Drug-target binding affinity prediction method based on graph neural network

Abstract

The invention provides a drug-target binding affinity prediction method based on a graph neural network. The drug-target binding affinity prediction method comprises the following steps: (1) obtaining an SMILES sequence of a drug and an amino acid sequence of a target protein; (2) converting the SMILES sequence of the drug into a drug molecular map, and performing feature extraction on the drug molecular map through a map convolutional neural network to obtain a drug feature vector; (3) inputting the amino acid sequence of the target protein into a Transform encoder for feature extraction to obtain a protein feature vector; (4) fusing the drug feature vector and the protein feature vector to obtain a fused feature vector; and (5) inputting the fusion feature vector into a multi-layer perceptron, and outputting a drug-target binding affinity prediction result. According to the method, the hybrid model combining the GCN and the Transform is used for DTA prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics and computational drug design, and particularly to a drug-target binding affinity prediction method based on a graph neural network. BACKGROUND

[0002] The drug discovery process has the significant characteristics of high cost and long cycle. According to statistics, it usually costs 500 million to 2 billion US dollars to launch a new drug, and it takes 10 to 15 years. A drug (usually as a ligand) produces a specific physiological effect by interacting with a target protein, and its therapeutic effect depends largely on the binding strength between the two. Binding affinity is a core indicator to measure the strength of drug-target interaction, and the technology to predict this indicator through experiments or computational methods is called drug-target binding affinity (DTA) prediction. According to the predicted affinity value, researchers can screen candidate drugs with strong interaction, providing direction for subsequent wet experiment verification, thereby accelerating drug discovery, design and reuse. Obviously, binding affinity is a key indicator to evaluate the strength of drug-target interaction, and accurate prediction of such interaction itself is also a core link of drug research and development.

[0003] However, the traditional wet experiment to determine the binding affinity between the candidate drug and the protein target is not only costly and time-consuming, but sometimes the results lack reproducibility. In order to accelerate this process, virtual screening technology is widely used. At the same time, the rapid development of computer technology has opened up a broad space for interdisciplinary cooperation. In particular, with the development of deep learning technology, deep learning-based methods have been increasingly applied to various stages of drug development. Deep learning can automatically learn feature representations from raw input data without relying on a large amount of domain knowledge, which makes it show a significant advantage in dealing with complex biological data. Deep learning-based methods can significantly accelerate the drug screening process, thereby greatly reducing development costs and time. Drug molecules and protein targets are essentially complex biological entities, containing rich feature information. Currently, a variety of deep learning models have been applied to DTA prediction: DeepDTA is based on convolutional neural network (CNN), which uses one-dimensional convolution to extract features from the simplified molecular-input line-entry system (SMILES) sequence of the drug and the amino acid sequence of the protein, and predicts the affinity through a multi-layer perceptron; WideDTA model integrates protein sequence, ligand SMILES, protein domain and motif information, and maximum common substructure vocabulary and other information sources for prediction; GraphDTA model represents the drug molecule as a graph structure, and uses graph neural network to extract drug features for prediction.

[0004] Although these methods have achieved certain results, in terms of protein sequence feature extraction, traditional CNN and other methods often have difficulty in fully capturing the global dependence and context information of long sequences. In view of the limitations of existing DTA prediction methods, it is an urgent need to develop more accurate and efficient computational models. SUMMARY

[0005] The present application provides a method for predicting drug-target binding affinity based on graph neural network, which uses a hybrid model combining GCN and Transformer for DTA prediction. The model uses GCN to extract features from the drug molecule graph, uses Transformer to encode the protein amino acid sequence, and fuses the features through a fully connected neural network to predict the binding affinity.

[0006] To achieve the above purpose, the present application adopts the following technical solutions:

[0007] The method for predicting drug-target binding affinity based on graph neural network comprises:

[0008] (1) obtaining the SMILES sequence of the drug and the amino acid sequence of the target protein;

[0009] (2) converting the SMILES sequence of the drug into a drug molecule graph, performing feature extraction on the drug molecule graph through a graph convolutional neural network to obtain a drug feature vector;

[0010] (3) inputting the amino acid sequence of the target protein into a Transformer encoder to perform feature extraction to obtain a protein feature vector;

[0011] (4) fusing the drug feature vector and the protein feature vector to obtain a fused feature vector;

[0012] (5) inputting the fused feature vector into a multi-layer perceptron to output a drug-target binding affinity prediction result.

[0013] In the specification, in step (2), the drug molecule graph takes atoms as nodes and chemical bonds as edges, and each node and edge contains a corresponding feature vector.

[0014] In the specification, in step (2), the graph convolutional neural network includes three consecutive graph convolutional layers, each layer uses a ReLU activation function, and a global maximum pooling layer is connected after the third graph convolutional layer, and the drug feature vector is output through the global maximum pooling layer.

[0015] In the specification, in step (3), the Transformer encoder is stacked by 10 layers with the same structure, each layer includes a multi-head self-attention mechanism module and a feedforward neural network module, and the number of heads of the multi-head self-attention mechanism is 4.

[0016] In the specification, the multi-head self-attention mechanism calculates the correlation between any two amino acids in the protein amino acid sequence, generates query, key, and value vectors, calculates attention based on the three vectors, and outputs after splicing the attention results of multiple heads.

[0017] In the specification, in step (3), the output of the Transformer encoder is processed through a global average pooling layer to obtain the protein feature vector, and the protein feature vector is a fixed-length vector that integrates global information of the protein sequence.

[0018] In the specification, in step (4), the fusion is achieved by splicing the drug feature vector and the protein feature vector, and the fused feature vector is the spliced vector.

[0019] In this specification, in step (1), the SMILES sequence of the drug and the amino acid sequence of the target protein are derived from the Davis dataset or the KIBA dataset. The Davis dataset contains interaction data of 68 drugs and 442 target proteins, and the KIBA dataset contains interaction data of 2111 drugs and 229 proteins.

[0020] In this specification, the prediction results are evaluated using metrics including mean squared error, consistency index, and corrected coefficient of determination. The mean squared error measures the average squared deviation between the predicted and actual values. The consistency index measures the probability that the predicted and actual values ​​are in the same order. The corrected coefficient of determination is used to evaluate the model's external predictive power and robustness.

[0021] In this specification, the graph convolutional neural network, Transformer encoder, and multilayer perceptron are trained using the Adam optimizer, with 6-fold cross-validation employed during training and a learning rate of 0.0001.

[0022] In summary, the present invention has at least the following beneficial effects:

[0023] The GCN-Transformer dual-stream architecture model proposed in this invention demonstrates significant technical advantages in experimental results on the Davis and KIBA benchmark datasets. Specific results are as follows:

[0024] 1. Improved prediction accuracy:

[0025] On the Davis dataset, the mean squared error (MSE) is 0.245±0.003, which is 5.7% lower than DeepDTA (0.261) and 3.5% lower than GraphDTA (0.254), indicating that the predicted values ​​deviate less from the true values.

[0026] On the KIBA dataset, the MSE drops to 0.133±0.003, a significant reduction of 31.4% compared to DeepDTA (0.194) and 4.3% compared to GraphDTA (0.139), demonstrating high-precision prediction capability for complex and heterogeneous data.

[0027] 2. Enhanced affinity ranking ability: The consistency index (CI) reached 0.887±0.004 on the Davis dataset, which is higher than DeepDTA (0.878) and GraphDTA (0.880); and reached 0.891±0.004 on the KIBA dataset, which is higher than GraphDTA (0.889) and DeepDTA (0.863), verifying the reliability of the model in prioritizing candidate drugs.

[0028] 3. Generalization and robustness improvement: external prediction capability metrics On Davis dataset, it is 0.6705±0.008, which is 6.4% higher than DeepDTA (0.630); on KIBA dataset, it is 0.6885±0.008, which is 2.3% higher than DeepDTA (0.673), proving that the model has strong generalization ability and stability on different datasets.

[0029] 4. Advantage of collaborative modeling: through GCN to capture the topological structure of drug molecules, and through Transformer to capture the long-range dependence of proteins, the collaborative modeling of drug spatial structure and protein global features is realized, breaking through the single feature extraction limitation of existing models. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0031] Figure 1 The schematic diagram of the drug-target binding affinity prediction method based on the graph neural network involved in the present application.

[0032] Figure 2 The model architecture schematic diagram of the DTA model involved in the present application. DETAILED DESCRIPTION

[0033] In the following, only some exemplary embodiments are simply described. As those skilled in the art can recognize, the described embodiments can be modified in various different ways without departing from the spirit or scope of the embodiments of the present application. Therefore, the drawings and the description are considered to be exemplary in nature rather than limiting.

[0034] The following disclosure provides many different embodiments or examples for implementing different structures of the embodiments of the present application. In order to simplify the disclosure of the embodiments of the present application, the components and settings of specific examples are described in the following. Of course, they are only examples, and the purpose is not to limit the embodiments of the present application. In addition, the embodiments of the present application can repeatedly refer to numerals and / or reference letters in different examples. Such repetition is for the purpose of simplification and clarity, and it does not indicate the relationship between the various embodiments and / or settings discussed.

[0035] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0036] As Figure 1As shown, the present embodiment provides a drug-target binding affinity prediction method based on a graph neural network, comprising:

[0037] (1) obtaining a SMILES sequence of a drug and an amino acid sequence of a target protein;

[0038] (2) converting the SMILES sequence of the drug into a drug molecule graph, extracting features of the drug molecule graph through a graph convolutional neural network, and obtaining a drug feature vector;

[0039] (3) inputting the amino acid sequence of the target protein into a Transformer encoder for feature extraction, and obtaining a protein feature vector;

[0040] (4) fusing the drug feature vector and the protein feature vector to obtain a fused feature vector;

[0041] (5) inputting the fused feature vector into a multi-layer perceptron to output a drug-target binding affinity prediction result.

[0042] In some embodiments, in step (2), the drug molecule graph takes atoms as nodes and chemical bonds as edges, and each node and edge contains a corresponding feature vector.

[0043] In some embodiments, in step (2), the graph convolutional neural network includes three consecutive graph convolutional layers, each layer uses a ReLU activation function, and a global maximum pooling layer is connected after the third graph convolutional layer, and the drug feature vector is output through the global maximum pooling layer.

[0044] In some embodiments, in step (3), the Transformer encoder is stacked by 10 layers with the same structure, each layer includes a multi-head self-attention mechanism module and a feedforward neural network module, and the number of heads of the multi-head self-attention mechanism is 4.

[0045] In some embodiments, the multi-head self-attention mechanism calculates the correlation between any two amino acids in the protein amino acid sequence, generates query, key, and value vectors, calculates attention based on the three vectors, and outputs after splicing the attention results of multiple heads.

[0046] In some embodiments, in step (3), the output of the Transformer encoder is processed by a global average pooling layer to obtain the protein feature vector, and the protein feature vector is a fixed-length vector that integrates global information of the protein sequence.

[0047] In some embodiments, in step (4), fusion is achieved by splicing the drug feature vector and the protein feature vector, and the fused feature vector is the spliced vector.

[0048] In some embodiments, in step (1), the SMILES sequence of the drug and the amino acid sequence of the target protein are derived from the Davis dataset or the KIBA dataset, wherein the Davis dataset contains interaction data of 68 drugs and 442 target proteins, and the KIBA dataset contains interaction data of 2111 drugs and 229 proteins.

[0049] In some embodiments, the prediction result is evaluated, and the evaluation indicators include mean squared error, concordance index and corrected determination coefficient, the mean squared error is used to measure the average squared deviation of the predicted value and the true value, the concordance index is used to measure the order consistency probability of the predicted value and the true value, and the corrected determination coefficient is used to evaluate the external prediction ability and robustness of the model.

[0050] In some embodiments, the graph convolutional neural network, the Transformer encoder and the multilayer perception are trained by an Adam optimizer, the training process adopts 6-fold cross-validation, and the learning rate is set to 0.0001.

[0051] The technical concept of the present application is as follows:

[0052] Drug-target binding affinity (DTA) prediction plays a key role in accelerating drug discovery and repositioning research. Deep learning methods have been widely applied in the field of DTA prediction, but the existing models still have limitations in modeling the long-range dependencies of proteins. Therefore, the present application proposes a dual-flow architecture model based on graph convolutional network (GCN) and Transformer. The model extracts the topological structure information of the drug molecule through GCN, captures the long-range dependencies and global context features of the protein sequence by using a deep Transformer encoder, splices the drug and protein representation vectors through a feature fusion module, and inputs a multilayer perception to realize end-to-end binding affinity prediction. Experiments on Davis and KIBA benchmark datasets show that, compared with baseline methods, the present application has significantly improved in key indicators such as mean squared error (MSE), concordance index (CI) and corrected determination coefficient (R2). The results verify the advantages of GCN-Transformer combination in modeling the spatial structure of drugs and the global features of proteins, and provide a robust computational framework for high-precision DTA prediction.

[0053] 1. Overall framework of the model

[0054] The DTA model proposed by the present application aims to predict the drug-target binding affinity score. As shown in Figure 2 The model architecture includes three core modules: data input module, drug and protein feature learning module, and affinity prediction module.

[0055] 1.1 Drug representation learning

[0056] Drug representation learning is based on GCN. Each drug is encoded as a graph, and a GCN model is used to learn the graph representation of the drug. The drug molecule graph is represented as a combination of nodes (atoms) and edges (chemical bonds), each node and edge has its corresponding feature vector. The drug graph is represented as G=(V, E), where V represents the set of atomic nodes, and E represents the set of chemical bond edges. The graph convolutional neural network of each layer collects and updates information. Each atom x i Collects local information from adjacent atoms and chemical bonds, and then updates the atom x i Node information:

[0057]

[0058] where W1, W2 are learnable weight matrices, N(i) is the set of adjacent nodes of node i, and σ represents the activation function ReLU. This process processes and updates the feature information of the atomic nodes at each iteration.

[0059] The GCN module designed by the present application includes three consecutive GCN layers, each of which uses a ReLU activation function. After the third GCN layer, a global max pooling layer is added, and finally a fixed-dimensional graph vector representation of each drug molecule is obtained

[0060] 1.2 Protein representation learning

[0061] The Transformer model has achieved great breakthroughs in the field of natural language processing. Its core multi-head self-attention mechanism can efficiently capture long-range dependencies and global context information between elements in the input sequence, breaking through the limitations of traditional sequence modeling methods such as RNN and CNN. Therefore, the present application uses a Transformer encoder to extract features from protein amino acid sequences.

[0062] The Transformer encoder is constructed by stacking multiple layers of the same structure, each layer containing two core sub-modules, namely the multi-head self-attention mechanism and the feedforward neural network. In the present application, the number of encoder layers is set to 10, and the number of heads of the multi-head attention module is set to 4.

[0063] Multi-head self-attention mechanism enables the model to focus on different positions in the sequence simultaneously, and can automatically identify functional modules and domains in the protein sequence by calculating the correlation between any two amino acids in the sequence. With the help of multi-head attention mechanism, the model's ability to capture global information can be enhanced. The multi-head attention layer contains three different vectors generated based on input transformation, which are generated by linear layers according to the input sequence. The self-attention mechanism enables the model to focus on different parts of the sequence, and its calculation method is shown in equation (2):

[0064]

[0065] where Q, K, V are the Query, Key and Value matrices, respectively, and d k is the dimension of the key vector. The multi-head self-attention mechanism divides the input into multiple heads, each of which independently calculates attention, and then splices the results, as shown in equation (3):

[0066] Multihead(Q,K,V)=Concat(head1,...,head h )W O ; (3)

[0067] where and W O are learnable parameter matrices.

[0068] Finally, the variable-length protein sequence output by the Transformer encoder is converted into a fixed-length feature vector by a global average pooling layer, which integrates the information of the entire protein sequence.

[0069] 1.3 Prediction output

[0070] The learned drug feature vector and the protein feature vector are spliced and then input into a multi-layer perceptron (MLP) to calculate the final binding affinity score, as shown in equation (4):

[0071]

[0072] 2 Experimental results

[0073] 2.1 Dataset

[0074] Two widely used standard datasets in the field of DTA prediction were adopted in this invention: Davis dataset and KIBA dataset. These two datasets were chosen because they comprehensively cover kinase inhibitors and their targets, and can capture different aspects of drug-target interactions. The statistical information of the two datasets is shown in Table 1.

[0075] Table 1. Experimental datasets

[0076] Datasets Drugs Proteins Interactions Davis 68 442 30,056 KIBA 2,111 229 118,254

[0077] Davis dataset: derived from a comprehensive analysis of kinase inhibitor selectivity by Davis et al. This dataset contains 68 drugs and 442 target proteins, covering more than 80% of the human catalytic protein kinase group, a total of 30,056 pairs of drug-protein interactions. Each data point contains the SMILES representation of the drug and the amino acid sequence of the target protein.

[0078] KIBA dataset: contains 118,254 pairs of drug-protein interactions involving 2,111 drugs and 229 proteins. Each data point contains the SMILES representation of the drug, the amino acid sequence of the protein, and the KIBA score as an indicator of binding activity. The KIBA score is calculated by a statistical-based ensemble method that integrates different types of biological activity data, including dissociation constant (K d ), inhibition constant (K i ) and half maximal inhibitory concentration (IC50).

[0079] 2.2 Evaluation metrics

[0080] This invention considers DTA prediction as a regression task, and uses the following metrics to evaluate model performance:

[0081] MSE: measures the average squared deviation between predicted and true values, which is a standard metric for regression tasks. The smaller the MSE value, the more accurate the prediction. The calculation formula is shown in (5):

[0082]

[0083] where output i and label i are the predicted and true values of the i-th sample, respectively, and n is the total number of samples.

[0084] CI: measures the probability that the predicted and true values of two random drug-target pairs are in the same order. The closer the CI value is to 1, the better the performance of the model.

[0085]

[0086] where, pi is the predicted value with higher affinity, p j is the predicted value with lower affinity, Z is a normalization constant, and h(x) is a step function.

[0087] Exponent: used to evaluate the external prediction ability and robustness of the regression model. Its calculation is based on the coefficient of determination r 2 and the coefficient of determination by origin regression

[0088]

[0089] It is generally considered that if r 2 > 0.5, the model reaches an acceptable performance standard. and close to r 2 indicates that the model has good prediction ability and robustness.

[0090] 2.3 Baseline model

[0091] To verify the effectiveness of the model proposed in the present application, we compared it with the current most advanced DTA prediction baseline model on the Davis and KIBA datasets, including: DeepDTA, WideDTA and GraphDTA.

[0092] 2.4 Model training and evaluation

[0093] The experimental environment is based on Python 3.9 and PyTorch 1.10 framework, using NVIDIA RTX4090 GPU for accelerated training. The Adam optimizer is used for model optimization, and the learning rate is set to 0.0001. To reduce the randomness caused by data division and fully evaluate the performance of the model, 6-fold cross-validation is used to train and validate the dataset. The final experimental results are the average values of the indicators calculated on the test set of 5 independent runs (standard deviation in parentheses).

[0094] Table 2. Performance comparison of Davis dataset

[0095]

[0096]

[0097] Table 3. Performance comparison of KIBA dataset

[0098]

[0099] As shown in Table 2, the model exhibits significant advantages on the Davis dataset. The precision is improved: the MSE is 0.245±0.003, which is reduced by 5.7% compared with DeepDTA (0.261) and by 3.5% compared with GraphDTA (0.254), indicating that the predicted value has smaller deviation from the true value. The CI index reaches 0.887±0.004, which is higher than all baselines (DeepDTA: 0.878, GraphDTA: 0.880), verifying the superiority of the model in affinity ranking ability. The index is 0.6705±0.008, which is increased by 6.4% compared with DeepDTA (0.630), proving that the model has stronger generalization ability and stability on the external test set.

[0100] The Davis dataset focuses on the selectivity of kinase inhibitors, and the global context of protein sequences (such as the kinase active site) is crucial to the binding mechanism. The Transformer module captures long-range residue dependencies through self-attention mechanisms, making up for the lack of global feature extraction in traditional CNNs (DeepDTA) and single graph networks (GraphDTA), thereby improving prediction accuracy.

[0101] As shown in Table 3, the model of the present application also performs better on the KIBA dataset. The MSE is reduced to 0.133±0.003, which is significantly reduced by 31.4% compared with DeepDTA (0.194) and further reduced by 4.3% compared with current GraphDTA (0.139). The CI index is improved to 0.891±0.004, which is higher than GraphDTA (0.889) and DeepDTA (0.863), highlighting the reliability of the model in prioritizing candidate drugs in high-throughput virtual screening. The index is increased by 2.3% compared with DeepDTA (0.673), maintaining the best robustness in complex heterogeneous data.

[0102] 3CONCLUSION

[0103] The present application proposes a hybrid model combining graph convolutional neural network (GCN) and Transformer encoder for predicting drug-target binding affinity (DTA). The model represents the drug molecule as a graph structure, extracts its topological and atomic features using a three-layer GCN; at the same time, the protein is regarded as an amino acid sequence, and a deep Transformer encoder is used to capture its long-range relationships and global information; finally, the feature vectors of the drug and protein are spliced, and a multi-layer perceptron (MLP) is used to predict the final binding affinity score. Experimental results on two widely used benchmark datasets (Davis and KIBA) show that the model has better performance in MSE, CI and The key indicators are superior to multiple mainstream baseline models, proving the effectiveness and superiority of the model.

[0104] In summary, the model proposed by the application provides a reliable and efficient solution for the DTA prediction task, and is expected to contribute to accelerating the process of new drug discovery and development.

[0105] In summary, the model proposed by the application provides a reliable and efficient solution for the DTA prediction task, and is expected to contribute to accelerating the process of new drug discovery and development.

[0106] From the above detailed description, it can be seen that the application can achieve the above-mentioned purposes, and has met the requirements of the Patent Law.

[0107] Although the preferred embodiments of the application have been described, those skilled in the art can make further changes and modifications to the embodiments once they know the basic inventive concept. Therefore, the appended claims are intended to include the preferred embodiments and all changes and modifications falling within the scope of the application. The above description is only the preferred embodiment of the application and is not intended to limit the application. It should be noted that any modification, equivalent replacement and improvement made within the spirit and principle of the application shall be included in the protection scope of the application.

[0108] It should be noted that the above description of the process is only for example and illustration, and does not limit the scope of the present application. Those skilled in the art can make various modifications and changes to the process under the guidance of the present application. However, these modifications and changes are still within the scope of the present application.

[0109] The above has described the basic concept, and it is obvious that the above application disclosure is only as an example and does not constitute a limitation to the present application for those skilled in the art after reading this application. Although it is not explicitly stated here, those skilled in the art can make various modifications, improvements and modifications to the present application. Such modifications, improvements and modifications are suggested in the present application, so such modifications, improvements and modifications still belong to the spirit and scope of the exemplary embodiments of the present application.

[0110] Meanwhile, specific words are used in the present application to describe the embodiments of the present application. For example, "one embodiment", "an embodiment", and / or "some embodiments" means a certain feature, structure or property related to at least one embodiment of the present application. Therefore, it should be emphasized and noted that the "an embodiment" or "one embodiment" or "an alternative embodiment" mentioned in different places in the specification does not necessarily refer to the same embodiment. In addition, some features, structures or characteristics in one or more embodiments of the present application can be properly combined.

[0111] Moreover, as will be appreciated by persons skilled in the art, the present application is capable of being embodied with several alternatives now known or later developed in the art, including any newly discovered equivalents, which do not depart from the scope and spirit of the present application. Accordingly, it is not intended that the application be limited as described herein, but rather that the application be limited only as set forth in the claims. For example, although the implementation of various components described above can be embodied in hardware, it can also be implemented as a software only solution (e.g., a computer program product which can include a computer-readable medium having stored computer program code thereon). Such software can be executed using a processor (e.g., a central processing unit (CPU), a microcontroller, etc.) to cause the functions and features described herein to be performed. The computer-readable medium can be a machine-readable storage device such as a floppy diskette, a hard disk, a CD-ROM, a DVD, a memory stick, a memory card, a ROM, a PROM, an EPROM, a FLASH-EPROM, or the like. The software can be provided in a single module or multiple modules, each of which can include computer program code or instructions. The software can be executed using one or more processors operating in a single thread or multiple threads.

[0112] Computer program code for carrying out operations of the present application can be written in any combination of one or more programming languages, including an object- oriented programming language such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, Python, or the like, conventional procedural programming languages, such as the C programming language, Visual Basic, Fortran 2103, Perl, COBOL 2102, PHP, ABAP, dynamic programming languages, such as Python, Ruby and Groovy, or another programming language. The program code can execute entirely on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer, or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic program code can be downloaded from an on-line software source or through a cloud computing environment, or as a service, such as Software as a Service (SaaS).

[0113] In addition, the order of execution or sequence of processing elements and sequences, unless otherwise specifically indicated, can vary from that described. For example, although the above-disclosed subject matter has been described in the context of certain illustrative embodiments and examples, it should be appreciated that those skilled in the art will recognize that additional implementations and use cases not specifically described can utilize many of the same concepts as those described above, without departing from the spirit and scope of the present application. For example, while the above- described implementations can be embodied in the context of a smart phone, mobile device, or other mobile computing device, the concepts taught herein can be applied to any computing device, such as a desktop computer, a server computer, a mobile phone, a smart phone, a tablet computer, a gaming console, a wearable computer, a vehicle computer, a computer in an appliance, or other mobile or non-mobile computer. Furthermore, although the above-disclosed subject matter has been described in the context of certain illustrative embodiments and examples, it should be appreciated that those skilled in the art will recognize that additional implementations and use cases not specifically described can utilize many of the same concepts as those described above, without departing from the spirit and scope of the present application. For example, while the above-described implementations can be embodied in the context of a smart phone, mobile device, or other mobile computing device, the concepts taught herein can be applied to any computing device, such as a desktop computer, a server computer, a mobile phone, a smart phone, a tablet computer, a gaming console, a wearable computer, a vehicle computer, a computer in an appliance, or other mobile or non-mobile computer.

[0114] For similar reasons, it is to be appreciated that the teachings of the present application provided herein can be applied to any embodiment of the present application, and that actual claims applied for or patent granted can be broader than any single, featured embodiment. Accordingly, a patent applicant has constructed and filed examples to particularly point out and distinctly claim those aspects which are regarded as novel and those aspects specifically shown.

Claims

1. A method for predicting drug-target binding affinity based on graph neural networks, characterized in that, include: (1) Obtain the SMILES sequence of the drug and the amino acid sequence of the target protein; (2) The SMILES sequence of the drug is converted into a drug molecule graph, and the features of the drug molecule graph are extracted by a graph convolutional neural network to obtain a drug feature vector. (3) Input the amino acid sequence of the target protein into the Transformer encoder for feature extraction to obtain the protein feature vector; (4) The drug feature vector and the protein feature vector are fused to obtain a fused feature vector; (5) Input the fused feature vector into the multilayer perceptron and output the drug-target binding affinity prediction result.

2. The drug-target binding affinity prediction method based on graph neural networks according to claim 1, characterized in that, In step (2), the drug molecule diagram is constructed with atoms as nodes and chemical bonds as edges, and each node and edge contains a corresponding feature vector.

3. The drug-target binding affinity prediction method based on graph neural networks according to claim 1, characterized in that, In step (2), the graph convolutional neural network includes three consecutive graph convolutional layers, each of which uses the ReLU activation function, and a global max pooling layer is connected after the third graph convolutional layer to output the drug feature vector through the global max pooling layer.

4. The drug-target binding affinity prediction method based on graph neural networks according to claim 1, characterized in that, In step (3), the Transformer encoder is composed of 10 stacked layers with the same structure. Each layer includes a multi-head self-attention mechanism module and a feedforward neural network module. The number of heads in the multi-head self-attention mechanism is 4.

5. The drug-target binding affinity prediction method based on graph neural networks according to claim 4, characterized in that, The multi-head self-attention mechanism calculates the correlation between any two amino acids in the protein amino acid sequence to generate three vectors: query, bond, and value. Attention is then calculated based on these three vectors, and the attention results from multiple heads are concatenated and output.

6. The drug-target binding affinity prediction method based on graph neural networks according to claim 1, characterized in that, In step (3), the output of the Transformer encoder is processed by a global average pooling layer to obtain the protein feature vector. The protein feature vector is a fixed-length vector that integrates global information of the protein sequence.

7. The drug-target binding affinity prediction method based on graph neural networks according to claim 1, characterized in that, In step (4), fusion is achieved by concatenating the drug feature vector with the protein feature vector, and the fused feature vector is the concatenated vector.

8. The drug-target binding affinity prediction method based on graph neural networks according to claim 1, characterized in that, In step (1), the SMILES sequence of the drug and the amino acid sequence of the target protein are derived from the Davis dataset or the KIBA dataset. The Davis dataset contains interaction data of 68 drugs and 442 target proteins, and the KIBA dataset contains interaction data of 2111 drugs and 229 proteins.

9. The drug-target binding affinity prediction method based on graph neural networks according to claim 1, characterized in that, The prediction results are evaluated using metrics including mean squared error, consistency index, and corrected coefficient of determination. The mean squared error measures the average squared deviation between the predicted and actual values. The consistency index measures the probability that the predicted and actual values ​​are in the same order. The corrected coefficient of determination is used to evaluate the model's external predictive power and robustness.

10. The drug-target binding affinity prediction method based on graph neural networks according to claim 1, characterized in that, The graph convolutional neural network, Transformer encoder, and multilayer perceptron are trained using the Adam optimizer with 6-fold cross-validation and a learning rate of 0.0001.

Images