Drug-target binding affinity prediction method based on cross-scale and cross-modal attention mechanism

Abstract

The present application relates to a drug-target binding affinity prediction method based on cross-scale and cross-modal attention mechanism, belonging to the field of drug research and development. The method comprises: the input drug and target protein are respectively extracted by a multi-element convolution kernel-based feature extraction module to obtain three-scale features, an interaction module obtains drug features and target protein features with key information according to the two different scale features of the drug and target protein, and a drug-target interaction module extracts features containing key binding sites based on the drug features and target protein features with key information; through the multi-element convolution kernel-based feature extraction module, the cross-scale attention-based interaction module and the cross-modal attention-based drug-target interaction module, features containing biological information are obtained, and the regression module is used to predict the binding affinity value of the drug-target based on the drug features and target protein features containing biological information.

Description

Technical Field

[0001] This invention belongs to the field of drug development and relates to a method for predicting drug-target binding affinity based on cross-scale and cross-modal attention mechanisms. Background Technology

[0002] With the rapid development of artificial intelligence (AI), it is accelerating the advancement of life sciences and biomedicine towards faster, more precise, safer, more economical, and more inclusive outcomes. Computer-aided drug design has become a crucial component of drug discovery and development, and is gradually taking a dominant position due to its superior performance. Among these, effectively identifying drug-target interactions and accurately predicting drug-target binding affinity are fundamental to various tasks in computer-aided drug design.

[0003] Therefore, accurately predicting drug-target binding affinity is an important fundamental task in order to accelerate the drug discovery and development process.

[0004] Task Definition: The task of drug-target binding affinity prediction aims to predict the affinity matrix I between drug D and target T. Each drug is represented by its SMILES string, and each target protein is represented by its amino acid sequence. The affinity matrix contains the affinity values ​​between the aforementioned drugs and targets. After a novel target protein is identified, the drug-target affinity prediction method can be used to efficiently screen potential drug structures that may be effective against this novel target protein structure.

[0005] Limitations of Related Methods: Several methods already exist that utilize one-dimensional sequence information of drugs and targets to predict the binding affinity of drug-target interactions. DeepDTA focuses on using one-dimensional data to predict the binding affinity of drug-target interactions, which is more efficient than methods using two-dimensional representations and three-dimensional structures as input. Its performance surpasses previous methods, but its network structure cannot further capture deep hidden information. DeepAffinity can efficiently extract key features of drugs or targets separately and fully utilize a large amount of unlabeled data, but this increases computational cost and significantly reduces network convergence. PCNN-DTA addresses the problem of semantic information gradually being lost in low-level features as the network expands. The PCNN-DTA method, based on Feature Pyramid Network (FPN), integrates features extracted from each layer of a multi-layer convolutional network, retaining more low-level feature information and thus improving prediction accuracy. However, it requires significant computational resources and has limitations in predicting and simulating structural changes. Furthermore, none of the above methods consider the biochemical properties of drugs and targets or the biological principles underlying their binding process, which can lead to the omission of some key information during feature extraction, affecting the final results. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a method for predicting drug-target binding affinity based on cross-scale and cross-modal attention mechanisms.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting drug-target binding affinity based on cross-scale and cross-modal attention mechanisms includes the following steps: S1. Preprocess the drug and target protein to be predicted, wherein the amino acid sequence of the target protein in the form of an amino acid sequence is encoded as a number, and the SMILES string of the drug in the form of a SMILES string is encoded as a number. S2. Input the preprocessed drug and target protein information into the feature extraction module based on multivariate convolution kernel of the prediction method to extract features of the drug and target protein at three scales respectively. S3. Input the features of the drug and target protein at two scales into the cross-scale attention-based interaction module. The cross-scale attention-based interaction module, combined with the attention mechanism, yields drug features and target protein features with key information. S4. Input the drug features and target protein features with key information into the drug-target interaction module based on cross-modal attention. The drug-target interaction module based on cross-modal attention, combined with the attention mechanism, extracts features containing key binding sites. S5. Features containing biological information are obtained through a feature extraction module based on multivariate convolution kernels, an interaction module based on cross-scale attention, and a drug-target interaction module based on cross-modal attention. The drug features and target protein features containing biological information are input into the regression module, which predicts the binding affinity of the drug and target through three fully connected operations.

[0008] Furthermore, in step S1, the drug, target protein, and drug-target affinity are pretreated, wherein... For target proteins represented by amino acid sequences, the amino acid sequences composed of letters are encoded as numbers. The numerical sequences representing target proteins are embedded into an array of length 1000. The sequences exceeding the maximum length are truncated, and the insufficient parts are padded with 0. The array of length 1000 is encoded into a vector with dimensions (1000, 128). For drugs represented by the string SMILES, the SMILES string composed of letters is encoded into numbers, and the sequence of numbers representing the drug is embedded into an array of length 100. During this process, the part of the sequence that exceeds the maximum length is truncated, and the part that is insufficient is padded with 0. The array of length 100 is encoded into a vector of dimension (100, 128). Regarding drug-target affinity, the minimum drug concentration required to inhibit the target protein is determined by drug concentration. The minimum drug concentration required to inhibit the target protein Perform the following mathematical transformation:

[0009] in, This indicates the affinity value between the drug and the target. The larger the value, the higher the affinity.

[0010] Furthermore, in step S2, the feature extraction module based on multivariate convolution kernels, combined with downsampling operators, extracts features at three scales from the input drug SMILES and target protein sequences. Among them, the deep features at the smallest scale contain key location information of biological structures, which can be used to find key sites for drug-target interaction; the shallow features at the largest scale are used to predict the specific value of binding affinity; and the features at the intermediate scale balance the errors caused by the extreme cases of the smallest and largest scale features and provide information. The feature extraction module based on multivariate convolutional kernels takes drug or target protein features as input. First, it obtains drug or target protein features under different receptive fields through three three-branch convolutional modules with different kernel sizes. Then, it uses the left and right residual branches to obtain supplementary features. Finally, it adds the three-branch features and the supplementary features, as shown in the following formula:

[0011]

[0012] in For the input characteristics of drugs or target proteins, and These represent the features obtained from the three-branch structure and the output features from the feature extraction module based on multi-convolutional kernels, respectively. It is the ReLU activation function. This represents convolution operations with different kernel sizes.

[0013] Furthermore, in step S3, the interaction module based on cross-scale attention, combined with the attention mechanism, processes the maximum-scale shallow features and intermediate-scale features of the drug and target protein, respectively; a learnable parameter matrix is ​​set in the attention mechanism. , and The process includes the following steps: Features at different scales and The Q and KV values ​​are used as inputs in the attention mechanism and are respectively compared with the learnable parameter matrix. , and Multiplication, the formula is expressed as:

[0014]

[0015]

[0016] Then the feature matrix and characteristic matrix Perform fully connected operations on each component and multiply them together, then normalize the results to obtain the final output. The formula is:

[0017] in, express Activation function Indicates a fully connected layer. This indicates a normalization operation; Next, the processing results will be... The feature matrix after the fully connected operation After multiplying, perform a full join operation to obtain... The specific formula is as follows:

[0018] Finally, and The results of a single convolution operation with a 1x1 kernel are summed to obtain the final drug feature containing key information. The formula is expressed as:

[0019] in, This indicates a convolution operation with a 1x1 kernel. Characteristics of drugs at two scales and In obtaining drug characteristics with key information At that time, and Each as input and In obtaining drug characteristics with key information At that time, and Each as input and ; Features of the target protein at two scales and In obtaining target protein features containing key information At that time, and Each as input and In obtaining drug characteristics with key information At that time, and Each as input and .

[0020] Furthermore, in step S4, the drug-target interaction module based on cross-modal attention combines attention mechanisms to analyze drug characteristics. and target protein characteristics Extracting features containing key binding sites includes the following steps: For drug characteristics containing key binding sites Drug characteristics and target protein characteristics Q and KV, which serve as the attention mechanism, are calculated according to the following formula:

[0021]

[0022]

[0023]

[0024]

[0025]

[0026] For target protein features containing key binding sites Characterize the target protein and drug characteristics Q and KV, which serve as the attention mechanism, are calculated according to the following formula:

[0027]

[0028]

[0029]

[0030]

[0031]

[0032] in Indicates a fully connected layer. This indicates a normalization operation. This represents a convolution operation with a 1x1 kernel. express Activation function.

[0033] Furthermore, in step S5, features containing biological information are obtained through a feature extraction module based on multivariate convolutional kernels, an interaction module based on cross-scale attention, and a drug-target interaction module based on cross-modal attention. The regression module then analyzes the drug features containing biological information. and target protein characteristics Perform several fully connected operations to predict drug-target binding affinity, using the following formula:

[0034] in For the predicted drug-target binding affinity, Indicates a fully connected layer. It is the ReLU activation function. Indicates a connection operation. This is the max pooling layer.

[0035] Furthermore, during the training process of the prediction method, mean squared error (MSE) is chosen to quantify the difference between the predicted affinity and its corresponding true value, and the loss function... The calculation process is as follows:

[0036] in, This represents the true value of drug-target binding affinity. This represents the predicted value of drug-target binding affinity.

[0037] The beneficial effects of this invention are as follows: This invention utilizes a multi-element convolutional kernel network (KTN) to extract key features based on the biochemical properties of functional groups formed by multiple atomic chemical bonds and peptide chains formed by the dehydration condensation of several amino acids, which play crucial roles in the properties of drugs and targets. Furthermore, a cross-scale interaction module (CSIM) further enriches and refines the extracted features. Considering the physiological interactions between drugs and targets, this invention employs a drug-target interaction module (DTIM) to utilize a cross-attention mechanism to find binding sites between drugs and targets. Extensive experiments on two public datasets demonstrate the superiority and stability of the proposed DTANet.

[0038] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0039] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is the overall network structure of the DTANet drug-target affinity prediction method of the present invention; Figure 2 This is a simplified flowchart illustrating the DTANet drug-target affinity prediction method of the present invention. Figure 3 This is a structural diagram of the feature extraction module based on multivariate convolution kernels in the prediction method of the present invention; Figure 4 This is a structural diagram of the interaction module based on cross-scale attention in the prediction method of the present invention; Figure 5 This is a structural diagram of the attention mechanism in the prediction method of the present invention; Figure 6 This is a structural diagram of the drug-target interaction module based on cross-modal attention in the prediction method of the present invention; Figure 7 This is a structural diagram of the regression module for predicting drug-target binding affinity in the prediction method of the present invention; Figure 8 This is a flowchart illustrating the training process of the prediction method of the present invention. Figure 9 This is a flowchart of the test process for the prediction method of the present invention; Figure 10 The results show the consistency index (CI) and standard deviation (std) of this invention compared with state-of-the-art methods on the KIBA dataset. Figure 11 The qualitative comparison results of the present invention and the method CoVAE on the KIBA dataset are shown in Figure (a), which shows the distribution of the true value of binding affinity and the predicted value of the method CoVAE, and Figure (b) shows the distribution of the true value of binding affinity and the predicted value of the present invention. Figure 12 The ROC curve of the present invention and method CoVAE on the KIBA dataset; Figure 13 The AUC value of the present invention and method CoVAE on the KIBA dataset varies with a threshold. Detailed Implementation

[0040] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0041] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0042] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0043] Please see Figures 1 to 13 This is a method for predicting drug-target binding affinity based on cross-scale and cross-modal attention mechanisms.

[0044] This invention proposes a method for predicting drug-target binding affinity based on cross-scale and cross-modal attention mechanisms, such as... Figure 1 and Figure 2 The overall framework shown includes the following steps: S1. Preprocess the drug and target protein to be predicted, wherein the amino acid sequence of the target protein in the form of an amino acid sequence is encoded as a number, and the SMILES string of the drug in the form of a SMILES string is encoded as a number. S2. Input the preprocessed drug and target protein information into the feature extraction module based on multivariate convolution kernel of the prediction method to extract features of the drug and target protein at three scales respectively. S3. Input the features of the drug and target protein at two scales into the cross-scale attention-based interaction module. The cross-scale attention-based interaction module, combined with the attention mechanism, yields drug features and target protein features with key information. S4. Input the drug features and target protein features with key information into the drug-target interaction module based on cross-modal attention. The drug-target interaction module based on cross-modal attention, combined with the attention mechanism, extracts features containing key binding sites. S5. Features containing biological information are obtained through a feature extraction module based on multivariate convolution kernels, an interaction module based on cross-scale attention, and a drug-target interaction module based on cross-modal attention. The drug features and target protein features containing biological information are input into the regression module, which predicts the binding affinity of the drug and target through three fully connected operations.

[0045] According to the drug-target binding affinity prediction method based on cross-scale and cross-modal attention mechanisms proposed in this invention, the following functional modules are proposed: A feature extraction module based on multivariate convolution kernels is used to extract multiscale features of drugs and / or target proteins; Interaction modules based on cross-scale attention are used to simulate the relationships between functional groups or peptide chains of different lengths through multi-scale features. A drug-target interaction module based on cross-modal attention is used to fuse drug features and target protein features to simulate the biological interaction between drugs and target proteins and obtain features containing key binding sites. The regression module is used to predict the binding affinity value of the drug-target based on features containing biological information obtained by the feature extraction module based on multivariate convolution kernels, the interaction module based on cross-scale attention, and the drug-target interaction module based on cross-modal attention. In this process, the drug and target protein input to the prediction method are extracted into features at three scales by a feature extraction module based on multi-convolutional kernels. A cross-scale attention-based interaction module, combined with an attention mechanism, obtains drug and target protein features with key information based on the drug and target protein features at two scales, respectively. A cross-modal attention-based drug-target interaction module, based on the drug and target protein features with key information and combined with an attention mechanism, simulates the biological interaction between the drug and target protein to obtain features containing key binding sites. These features, containing biological information, are then used by a regression module to predict the drug-target binding affinity based on the biologically informative drug and target protein features.

[0046] The drug-target affinity prediction method DTANet is essentially a multi-kernel convolutional network (KTN). This prediction method extracts features from both the drug and the target protein separately. Specifically, it is processed by a feature extraction module based on multi-kernel convolution (KFEB), an interaction module based on cross-scale attention (CSIM), and a drug-target interaction module based on cross-modal attention (DTIM). In addition, it uses deconvolution operation (Convt module) for upsampling. Finally, it obtains key features with rich information as input to the regression module, thereby completing the drug-target affinity prediction.

[0047] Furthermore, in the prediction method, this invention designs convolutional kernels of different sizes based on the biological structural features of the drug and target protein, respectively. Combining residual and inception structures, this invention constructs a feature extraction module (KFEB) based on multi-kernel convolution, and uses the KFEB and downsampling operators to extract features at three scales from the input drug SMILES and target protein sequences. The deep features at the smallest scale typically contain key location information of the entire biological structure and can be used to find key sites for drug-target binding. The shallow features at the largest scale contain detailed information, which helps predict the specific value of binding affinity. Features at intermediate scales can balance the errors caused by the extreme cases of the smallest and largest scale features and provide richer information. More specifically, as... Figure 3The diagram shown illustrates the structure of the feature extraction module based on multi-kernel convolution. It takes drug or target protein features as input. First, it uses three three-branch convolutional modules with different kernel sizes to obtain drug or target protein features under different receptive fields. Inspired by residual structures, it designs left and right residual branches to obtain supplementary features. Finally, the three-branch features and the supplementary features are added together. The processing details of the multi-kernel convolution feature extraction module (KFEB) are shown below:

[0048]

[0049] in For the input characteristics of drugs or target proteins, and These represent the features obtained from the three-branch structure and the output features from the feature extraction module based on multi-convolutional kernels, respectively. It is the ReLU activation function. This represents convolution operations with different kernel sizes.

[0050] Furthermore, in prediction methods, such as Figure 4 The diagram shown illustrates the architecture of the Cross-Scale Attention-Based Interaction Module (CSIM), which integrates features from two scales obtained through a feature extraction module based on multivariate convolutional kernels and a downsampling operator. and Simultaneously, the features are input into the Cross-Scale Attention Interaction Module (CSIM), where the features from the two scales serve as Q and KV in the attention mechanism, respectively, complementing each other's important feature information and thus outputting key information features at different scales. Through this operation, the relationships between functional groups (or peptide chains) of different lengths can be simulated more effectively, thereby enabling the extraction of more reliable features. More specifically, such as... Figure 5 The illustrated structure diagram of the attention mechanism includes at least a feature matrix. , and The designed learnable parameter matrix , and ; When the attention mechanism is used in an interaction module based on cross-scale attention, the input to the attention mechanism is... and Among them, and If we represent the characteristics of a drug at different scales, then... With learnable parameter matrix Multiplication yields the characteristic matrix , Learnable parameter matrices Multiplication yields the characteristic matrix and The formula is expressed as:

[0051]

[0052]

[0053] Then the feature matrix and characteristic matrix Perform fully connected operations on each component and multiply them together, then normalize the results to obtain the final output. The specific formula is as follows:

[0054] in, express Activation function Indicates a fully connected layer. This indicates a normalization operation. This indicates an interest in drug characteristics. hour, More important feature regions; Next, the processing results will be... The feature matrix after the fully connected operation After performing a full connect operation, we get... The specific formula is as follows:

[0055] in, This indicates an interest in drug characteristics. hour, More important feature information; Finally, and The results of a single 1x1 convolution operation are summed to obtain the final drug feature with key information obtained through the Cross-Scale Attention-Based Interaction Module (CSIM). The formula is expressed as:

[0056] in, This indicates a convolution operation with a 1x1 kernel.

[0057] Similarly, to obtain drug characteristics containing key information At that time, Q, as an attention mechanism, K and V can be used as the attention mechanism; the overall process is the same, and the specific formula is as follows:

[0058]

[0059]

[0060]

[0061]

[0062]

[0063] Similarly, when and When representing target protein characteristics at different scales, and The process remains consistent to obtain target protein features with key information through the Cross-Scale Attention-Based Interaction Module (CSIM).

[0064] Furthermore, in prediction methods, the binding of a drug to a target is a complex biochemical reaction in biology or physiology. Most existing methods simply add or connect the features obtained from the drug and target into a fusion feature before predicting binding affinity, thus neglecting the biological or physiological interactions between the drug and target. This invention proposes... Figure 5 The diagram shown illustrates the structure of the cross-modal attention-based drug-target interaction module (DTIM), which integrates drug features. and target protein characteristics Q and KV, respectively, serve as attention mechanisms or target protein features and drug characteristics Q and KV, respectively, serve as attention mechanisms, simulating the biological interaction between drugs and target proteins and obtaining features including key binding sites. The processing details of the drug-target interaction module (DTIM) based on cross-modal attention are shown below:

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071]

[0072]

[0073]

[0074]

[0075]

[0076] in Indicates a fully connected layer. This indicates a normalization operation. This represents a convolution operation with a 1x1 kernel. express Activation function. Furthermore... and These represent the characteristics of the drug and the target protein, respectively. , and It is the learnable parameter matrix in the attention mechanism. and This is a drug signature and target protein signature containing key binding sites obtained through the drug-target interaction module (DTIM) based on cross-modal attention.

[0077] Furthermore, in prediction methods, such as Figure 7 The diagram shows the structure of the regression module for predicting drug-target binding affinity. Key features of the drug and target are input into the regression module, and the binding affinity is predicted through three fully connected operations. The processing details of the regression module are shown below:

[0078] in For the predicted drug-target binding affinity, Indicates a fully connected layer. It is the ReLU activation function. Indicates a connection operation. This is a max pooling layer; and These represent the drug features and target protein features containing biological information obtained through the feature extraction module based on multivariate convolution kernels, the interaction module based on cross-scale attention, and the drug-target interaction module based on cross-modal attention, respectively.

[0079] Meanwhile, this invention chooses mean squared error (MSE) to quantify the difference between the predicted affinity and its corresponding true value. Loss function The specific calculation process is as follows:

[0080] in This represents the true value of drug-target binding affinity. This represents the predicted value of drug-target binding affinity.

[0081] Furthermore, during model training, the model is trained using drugs, target proteins, and affinity values ​​between the drugs and targets. Each drug is represented by its SMILES string, and each target protein is represented by its amino acid sequence. The affinity matrix for drugs and target proteins contains the aforementioned affinity values ​​between the drugs and targets.

[0082] When testing the model, once a new target protein is identified, it is represented by its amino acid sequence. The aim of this invention is to enable rapid searching of lead compounds that interact with the aforementioned target proteins using existing databases; specifically, to quickly identify compounds with high binding affinity to target proteins and potential pharmaceutical value by searching drug databases. Each drug in the drug database is represented by its SMILES string. The main contribution of this invention is the use of deep neural networks to predict the binding affinity between drugs and target proteins, reducing the human, material, and time costs required for traditional pharmaceutical experiments to determine drug-target binding affinity, and accelerating the drug discovery and development process.

[0083] For target proteins represented by amino acid sequences, in Figure 2 In the embedding operation, the amino acid sequence composed of letters is first encoded as numbers, that is, different numbers represent different letters—this format is more convenient for subsequent computer calculations. Then, the numerical sequence representing the target protein is embedded into an array of length 1000. During this process, sequences exceeding the maximum length are truncated, and insufficient sequences are padded with 0s. Finally, the array of length 1000 is encoded into a vector of dimension (1000, 128).

[0084] For drugs represented as the string SMILES, in Figure 2In the embedding operation, the SMILES string, composed of letters, is first encoded into numbers, with different numbers representing different SMILES letters—this format facilitates subsequent computer calculations. Then, the numerical sequence representing the drug is embedded into an array of length 100. During this process, any part of the sequence exceeding the maximum length is truncated, and insufficient parts are padded with 0s. Finally, the array of length 100 is encoded into a vector of dimension (100, 128).

[0085] To determine the affinity between a drug and its target, data preprocessing is performed before model training. The affinity value can be used to determine the minimum drug concentration required to inhibit the target protein. Perform the following mathematical transformation:

[0086] in, This indicates the affinity value between the drug and the target. The larger the value, the higher the affinity, which means the greater the potential medicinal value.

[0087] When training the model, such as Figure 8 The training flowchart shown illustrates how preprocessed drug and target features are input into a prediction method for prediction. The predicted affinity output by the prediction method is compared with the true affinity, and the model is optimized using a loss function until training is complete. The loss function is expressed as: .

[0088] Furthermore, such as Figure 9 The test flowchart shown also requires each drug to be represented by its SMILES string and each target protein by its amino acid sequence. Drugs and target proteins are used as input data, and the DTANet model is used to predict drug-target binding affinity.

[0089] In summary, this invention utilizes deep convolutional neural network algorithms to design a deep architecture called DTANet for drug-target affinity prediction. This architecture includes a multi-kernel convolutional network (KTN), a cross-scale attention-based interaction module (CSIM), a cross-modal attention-based drug-target interaction module (DTIM), and a regression module for predicting drug-target binding affinity. In biochemistry, functional groups formed by multiple atomic chemical bonds and peptide chains formed by the dehydration condensation of several amino acids play crucial roles in the properties of drugs and targets. Therefore, this invention designs a multi-kernel convolutional network (KTN) based on these biochemical characteristics to extract key features. Furthermore, the cross-scale attention-based interaction module (CSIM) further enriches and refines the extracted features. Considering the physiological interactions between drugs and targets, the cross-modal attention-based drug-target interaction module (DTIM) utilizes a cross-attention mechanism to find the binding sites of drugs and targets.

[0090] Example This embodiment compares the model's prediction results with existing state-of-the-art methods.

[0091] Figure 10 The results show a comparison of the consistency index (CI) and standard deviation (std) of the DTANet prediction method of this invention with state-of-the-art methods on the KIBA dataset. A higher CI value indicates higher prediction accuracy, while a lower std value indicates higher stability. The DTANet prediction method of this invention achieves higher CI and lower std, highlighting its superior performance.

[0092] Figure 11 The qualitative comparison results of the DTANet prediction method of the present invention and the CoVAE method on the KIBA dataset are shown. Figure 11 (a) shows the distribution of the true value of binding affinity and the predicted value of method CoVAE. Figure 11 (b) shows the distribution of the true values ​​of affinity and the predicted values ​​of the DTANet prediction method of the present invention. The more concentrated the results are in the middle of the dashed line, the better the performance. It can be seen intuitively that the results obtained by the DTANet prediction method of the present invention are more concentrated, that is, the prediction results are more accurate.

[0093] Figure 12 The diagram shows the ROC curves of the DTANet prediction method and the CoVAE method of this invention on the KIBA dataset. The area under the ROC curve can directly evaluate the performance of the classifier. The larger the value, the better the model performance. As can be seen, the DTANet prediction method of this invention achieves better results.

[0094] Figure 13The AUC values ​​obtained by the DTANet prediction method and the method CoVAE of the present invention on the KIBA dataset are shown to change with the threshold. It can be seen that the DTANet prediction method of the present invention performs better under different thresholds.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting drug-target binding affinity based on cross-scale and cross-modal attention mechanisms, characterized in that: It includes the following steps: S1. Preprocess the drug and target protein to be predicted, wherein the amino acid sequence of the target protein in the form of an amino acid sequence is encoded as a number, and the SMILES string of the drug in the form of a SMILES string is encoded as a number. For a target protein represented by an amino acid sequence, the amino acid sequence composed of letters is encoded into numbers, and the numerical sequence representing the target protein is embedded into an array of a first preset length. The part of the sequence that exceeds the maximum length is truncated, and the part that is insufficient is padded with 0. The array of the first preset length is encoded into a vector of a first preset dimension. For drugs represented by the string SMILES, the SMILES string composed of letters is encoded into numbers, and the sequence of numbers representing the drug is embedded into an array of a second preset length. During this process, the part of the sequence that exceeds the maximum length is truncated, and the part that is insufficient is padded with 0; the array of the second preset length is encoded into a vector of a second preset dimension. S2. Input the preprocessed drug and target protein information into the feature extraction module based on multivariate convolution kernel of the prediction method to extract features of the drug and target protein at three scales respectively. S3. Input the features of the drug and target protein at two scales into the cross-scale attention-based interaction module. The cross-scale attention-based interaction module, combined with the attention mechanism, yields drug features and target protein features with key information. S4. Input the drug features and target protein features with key information into the drug-target interaction module based on cross-modal attention. The drug-target interaction module based on cross-modal attention, combined with the attention mechanism, extracts features containing key binding sites. S5. Features containing biological information are obtained through a feature extraction module based on multivariate convolution kernels, an interaction module based on cross-scale attention, and a drug-target interaction module based on cross-modal attention. The drug features and target protein features containing biological information are input into the regression module, which predicts the binding affinity of the drug and target through three fully connected operations.

2. The drug-target binding affinity prediction method based on cross-scale and cross-modal attention mechanisms according to claim 1, characterized in that: In step S1, the drug, target protein, and drug-target affinity are pretreated, wherein... Regarding drug-target affinity, the minimum drug concentration required to inhibit the target protein is determined by drug concentration. The minimum drug concentration required to inhibit the target protein Perform the following mathematical transformation: in, This indicates the affinity between the drug and the target. The larger the value, the higher the affinity.

3. The drug-target binding affinity prediction method based on cross-scale and cross-modal attention mechanisms according to claim 2, characterized in that: In step S2, the feature extraction module based on multivariate convolution kernels, combined with downsampling operators, extracts features at three scales from the input drug SMILES and target protein sequences. The deep features at the smallest scale contain key location information of biological structures, which can be used to find key sites for drug-target interaction. The shallow features at the largest scale are used to predict the specific value of binding affinity. The features at the intermediate scale balance the errors caused by the extreme cases of the smallest and largest scale features and provide information. The feature extraction module based on multivariate convolutional kernels takes drug or target protein features as input. First, it obtains drug or target protein features under different receptive fields through three three-branch convolutional modules with different kernel sizes. Then, it uses the left and right residual branches to obtain supplementary features. Finally, it adds the three-branch features and the supplementary features, as shown in the following formula: in For the input characteristics of drugs or target proteins, and These represent the features obtained from the three-branch structure and the output features from the feature extraction module based on multi-convolutional kernels, respectively. It is the ReLU activation function. This represents convolution operations with different kernel sizes.

4. The drug-target binding affinity prediction method based on cross-scale and cross-modal attention mechanisms according to claim 3, characterized in that: In step S3, the interaction module based on cross-scale attention combines the attention mechanism to process the maximum-scale shallow features and intermediate-scale features of the drug and target protein, respectively. Setting a learnable parameter matrix in the attention mechanism , and The process includes the following steps: Features at different scales and The Q and KV values ​​are used as inputs in the attention mechanism and are respectively compared with the learnable parameter matrix. , and Multiplication, expressed by the formula: Then the feature matrix and characteristic matrix Perform fully connected operations on each component and multiply them together, then normalize the results to obtain the final output. The formula is: in, express Activation function Indicates a fully connected layer. This indicates a normalization operation; Next, the processing results will be... The feature matrix after the fully connected operation After multiplying, perform a full join operation to obtain... The specific formula is as follows: Finally, and The results of a single convolution operation with a 1x1 kernel are summed to obtain the final drug feature containing key information. The formula is expressed as: in, This indicates a convolution operation with a 1x1 kernel. Characteristics of drugs at two scales and In obtaining drug characteristics with key information At that time, and Each as input and In obtaining drug characteristics with key information At that time, and Each as input and ; Features of the target protein at two scales and In obtaining target protein features containing key information At that time, and Each as input and In obtaining drug characteristics with key information At that time, and Each as input and .

5. The drug-target binding affinity prediction method based on cross-scale and cross-modal attention mechanisms according to claim 4, characterized in that: In step S4, the drug-target interaction module based on cross-modal attention combines attention mechanisms to analyze drug features. and target protein characteristics Extracting features containing key binding sites includes the following steps: For drug characteristics containing key binding sites Drug characteristics and target protein characteristics Q and KV, which serve as the attention mechanism, are calculated according to the following formula: For target protein features containing key binding sites , target protein characteristics and drug characteristics Q and KV, which serve as the attention mechanism, are calculated according to the following formula: in Indicates a fully connected layer. This indicates a normalization operation. This represents a convolution operation with a 1x1 kernel. express Activation function.

6. The drug-target binding affinity prediction method based on cross-scale and cross-modal attention mechanisms according to claim 5, characterized in that: In step S5, features containing biological information are obtained through a feature extraction module based on multivariate convolutional kernels, an interaction module based on cross-scale attention, and a drug-target interaction module based on cross-modal attention. The regression module then analyzes the drug features containing biological information. and target protein characteristics Perform several fully connected operations to predict drug-target binding affinity, using the following formula: in For the predicted drug-target binding affinity, Indicates a fully connected layer. It is the ReLU activation function. Indicates a connection operation. This is the max pooling layer.

7. The drug-target binding affinity prediction method based on cross-scale and cross-modal attention mechanisms according to any one of claims 1-6, characterized in that: During the training process of the prediction method, the mean squared error (MSE) is chosen to quantify the difference between the predicted affinity and its corresponding true value, and the loss function is... The calculation process is as follows: in, This represents the true value of drug-target binding affinity. This represents the predicted value of drug-target binding affinity.

Images