93,281 drug small molecules. After preprocessing this dataset, we obtained 47,247 drug-target pairs, including 23,481 positive samples and 23,766 negative samples. These drug-target pairs serve as the baseline dataset for this study. The dataset preprocessing method is as follows:

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

[0035] (2) Cluster the compounds in each subset of data based on similarity, with a similarity threshold of 0.3, thereby dividing each subset of data into multiple compound clusters;

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

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

[0038] After these partitioning 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 a balance between positive and negative samples in the dataset (Table 1).

[0039] Table 1 Statistical Analysis of the Dataset

[0040]

[0041] Drug-target interaction data also exhibited diversity (Table 2). For example, 292 targets had 10 or more known data points (the relationship between the target and related drugs (interaction or non-interaction) has been validated), 102 targets had over 100 data points, and 19 targets had over 500 data points. Furthermore, 183 targets participated in the formation of 10 or more positive samples, 69 targets participated in the formation of over 100 positive samples, and 5 targets participated in the formation of over 500 positive samples; 276 targets participated in the formation of 10 or more negative samples, 48 ​​targets participated in the formation of over 100 negative samples, and 4 targets participated in the formation of over 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 model's target recognition performance, we used the same comparative dataset as the Chemogenomic-Model. This comparative dataset is derived from the validation set of the SwissTargetPrediction target recognition method (from ChEMBL24), and contains 67 different targets and 500 different drugs. Among these, 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. All data were also annotated as direct binders with high activity (Ki, KD, IC50, or EC50) <1 nM, and associated with 1061 ligand-target interactions.

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

[0046] MCNN was used to extract multi-scale features from target sequences and drug samples. The three branches of MCNN consist of convolutional layers with different receptive fields. We stacked the three convolutional layers and passed the outputs of each layer through a max-pooling layer. The sum of all outputs was then used as the multi-scale feature of the sequence. First, we applied an embedding layer to map the sequence's label encoding representation into a feature matrix s. Then, we input s into the feature extraction network composed of convolutional layers with different receptive fields. Finally, we summed the results of the three different branches to obtain the multi-scale feature Y of the sequence (Equation 1).

[0047]

[0048] Where 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 properties of the drug data are represented by extracting them from the drug's SMILES expression using the molecular fingerprinting tool (ECFP4). Specifically, an extended linkage fingerprint (ECFP4) with a bond diameter of 2 is used. Each atom is centered on an adjacent atom within a radius of 2, and these adjacent atoms are linked in a specific order to form a feature matrix constructed from the substructure.

[0050] The physicochemical properties of the target data are represented by seven types of descriptors, including amino acid composition descriptors, autocorrelation coefficient descriptors, CTD (composition, transition, and distribution) descriptors, joint ternary descriptors, quasi-sequence order descriptors, pseudo-amino acid composition descriptors, and target chemometric descriptors. These descriptors were all subjected to principal component analysis (PCA) to obtain the key features of the target.

[0051] Step 4: The graph structure feature representation of the drug data is extracted from the SMILES expression of the drug using the RDKit tool. The vertices of the graph consist of atoms in the drug molecule, while the edges represent the bond relationships between atoms. Each vertex's feature is a 44-dimensional vector, including information such as atom symbol, degree, charge, number of free radical electrons, atomic hybridization type, whether it is aromatic, total number of hydrogen atoms, and atomic chirality. The edge features of the drug graph structure are 10-dimensional, consisting of four categories: bond type between atoms, whether it is a conjugated bond, whether the bond is in a ring, and 3D configuration.

[0052] A graph transformer framework is used as an encoder to extract graph structure modal features of the drug. Each graph transformer layer includes a multi-head attention mechanism and a feedforward neural network for transforming and updating node features.

[0053] In drug diagram G D In the diagram, the features of node i are used... (d n (This is represented by the node feature dimension), and the edge features between nodes i and j are represented by... (d e The hidden representation is obtained by linearly projecting the features of nodes and edges (which are the edge feature dimensions). (Formula 2) and (Formula 3):

[0054]

[0055] in, These are learnable parameters, a0, b0 ∈ R. d The bias is denoted by d (where d represents the dimension). To further encode the position λ... i (Formula 4) Add to node features In Equation 5, we use the Laplacian eigenvector (Equation 6) as the positional encoding in the graph transformation network. The Laplacian eigenvector Δ for each graph can be pre-calculated through factorization of the graph's Laplacian matrix.

[0056]

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

[0058]

[0059]

[0060] Among them, Q k，l ,K k，l V k，l , These are the parameters of the linear layer. k∈[1,gt] represents the number of attention heads. d k This represents the dimension of each attention head. This represents the self-attention weight matrix after incorporating edge features. Following this, the node features of layer l+1... Sum of edge features The data are fed into the feedforward network, and after normalization (Equation 12) and ReLU activation, the node features of the final l+1 layer are obtained. (Formula 10) and edge features (Formula 11):

[0061]

[0062] in, and This represents the learnable parameters.

[0063]

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

[0065]

[0066] in, This represents the eigenvector of the j-th column in the input matrix. This indicates the number of columns in the characteristic matrix.

[0067] Step 5: The graph structure feature representation of the target data is constructed based on the target's tertiary structure. First, the carbon α atom of each amino acid residue is used as the vertex of the graph. Then, the existence of an edge is determined by whether the Euclidean distance between any two carbon α atoms is less than or equal to 8.0 Å. In the adjacency matrix of the graph, positions with existing edges are set to 1, and those without are 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 / basicity, hydrophobicity, whether it is fatty, and whether it is aromatic. The graph structure feature of the target uses not only vertex features but also edge features. The edge features of the target graph structure are three-dimensional, representing the similarity, distance, and cosine value of the angle between carbon α atoms.

[0068] A Residual EW-GCN architecture is employed to extract structural modal features from the target graph. Adapted from GPCNDTA, the Residual EW-GCN architecture consists of three EW-GCN layers. Unlike GCN, EW-GCN updates node features using the weights of all edges connected to the nodes. The input to EW-GCN is the graph, node features, and the weight matrix of edges connected to the nodes. To accelerate model convergence, a normalization layer is appended after each EW-GCN layer.

[0069] During the node feature update process, we select two Cs by row. α The edge weight matrix is ​​constructed using the maximum value among the cosine similarity of atomic eigenvectors, the reciprocal of the Euclidean distance, and the cosine of the included angle.

[0070] This yields the adjacency matrix of the Laplace edge-weighted node x. (Formula 14):

[0071]

[0072] in, for diagonal matrix, and These are the adjacency matrix and identity matrix of the nodes, respectively. Let be the dimension of the feature matrix, ew represent the edge weights, and T represent the target. The adjacency matrix of nodes is calculated by weighting the edges. and node features of layer l Thus, the features of the (l+1)th layer node are obtained. (Formula 15).

[0073]

[0074] in, This represents the learnable parameters. For the target image, we use node features as the residuals of the EW-GCN. First, the residuals of the first layer of the EW-GCN are concatenated with the vertex features after linear transformation, and this is used as the input to the second layer of the EW-GCN. The output of the second layer is used as the input to the third layer. In the residual EW-GCN, we aggregate the features of all nodes in the target image using element-wise summation.

[0075] Step 6: MM-IDTarget utilizes sequence modal features and physicochemical modal features of drugs and targets. To test the impact of these modal features on the performance of MM-IDTarget, we conducted seven variant experiments of MM-IDTarget based on the physicochemical and sequence modal features of drugs and targets. During these variant experiments, the extraction process of structural modal features of drugs and targets remained unchanged. From the experimental results (Table 3), it is evident that when only sequence modal features of drugs and targets are available, the performance indicators of the variant models did not exceed 50%. However, when only drug molecular fingerprints or target physicochemical properties are available, the variant models all exhibited better performance. Especially when based solely on the physicochemical properties of targets, the performance of the variant models was only slightly different from that of the MM-IDTarget model, indicating that the physicochemical features of targets play a crucial role in target identification. However, by integrating the sequence modal features of both the drug and the target based on the physicochemical modal features of the target, the constructed MM-IDTarget method achieved better performance than all variants. The Top-K metrics (K = 1, 3, 5, 7, 10) were improved by 0.83%, 1.41%, 2.56%, 1.41%, and 0.84% ​​respectively compared to the best performance obtained by the variants. These experimental results demonstrate that integrating the sequence and physicochemical modal features of both the drug and the target facilitates the complementary fusion of key information influencing drug-target interactions, thereby obtaining comprehensive information on drug-target interactions and improving the performance of the target identification model.

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

[0077]

[0078] A: Represents the molecular fingerprint ECPF4; B: Represents the physicochemical properties of the target; C: Sequence modality features extracted from drug smiles and target sequences.

[0079] Step 7: In addition to the sequence modalities and physicochemical modalities of the drug and target, the structural modalities of the drug and target are also an important component of the multimodal features of MM-IDTarget. Therefore, we also comprehensively evaluated the variant models extracted from the structural modalities of the drug and target. While retaining the physicochemical and sequence modalities of the drug and target, we conducted experiments with MM-IDTarget and three variant models. The experimental results (Table 4) show that without target structural modalities, the two variant models (Variant 2 and Variant 3) achieved similar performance across all Top-K metrics. However, after introducing only target structural modalities, the variant model (Variant 1) showed improved performance in all Top-K metrics compared to the variant model without target structural modalities, except for a slightly lower Top-1 metric value. However, when structural modal features of both drugs and targets are integrated simultaneously, MM-IDTarget exhibits superior performance compared to all variant models. For example, Top-1, Top-3, Top-5, Top-7, and Top-10 show improvements of 0.95%, 3.19%, 3.59%, 3.3%, and 3.49%, respectively. This demonstrates that structural modal features of drugs and targets are an indispensable component of the multimodal features employed in target recognition models, effectively enhancing their performance.

[0080] Table 4 Comparison of target recognition performance of variant models generated by structural modalities

[0081]

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

[0083] Step 8: In this study, we employed both inner-layer and outer-layer crossover to fuse the sequence and structural modalities of the drug and target. The inner-layer crossover part completed the feature complementation of the drug's (or target's) sequence and structural modalities, while the outer-layer crossover part completed the feature fusion of the drug and target. To evaluate the impact of inner-layer and outer-layer crossover on the performance of the target recognition model, we conducted experiments on corresponding variant models. The experimental results show (Table 5) that the performance of variant models without cross-attention and those using only inner-layer or outer-layer crossover was not significantly different, while the variant model using only inner-layer crossover achieved the best performance. Based on the variant model using only inner-layer crossover, we further introduced outer-layer crossover to form the MM-IDTarget method, which showed better performance than all variant models. The experimental results of these variant models demonstrate that the feature complementation of the same object completed by inner-layer crossover and the feature fusion of different objects completed by outer-layer crossover help improve the performance of the target recognition model.

[0084] Table 5. Performance comparison of cross-attention variant models in target recognition

[0085]

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

[0087] Step 9: To further evaluate the generalization ability of MM-IDTarget, we compared its generalization performance with that of the state-of-the-art Chemogenomic-Model, which has the best target recognition 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 in Chemogenomic-Model to verify the generalization ability of the model. The PDSP Ki(778) dataset consists of 778 compound-target interaction pairs consisting of 442 compounds, while NPASS(181) consists of 181 compound-target interaction pairs consisting of 122 compounds. The experimental results (Table 6) show that, except for a slight weakness in the Top-1 metric compared to Chemogenomic-Model on the PDSP Ki(778) dataset, MM-IDTarget outperforms Chemogenomic-Model on all other metrics on the PDSP Ki(778) dataset and on all metrics on the NPASS(181) dataset, with each metric showing an improvement of more than 10%. Furthermore, on two different additional test sets, MM-IDTarget achieved more than 50% in most evaluation metrics, demonstrating its strong practical application capabilities.

[0088] Table 6 Performance Comparison of MM-IDTarget and Advanced Methods

[0089]

[0090] Step 10: To comprehensively evaluate the practical application capabilities of MM-IDTarget, in addition to evaluating its performance on a generalized dataset, we also used MM-IDTarget to identify targets of approved drugs. In this study, the benchmark dataset we used contained 859 targets, some of which already had corresponding approved therapeutic drugs. To obtain samples for the case study, we searched the TTD (Therapeutic Target Database) for the drugs corresponding to the 859 targets, thus obtaining samples of approved drugs and targets. To avoid obtaining drug-target pairs that already appeared in the benchmark dataset, we deleted 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. Table 7 shows the experimental results, indicating that MM-IDTarget also demonstrated good performance in identifying targets of approved drugs, with Top-5 and Top-7 scores exceeding 50%, and Top-10 scores exceeding 70%. The case study results demonstrate that MM-IDTarget effectively integrates multimodal features of drugs and targets, thereby obtaining comprehensive features that influence the performance of the target identification model, making it an effective and reliable target identification tool.

[0091] Table 7 Case Performance Analysis of the MM-IDTarget Method

[0092]

[0093] Step 11: Compare the true labels with the predicted results and minimize the cross-entropy loss function. The calculation process of the loss function is shown in Equation 16.

[0094]

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

[0096] Compared with state-of-the-art target identification methods, including HitPickV2, PPB2, PPB, TargetNet, SwissTargetPrediction, and Chemogenomic-model, MM-IDTarget uses a combination of methods to predict targets. In this study, we only compare its performance with the best-performing method. Furthermore, the Top-K metrics for target identification for these methods are derived from the Chemogenomic-model. While MM-IDTarget's performance is slightly lower than HitPickV2 (0.86% lower for Top-3 metrics), it improves performance by 6.68%, 1.39%, 1.24%, and 1.32% for Top-1, Top-5, Top-7, and Top-10 metrics, respectively. It is noteworthy that although these state-of-the-art methods use the same dataset to evaluate their target identification performance as MM-IDTarget, MM-IDTarget has only 47,247 samples, approximately one-third the size of these state-of-the-art methods. However, given that the training and testing datasets for the target recognition model are significantly smaller than those of other state-of-the-art methods, MM-IDTarget achieved performance close to or exceeding that of the state-of-the-art methods across all evaluation metrics. This demonstrates that MM-IDTarget is an efficient and reliable target recognition method. The metrics 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-level sequence and structural modal features from drugs and targets. Subsequently, a cross-attention mechanism is used to complement the extracted multimodal features within both the drug and target domains, as well as to fuse features between the drug and target. Simultaneously, the molecular fingerprint of the drug and the physicochemical features of the target are integrated to obtain key features influencing drug-target interactions. Ultimately, although the benchmark dataset we used is only one-third the size of the benchmark dataset used by the compared state-of-the-art methods, MM-IDTarget demonstrates performance close to or better than the compared state-of-the-art methods on all Top-K evaluation metrics on the same target identification test dataset. Furthermore, we comprehensively tested the application capabilities of MM-IDTarget on generalization test datasets and in real-world application cases, finding that MM-IDTarget still exhibits good target identification performance, thus demonstrating that MM-IDTarget is an effective and reliable target identification tool.

[0100] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0101] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments 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, Includes the following steps: S1: Obtain the drug-target interaction Chemogenomic-Model dataset; S2: Obtain the target's structure file, and use the target's structure file to obtain the target's structural feature representation and edge feature representation; S3: Obtain sequence tag encoding feature representations, graph structure feature representations, and physicochemical feature representations of drugs and targets; S4: The sequence label encoding feature representation of the drug is fed into a three-layer multi-scale convolutional neural network with a stride of 1 and padding of 0 for each layer. The kernel sizes are 3x3, 5x5 and 7x7 respectively. A normalization operation is followed by each convolutional layer to obtain the three-layer features of the drug in the multi-scale convolutional neural network. These features are then concatenated and a linear transformation is performed to generate the sequence label encoding feature. S5: The structural feature representation of the drug graph is fed into a three-layer graph converter framework to extract the structural modal features of the drug. Each layer of the graph converter has 8 heads, each hidden layer has 128 neurons, each hidden layer is followed by a normalization operation, and finally a max pooling operation is connected. S6: The target image structural feature representation is fed into the Residual EW-GCN graph convolutional network. The Residual EW-GCN contains three GCN layers and one fully connected network. First, the target image 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: The drug-target cross-attention fusion feature representation is obtained by fusing the sequence tag encoding feature representation and graph structure feature representation of the drug with cross-attention fusion, and then fusing the fusion feature representation of the drug and the target. S8: The drug-target cross-attention fusion feature representation and the drug physicochemical feature representation are concatenated and used as a comprehensive feature of drug-target interaction to be fed into the fully connected layer to predict drug-target interaction. S9: Compare the true labels with the predicted results, update the network parameters to minimize the cross-entropy loss function, until the maximum number of iterations is reached, training ends, and the trained MM-IDTarget model is obtained; S10: For the trained MM-IDTarget model, Top-K based recall is used to evaluate the model's drug target recognition performance; S11: Compare the true labels with the predicted results and minimize the cross-entropy loss function.

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

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

5. The drug target identification method 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 ranking list of interactions with 859 targets from the model; (2) If a drug molecule has the same interaction score with multiple targets, DeepPurpose is used to predict the interaction score between the drug molecule and the targets, and the obtained scores are sorted. (3) If the true target of a drug small molecule is within the range of the Top-K targets identified by the model, then the target identification of the drug small molecule is defined as correct. The Top-K based recall evaluation index is used to measure the proportion of drug small molecules that are correctly identified as targets in a given test drug small molecule dataset.

6. The drug target identification method 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 y pred This represents the true label and the predicted label.

Images