DTA prediction method and model based on multi-scale features
By combining the molecular structure characteristics and interaction network characteristics of drugs and targets in DTA prediction and performing feature adaptive fusion, the problem of lack of topological structure information and network topological information in the prior art is solved, and the accuracy of DTA prediction is significantly improved.
Patent Information
- Application Number
- CN202411713540.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-27
- Publication Date
- 2025-03-07
AI Technical Summary
When predicting the binding affinity of drugs to targets (DTA), the lack of topological information of molecules and topological information of drug target interaction networks leads to poor prediction results.
A DTA prediction method based on multi-scale features is proposed, and the molecular structure characteristics and network structure characteristics of drugs and targets are learned through graph neural networks, and composite features are obtained through feature adaptive fusion to improve the DTA prediction effect.
By combining the feature information of the molecular structure scale and the interaction network scale, the accuracy of DTA prediction is significantly improved, which is better than existing methods, and the importance of multi-scale feature fusion is verified through ablation experiments.
Smart Images
Figure CN119580824A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of medical data processing, and in particular relates to a DTA prediction method and model based on multi-scale features. Background Art
[0002] The development of new drugs is a long and expensive process. According to statistics, it takes an average of 10-13 years and costs $600-2.6 billion to develop a new drug. This has prompted researchers to turn to computational methods to reduce time and costs. A recent report showed that most experimental failures in clinical trials are caused by inaccurate validation of drug targets, which shows that accurate prediction of drug-target binding affinity (DTA) is crucial to drug development.
[0003] In recent years, artificial intelligence, as a technology that uses deep learning and machine learning, has achieved satisfactory results in the task of drug-target affinity prediction. These methods can be roughly divided into structure-based methods and network-based methods. The structure-based method inputs the drug and target into the constructed encoder to extract their respective abstract features, and then sends the features to the multi-layer perceptron (MLP) to calculate the prediction results. DeepDTA proposed by Hakime et al. uses the sequence information of drugs and targets based on convolutional neural networks (CNN) to predict DTA and achieved relatively satisfactory results. However, one-dimensional sequence information lacks molecular topological structure information, which may impair the prediction ability of the model. Graph neural networks (GNNs) are used to solve this problem because the two-dimensional representation of molecules can be naturally converted into graph structure data. Tsubaki et al. proposed using GNN and CNN to learn the two-dimensional graph structure of molecules and the amino acid sequence of targets respectively. The experimental results on three data sets are significantly improved compared with sequence-based methods. Molecular structure-based methods do not consider the topological information of the drug-target interaction network, which is very important for DTA prediction.
[0004] Network-based methods explore potential interaction patterns by modeling complex drug-target relationship networks. This type of method not only takes into account the inherent characteristics of drugs and targets, but also infers potential affinity through their structural position and connection relationship in the network. Zhao et al. predicted unknown drug-target interaction information based on the constructed network by considering drugs and targets as nodes in the network and their interactions as edges. The molecular properties of drug targets and network topology information can be regarded as information of different scales. However, how to effectively capture and integrate this information remains to be solved.
[0005] Based on this, the present invention proposes a new method and model for predicting drug target affinity, which is based on multi-scale features and adaptive fusion. Summary of the invention
[0006] The first inventive object of the present invention is to provide a DTA prediction method based on multi-scale features, which combines the multi-scale methods of the molecular structure scale of drugs and targets and the interaction network scale to capture feature information, and then obtains composite features through feature adaptive fusion, thereby improving the DTA prediction effect.
[0007] In order to achieve the above purpose, the technical solution adopted is:
[0008] A DTA prediction method based on multi-scale features comprises the following steps:
[0009] After modeling drugs and targets as undirected graphs, the molecular scale features of drugs and targets are obtained through graph neural network learning;
[0010] For the bipartite network of the drug and the target, learn from the network scale to obtain the network scale characteristics of the drug and the target;
[0011] After dynamically fusing the molecular scale characteristics and network scale characteristics of the drug, and the molecular scale characteristics and network scale characteristics of the target, a prediction result is output.
[0012] Furthermore, the learning process of the molecular scale features of the drug is as follows: after initializing the atomic features and edge features of the drug, the embedding vectors of atoms and bonds are learned by using a graph neural network in a neighborhood aggregation manner, and then the molecular representation of the drug is obtained by reading out the function, which is the embedding of the drug, i.e., the molecular scale features of the drug;
[0013] The polymerization operation adopts the following formula:
[0014]
[0015] In the formula, represents the feature vector of atom i in layer l+1, σ is the activation function, is the set of neighboring nodes of atom i, is the learnable parameter matrix of the lth layer of GNN, c i and c j is the normalization constant.
[0016] The learning process of the molecular scale features of the target is as follows: after initializing the node features of the target, learning is performed by means of neighborhood aggregation through a graph neural network, and then the molecular representation of the target is obtained by a readout function, which is the embedding of the target, i.e., the molecular scale features of the target;
[0017] The polymerization operation adopts the following formula:
[0018]
[0019] In the formula, represents the feature vector of amino acid residue a in the l+1 layer, σ is the activation function, is the set of neighboring nodes of residue a, is the learnable parameter matrix of the lth layer of GNN, c a and c b is the normalization constant.
[0020] Furthermore, in the process of learning the molecular scale features of the drug, the atomic features include atomic type, atomic degree, implicit valence, and aromatic ring; the edge features include bond type, conjugation, and ring;
[0021] In the process of learning the molecular scale features of the target, the node features include disability type, whether it is aromatic hydrocarbon, position specific score and residue weight.
[0022] Furthermore, the learning process of the network scale characteristics of the drug and target is as follows: for the constructed drug target bipartite network, after extracting network information through GNN, the graph signal is filtered in the spectral domain to obtain the potential feature embedding of the drug and target, that is, the network scale characteristics of the drug and target;
[0023] Among them, the network node features are updated by the following formula in extracting network scale features:
[0024] F (l) =σ(Lθ (l) σ(…σ(Lθ 1 F)…));
[0025] In the formula, σ is the activation function, is the normalized Laplace matrix, θ (l) is the learnable parameter of the lth layer of GNN.
[0026] Furthermore, the dynamic fusion process is as follows:
[0027] (1) The molecular scale characteristics and network scale characteristics of the drug, the molecular scale characteristics and network scale characteristics of the target are connected into a multi-scale representation to obtain a multi-scale representation of the drug and the target, using the following formula:
[0028]
[0029] Where, d m ,t m Represent the molecular scale characteristics of drugs and targets, d n ,t n Represents the network-scale characteristics of drugs and targets, Represents a splicing operation;
[0030] (2) Calculate the weighted feature attention matrix of drugs and targets respectively Use the following formula:
[0031]
[0032] Where W1 and W2 are learnable parameter matrices;
[0033] (3) The weighted feature attention matrix of the drug and target Multiplying with the multi-scale representation of the drug and target, the final joint representation is obtained, that is, the fusion embedding representation of the drug and target, which adopts the following formula:
[0034]
[0035] Furthermore, the process of outputting the prediction result is: after the fusion embedding representation connection of the drug and the target, it is sent to MLP to output the predicted affinity result.
[0036] Furthermore, in the process of outputting the prediction results, the root mean square error is used as the loss function, which adopts the following formula:
[0037]
[0038] In the formula, n is the sample size, P i is the prediction result, Y i is the real affinity value.
[0039] The second inventive object of the present invention is to provide a DTA prediction model based on multi-scale features, which can implement the DTA prediction method mentioned in the claims and has a better DTA prediction effect.
[0040] In order to achieve the above purpose, the technical solution adopted is:
[0041] A DTA prediction model based on multi-scale features can implement the above-mentioned DTA prediction method, which includes:
[0042] Drug molecule graph feature extraction module: Encode drug molecule graph into drug embedding through graph neural network;
[0043] Drug and target network feature learning module: learn the embedding of drugs and targets from the network scale;
[0044] Target molecular graph feature extraction module: Use GNN to learn target embedding of target molecular graph;
[0045] Feature fusion and prediction module: The learned molecular structure features and network information features are dynamically fused and then sent to the prediction module to output the prediction results.
[0046] A third inventive object of the present invention is to provide a computer device for DTA prediction with better accuracy and adaptability.
[0047] In order to achieve the above purpose, the technical solution adopted is:
[0048] A computer device includes a processor and a memory, wherein the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement the above-mentioned DTA prediction method.
[0049] A fourth object of the present invention is to provide a computer-readable storage medium that can be used for DTA prediction.
[0050] In order to achieve the above purpose, the technical solution adopted is:
[0051] A computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement the above-mentioned DTA prediction method.
[0052] Compared with the prior art, the present invention has the following beneficial effects:
[0053] The present invention proposes a DTA prediction method and model based on multi-scale features, which is a method and model (MSFDTA) that combines the molecular structure scale and the drug-target interaction network scale, and has the following advantages:
[0054] 1. In the technical solution of the present invention, a more comprehensive and sufficient feature representation is obtained by learning the molecular-level embedding of drug molecule graphs and target molecule graphs, and learning the potential associations of the drug-target interaction network.
[0055] 2. In the technical solution of the present invention, the feature fusion module features attention dynamically adjusts the weights of different modalities to perform multimodal fusion.
[0056] The experimental results of the present invention on different data sets show that MSFDTA can effectively improve the prediction accuracy of DTA, and MSFDTA can be used as a favorable tool for DTA prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0057] Figure 1 This is the overall framework diagram of MSFDTA. DETAILED DESCRIPTION
[0058] In order to further explain the DTA prediction method and model based on multi-scale features of the present invention and achieve the intended purpose of the invention, the following is a detailed description of the DTA prediction method and model based on multi-scale features proposed in accordance with the present invention, its specific implementation, structure, features and efficacy, in combination with the preferred embodiments. In the following description, different "one embodiment" or "embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.
[0059] The following will further introduce in detail a DTA prediction method and model based on multi-scale features of the present invention in conjunction with specific embodiments:
[0060] Accurate prediction of drug-target affinity (DTA) is crucial for drug development and drug reuse. However, most deep learning-based methods only make predictions based on the molecular structure information of drugs or targets or the network information of their interactions. This may ignore the specificity of features at different scales, thereby reducing the effectiveness of the experiment. The present invention proposes a DTA prediction method and model based on multi-scale features, which is a DTA prediction model MSFDTA that integrates information at different scales of drug targets. The framework consists of a molecular scale feature extraction module, a network scale feature extraction module, and a feature fusion and prediction module. The molecular scale features and the bipartite network interaction features of drug targets are obtained respectively through a graph neural network, and finally the features processed by the feature fusion module are sent to the DTA prediction module for DTA prediction. The experimental results of this experiment on two benchmark data sets show that there is a significant improvement compared with the existing methods. The proposed MSFDTA is superior to the most advanced methods. The ablation experiment proves that the combination of multi-scale feature information has a significant improvement in the DTA prediction performance. The combination of different scale features and feature fusion modules are indispensable for improving model performance. The specific embodiments are as follows:
[0061] Example 1.
[0062] The specific steps are as follows:
[0063] A Method
[0064] 1. Basic theory and methods
[0065] In this example, the DTA prediction task is regarded as a regression task, and then the pre-defined in the experiment is given. The DTA prediction task, drug-target bipartite network, drug molecule graph and protein target molecule graph are defined as follows:
[0066] (1) DTA prediction. Given a drug set D and a target set T, the drug-target affinity value matrix can be defined as A∈R| D | ×|T|, the purpose of this embodiment is to find a matrix To approximate A.
[0067] (2) Drug-target bipartite network. Define B = (D, T, E) as the drug-target bipartite network, where E represents the affinity set between the drug and the target, and the corresponding adjacency matrix G A ∈R (|D|+|T,|D|+|T) It is defined as shown in formula (1).
[0068]
[0069] (3) Drug molecule graph. A given drug molecule graph can be represented as G d =(V,E), where represents the atomic set, E={e ij |(v i ,v j )} represents the set of edges between atoms.
[0070] (4) Target molecular graph. Considering the amino acids of the target as nodes, the molecular graph of the target can be expressed as G t =(N,C), where N is a node set consisting of amino acid residues, and C = {c ab |(n a ,n b )} indicates that there are contacts between the amino acid residues.
[0071] 2. Overall framework
[0072] Figure 1 The overall framework of MSFDTA is given. Among them, a is the drug molecule graph feature extraction module, which encodes the drug molecule graph into drug embedding through GNN. b is the drug target network feature learning module, which learns the embedding of drugs and targets from the network scale. c is the target molecule graph feature extraction module, which also uses GNN to learn the target embedding of the target molecule graph. d is the feature fusion and prediction module, which dynamically fuses the learned molecular structure features and network information features, and then sends them to the prediction module to output the prediction results.
[0073] Combination Figure 1 , a DTA prediction method based on multi-scale features, comprising the following steps:
[0074] After modeling drugs and targets as undirected graphs, the molecular scale features of drugs and targets are obtained through graph neural network learning;
[0075] For the bipartite network of drugs and targets, learn from the network scale to obtain the network scale characteristics of drugs and targets;
[0076] After dynamically fusing the molecular scale characteristics and network scale characteristics of the drug, and the molecular scale characteristics and network scale characteristics of the target, the prediction results are output.
[0077] A DTA prediction model based on multi-scale features can implement the above-mentioned DTA prediction method, which includes:
[0078] Drug molecule graph feature extraction module: Encode drug molecule graph into drug embedding through graph neural network;
[0079] Drug and target network feature learning module: learn the embedding of drugs and targets from the network scale;
[0080] Target molecular graph feature extraction module: Use GNN to learn target embedding of target molecular graph;
[0081] Feature fusion and prediction module: The learned molecular structure features and network information features are dynamically fused and then sent to the prediction module to output the prediction results.
[0082] The key points of the DTA prediction method and model in this embodiment are as follows:
[0083] (1) Learning of molecular structure scale features
[0084] After modeling drugs and targets as undirected graphs, the molecular graphs are encoded using graph neural networks to obtain molecular representations of drugs and targets, which are embeddings of drugs and targets, i.e., molecular scale features of drugs and targets. Specifically:
[0085] For a given drug G d =(V,E), this embodiment can initialize the atomic features, including atomic type, atomic degree, implicit valence, aromatic ring, etc. Edge feature initialization includes bond type, conjugation, ring, etc. Then use GNN to learn the embedding vectors of atoms and bonds in a neighborhood aggregation manner, and finally obtain the molecular representation of the drug by reading out the function, that is, the molecular scale feature d of the drug m The feature aggregation operation can be defined as shown in formula (2):
[0086]
[0087] in, represents the feature vector of atom i in layer l+1, σ is the activation function, is the set of neighboring nodes of atom i, is the learnable parameter matrix of the lth layer of GNN, c i and c j is the normalization constant.
[0088] The operation for the target molecule graph is similar. Given a target molecule graph Gt =(N,C), the initial features of the amino acid residues, i.e. the nodes, include disability type, whether it is aromatic, position-specific score and residue weight, etc. Then GNN is used to learn the embedding vector in a neighborhood aggregation manner, and finally the molecular representation of the target is obtained through the readout function, i.e. the molecular scale feature t of the target m The aggregation definition of the node features of the target molecule is shown in formula (3):
[0089]
[0090] in, represents the feature vector of amino acid residue a in the l+1 layer, σ is the activation function, is the set of neighboring nodes of residue a, is the learnable parameter matrix of the lth layer of GNN, c a and c b is the normalization constant.
[0091] (2) Network-scale feature learning
[0092] In the drug and target network feature learning module, the present invention uses GNN to extract network information from the constructed drug target bipartite network, and then obtains the potential feature embedding of drugs and targets by filtering the graph signal in the spectral domain, that is, the network scale feature d of drugs and targets. n ,t n .
[0093] Let G A ∈R (|D|+|T|,|D|+|T|) and Represent the adjacency matrix of the drug and target bipartite network and the initial feature matrix of the drug and target, respectively, where F∈R (|D|+|T,1) , F d is the initial feature matrix of the drug, F t is the initial feature matrix of the target. In order to extract the network scale features, the network node features are updated by the method shown in formula (4).
[0094] F (l) =σ(Lθ (l0 σ(…σ(Lθ 1 F)…))(4)
[0095] Among them, σ is the activation function, is the normalized Laplace matrix, θ (l) is the learnable parameter of the lth layer of GNN. The high-order adjacent node features of the aggregation nodes are iteratively learned through formula (4) to obtain the embedding vectors of drugs and targets, that is, the network scale features d of drugs and targets n ,t n .
[0096] (3) Feature fusion and prediction module
[0097] After obtaining the molecular-scale features and network-scale features of drug targets, in order to measure the importance of different modalities, the present invention designs a feature-based attention module to dynamically adjust the weight of each modality to achieve multimodal fusion.
[0098] First, the different modal features of drugs and targets are concatenated into a multi-scale representation, Here m ,t m Represent the molecular scale characteristics of drugs and targets, respectively, n ,t n Respectively represent the network scale characteristics of drugs and targets, Represents the concatenation operation. Then the weighted feature attention matrix of the drug is calculated separately Weighted feature attention matrix of the target The calculation method is shown in formula (5)-(6):
[0099]
[0100] Among them, W1 and W2 are learnable parameter matrices. In addition, the value of the attention matrix is constrained between 0 and 1 through the Sigmoid function.
[0101] Afterwards, the attention matrix is multiplied with the original input features (i.e., the multi-scale representation of drugs and targets) to obtain the final joint representation, i.e., the fused embedding representation of drugs and targets, as shown in formulas (7)-(8):
[0102]
[0103] Finally, after obtaining the fusion embedding representation of the drug and the target, the two representations are concatenated column by column to obtain the joint representation and then sent to the MLP to output the predicted affinity result. Since DTA is a regression task, this embodiment selects the root mean square error (MSE) as the loss function, as shown in formula (9):
[0104]
[0105] Where n is the sample size, P i is the prediction result, Y i is the real affinity value.
[0106] B. Experimental Setup and Results
[0107] 1. Dataset
[0108] In order to comprehensively evaluate the effectiveness of the proposed method, this example uses two public benchmark datasets, the Davis dataset and the KIBA dataset. The Davis dataset consists of 68 drugs and 442 protein kinases, and the KIBA dataset consists of 2111 drugs and 229 proteins. Detailed information about the dataset is listed in Table 1.
[0109] Table 1 Statistics of the used datasets
[0110]
[0111] 2. Evaluation strategies and indicators
[0112] This example uses five-fold cross validation to evaluate the data set. Each data set is evenly divided into six subsets of equal size, five of which are used for model training and the remaining subsets are used for testing. As an evaluation indicator.
[0113] 3. Experimental results
[0114] (1) Experimental results on Davis and KIBA datasets
[0115] The proposed MSFDTA is compared with five baseline methods: DeepDAT, GrapDTA, DGraphDTA, MATT_DTI and FusionDTA. The experimental results are shown in Table 2. It can be seen that MSFDTA is generally better than other baseline models. On the Dvais dataset, it improves the MSE by 0.039 compared with the second-best FusionDTA. It also improved by 0.033. In particular, the performance of DeepDTA, which only uses sequence information as input, and GraphDTA, which uses drug molecular graph information, has been significantly improved. The MSFDTA proposed in this invention also achieved the best performance on the KIBA dataset. In general, compared with the method that only uses a single scale, MSFDTA comprehensively utilizes feature information at different scales to fully characterize drugs and targets, which shows that combining molecular structure scale information and interaction network information can effectively improve the DTA prediction accuracy.
[0116] Table 2 Experimental results of MSFDTA and baseline methods on Davis and KIBA datasets
[0117]
[0118] 4. Ablation experiment
[0119] In order to verify the influence of each module in MSFDTA on the experimental results, this embodiment designs the following MSFDTA variants to conduct ablation experiments on the Davis dataset, and the experimental settings remain unchanged.
[0120] MSFDTA_w / o_M: Delete the molecular structure scale encoding module in MSFDTA.
[0121] MSFDTA_w / o_N: Remove the drug-target interaction network scale encoding module in MSFDTA.
[0122] MSFDTA_w / o_F: Delete the feature fusion encoding module in MSFDTA.
[0123] The experimental results are shown in Table 3. It can be seen that using only the molecular structure scale MSFDTA_w / o_N and using the network scale MSFDTA_w / o_M will lead to a decrease in prediction performance. In particular, the performance of MSFDTA_w / o_N is greatly affected, with an MSE of only 0.235. The ablation experiment results show that different components in MSFDTA have a positive impact on DTA prediction.
[0124] Table 3 Ablation experiment results of MSFDTA on Davis dataset
[0125]
[0126] The above is only a preferred embodiment of the embodiment of the present invention, and does not impose any form of limitation on the embodiment of the present invention. Any simple modification, equivalent changes and modifications made to the above embodiment based on the technical essence of the embodiment of the present invention are still within the scope of the technical solution of the embodiment of the present invention.
Claims
1. A DTA prediction method based on multi-scale features, characterized in that: The following steps are involved: After modeling drugs and targets as undirected graphs, the molecular scale features of drugs and targets are obtained through graph neural network learning; For the bipartite network of the drug and the target, learn from the network scale to obtain the network scale characteristics of the drug and the target; After dynamically fusing the molecular scale characteristics and network scale characteristics of the drug, and the molecular scale characteristics and network scale characteristics of the target, a prediction result is output.
2. The DTA prediction method according to claim 1, characterized in that: The learning process of the molecular scale features of the drug is as follows: after initializing the atomic features and edge features of the drug, the embedding vectors of atoms and bonds are learned by using a graph neural network in a neighborhood aggregation manner, and then the molecular representation of the drug is obtained by reading out the function, which is the embedding of the drug, i.e., the molecular scale features of the drug; The polymerization operation adopts the following formula: In the formula, represents the eigenvector of atom i at level l+1, is the set of neighboring nodes of atom i, is the learnable parameter matrix of the lth layer of GNN, c i and c j is the normalization constant, σ is the activation function; The learning process of the molecular scale features of the target is as follows: after initializing the node features of the target, learning is performed by means of neighborhood aggregation through a graph neural network, and then the molecular representation of the target is obtained by a readout function, which is the embedding of the target, i.e., the molecular scale features of the target; The polymerization operation adopts the following formula: In the formula, represents the feature vector of amino acid residue a at the l+1th layer, is the set of neighboring nodes of residue a, is the learnable parameter matrix of the lth layer of GNN, c a and c b is the normalization constant and σ is the activation function.
3. The DTA prediction method according to claim 2, characterized in that: In the process of learning the molecular scale characteristics of the drug, the atomic characteristics include atomic type, atomic degree, implicit valence, and aromatic ring; Edge features include bond type, conjugation, and rings; In the process of learning the molecular scale features of the target, the node features include disability type, whether it is aromatic hydrocarbon, position specific score and residue weight.
4. The DTA prediction method according to claim 1, characterized in that: The learning process of the network scale characteristics of the drug and target is as follows: for the constructed drug target bipartite network, after extracting network information through GNN, the graph signal is filtered in the spectral domain to obtain the potential feature embedding of the drug and target, that is, the network scale characteristics of the drug and target; Among them, the network node features are updated by the following formula in extracting network scale features: F (l) =σ(Lθ (l) σ(…σ(Lθ 1 F)…)); In the formula, σ is the activation function, is the normalized Laplace matrix, θ (l is the learnable parameter of the lth layer of GNN.
5. The DTA prediction method according to claim 1, characterized in that: The dynamic fusion process is as follows: (1) The molecular scale characteristics and network scale characteristics of the drug, the molecular scale characteristics and network scale characteristics of the target are connected into a multi-scale representation to obtain a multi-scale representation of the drug and the target, using the following formula: Where, d m ,t m Represent the molecular scale characteristics of drugs and targets, d n ,t n Represents the network-scale characteristics of drugs and targets, Represents a splicing operation; (2) Calculate the weighted feature attention matrix of drugs and targets respectively Use the following formula: Where W1 and W2 are learnable parameter matrices; (3) The weighted feature attention matrix of the drug and target Multiplying with the multi-scale representation of the drug and target, the final joint representation is obtained, that is, the fusion embedding representation of the drug and target, which adopts the following formula:
6. The DTA prediction method according to claim 5, characterized in that: The process of outputting the prediction result is as follows: after the fusion embedding representation connection of the drug and the target is connected, it is sent to the MLP to output the predicted affinity result.
7. The DTA prediction method according to claim 6, characterized in that: In the process of outputting the prediction results, the root mean square error is used as the loss function, which adopts the following formula: In the formula, n is the sample size, P i is the prediction result, Y i is the real affinity value.
8. A DTA prediction model based on multi-scale features, characterized in that: The DTA prediction model can implement the DTA prediction method according to any one of claims 1 to 7, which comprises: Drug molecule graph feature extraction module: Encode drug molecule graph into drug embedding through graph neural network; Drug and target network feature learning module: learn the embedding of drugs and targets from the network scale; Target molecular graph feature extraction module: Use GNN to learn target embedding of target molecular graph; Feature fusion and prediction module: The learned molecular structure features and network information features are dynamically fused and then sent to the prediction module to output the prediction results.
9. A computer device, characterized in that: It comprises a processor and a memory, wherein the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement the DTA prediction method according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement the DTA prediction method according to any one of claims 1 to 7.