Drug target binding affinity prediction method based on drug bimodal characteristics

Abstract

The present application relates to a drug target binding affinity prediction method based on drug bimodal characteristics, and belongs to the field of drug target affinity prediction. First, a three-channel deep learning network is constructed to process the sequence of biological targets, the sequence of drug molecules and their chemical structure diagram respectively. After preprocessing, the target sequence is embedded into a high-dimensional feature vector, and XLNet-DSC is used to extract deep biological information. At the same time, the SMILES sequence of the drug molecule is converted into a chemical structure diagram, and key chemical features are extracted through graph neural network technologies such as graph attention network and message passing neural network. The sequence of the drug molecule is processed through a specific encoding and neural network to obtain further feature representation. By fusing these multi-dimensional features and applying a multi-head attention mechanism for detailed feature integration, the method can accurately predict the binding affinity between drugs and targets.

Description

Technical Field

[0001] This invention belongs to the field of drug target affinity prediction and relates to a method for predicting drug target binding affinity based on drug bimodal characteristics. Background Technology

[0002] Drug discovery is the process of finding potential novel drugs, involving multiple fields such as pharmacology, chemistry, and biology, and typically requires significant economic and time costs. Deep learning-based drug screening can eliminate the financial, energy, and time consumption associated with traditional experimental screening, reducing investment in drug development and making it a hot research direction in current new drug development technologies. In modern drug discovery and bioinformatics, accurately predicting the binding affinity between drug molecules and their biological targets is crucial for effective drug design. This process involves understanding complex intermolecular interactions, which is essential for screening and optimizing potential drug candidates. Traditional methods often rely on simplified molecular models and limited experimental data, which restricts their accuracy and efficiency when dealing with complex biological systems.

[0003] Accurate prediction of the affinity between drug molecules and protein targets is crucial in developing new drugs and understanding the interactions between drugs and their targets. In recent years, deep learning techniques, particularly graph neural networks (GNNs), natural language processing (NLP) models, and attention mechanisms, have become a research hotspot in this field. These methods can extract valuable features from complex biomolecular data, providing profound insights for drug design and biological target identification. They can learn complex molecular features from large amounts of data, improving the performance of predictive models. While these methods excel in handling standard molecular structures, NLP models often generate large amounts of high-dimensional feature data when extracting protein features, making training and inference inefficient. Furthermore, each node of a drug lacks rich feature representations, failing to capture the complex structural and interaction information within the molecule.

[0004] Therefore, there is an urgent need to develop a new method that can more comprehensively and accurately extract drug and target features, and effectively fuse these features to improve the accuracy and efficiency of binding affinity prediction. Furthermore, this method should be able to provide interpretable predictions while maintaining high computational efficiency, so that scientists and researchers can better understand and interpret the results. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method for predicting drug target binding affinity based on drug bimodal characteristics.

[0006] To achieve the above objectives, the present invention provides the following technical solution: S1: Construct a three-channel deep learning network, with the target sequence, drug molecule sequence, and drug molecule map as inputs respectively.

[0007] S2: Constructing the target feature extraction network: S11: Preprocess the target data, transforming it into a format suitable for the XLNet model. Perform sequence labeling for a given protein sequence. Use one-hot encoding for each Mapped to a unique tag ; S12: For the labeled sequence Convert to embedded representation:

[0008] in It is the interrelationship of each amino acid in sequence T. It is the corresponding position embedding.

[0009] S13: Feed the embedded representation into the self-attention layer. For each location Calculate its query ( ),key( ) and value ( ): ;in These are three learnable weight matrices. A self-attention mechanism is then applied to obtain a globally sensitive representation for each position. ,in This is the dimension of the key (K) vector. The layer output is obtained through two inter-layer interactions and a feedforward network. First, LayerNorm is applied, then FFN is used, and LayerNorm is applied again. This process can be represented as:

[0010]

[0011]

[0012] in The specific calculation method is as follows: , It is input. and They are The mean and variance, It is a very small number added for numerical stability. and These are learnable parameters used to recover the proportions and positions of normalized data.

[0013] S14: Will Depth-separable convolution can be represented as: in This represents pointwise convolution. This represents depthwise convolution.

[0014] S2: Construct a drug molecule map feature extraction network. The specific implementation steps are as follows: S21: Use the cheminformatics tool RDKit to analyze the SMILES sequence of compound SMILES. Convert the SMILES string into a graphical representation of its chemical structure (molecular diagram). ,in It is a set of nodes (atoms). It is an edge set (chemical bond).

[0015] S22: Molecular diagram Input message passing neural network, each node of the molecular graph Features Updated in each round of message passing as follows: Among them It is a node The set of neighboring nodes, It is a node neighboring nodes The current characteristics, It is an average aggregation function. Update for fully connected layers.

[0016] S23: Molecular diagram An input graph isomorphic network can be represented as:

[0017] in It is a multilayer perceptron. These are learnable parameters.

[0018] S24: The features updated via MPNN are represented as follows The features updated via GIN are represented as The combined features are calculated using a weighted summation method. , to be represented as: , It is a weight coefficient that can learn a range from 0 to 1. (The graph...) The global information of the entire graph is obtained by summing the feature vectors of all nodes using the summation formula, and can be represented as follows: .

[0019] S3: Construct a drug sequence feature extraction network, as follows: S31: Drug molecules The input embedding layer obtains a fixed-length vector, which can be represented as: .

[0020] S32: Embed the vector Input a bidirectional RNN and compute the forward feeds respectively. and backward The hidden states are then concatenated with the forward and backward hidden states to form the bidirectional RNN output at each time step. Then will Input convolutional layer, set the kernel size to In position The convolution operation on can be represented as , It is the weight matrix of the convolution kernel. It is a bias term. It is an activation function.

[0021] S33: Output the convolutional layer The input maximal layer reduces the feature dimensionality and captures the most important features, which can be represented as: The drug sequence features are then obtained by inputting them into a fully connected layer. .

[0022] The beneficial effects of this invention are as follows: By utilizing models specifically designed for different data types (such as GIN and MPNNs for molecular graphs, and XLNet for sequence data), this method can achieve a deeper understanding of the structure and function of protein and drug molecules. This enhanced feature representation capability helps capture complex interactions between biomolecules, thereby improving the accuracy of predicting the affinity between drugs and protein targets. Simultaneously, employing deep separable convolutions to process protein sequence features retains the advantages of convolutional neural networks in capturing local patterns while improving model efficiency by reducing the number of parameters and computational complexity. This balance between efficiency and accuracy is particularly important for processing large-scale biological datasets, making this method both practical and efficient. Finally, by fusing features from different sources through a multi-head attention mechanism, intermolecular interactions can be captured from multiple perspectives. This not only increases the model's depth of understanding of potential patterns in the data but also helps improve the model's generalization ability on unseen data. This means the model can better predict the affinity between new drug molecules and proteins, accelerating the drug discovery and development process. The method proposed in this invention has high accuracy, efficiency, and effectiveness in predicting drug target binding affinity, outperforming traditional methods.

[0023] 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

[0024] 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 a flowchart of the present invention.

[0025] Figure 2 This is a network structure diagram of the present invention. Detailed Implementation

[0026] 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.

[0027] 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.

[0028] 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.

[0029] Please see Figures 1-2 The specific implementation details of each part of the present invention are as follows: 1. Data sets for the DTA (Digital Transmission and Acquisition) model were collected. Three public datasets from the DTA direction were used: the Davis dataset, the Metz dataset, and the KIBA dataset, which were divided into training and testing sets. Binding affinity in these datasets is represented by the dissociation constant (Kd), inhibition constant (Ki), and KIBA score, which indicate the binding strength between the protein and the compound. In the Davis and Metz datasets, binding affinity increases with increasing Kd and Ki values, while in the KIBA dataset, binding affinity decreases with increasing KIBA score. For computational convenience, Kd and Ki are converted to pKd and pKi, respectively, and can be expressed as:

[0030]

[0031] pKd means converting the corresponding kd to a logarithmic value, and pKi means converting the corresponding ki to a logarithmic value; The Davis dataset contains 30,056 interaction pairs between 442 proteins and 68 compounds. The Metz dataset contains 35,259 interaction pairs between 1,423 proteins and 170 compounds, and the KIBA dataset contains 118,254 interaction pairs between 2,111 proteins and 322 compounds.

[0032] 2. Input the target sequence into the XLNet-DSC module for feature extraction. Preprocess the target data to convert it into a format suitable for the XLNet model. Perform sequence labeling for a given protein sequence. Use one-hot encoding for each Mapped to a unique tag Protein sequence S The sequence after one-hot encoding is the labeled sequence. Convert to embedded representation The embedded representation is fed into the self-attention layer. For each location... Calculate its query ( ),key( ) and value ( Then, a self-attention mechanism is applied to obtain a globally sensitive representation of each location. ; ;in There are three learnable weight matrices. It is the dimension of the key (K) vector; Will The layer output is obtained through two inter-layer interactions and a feedforward network. First, LayerNorm is applied, then FFN is used, and LayerNorm is applied again to obtain the vector. Finally, the vector Input depthwise separable convolution to obtain target sequence features .

[0033] 3. Use the cheminformatics tool RDKit to analyze the SMILES sequence of the compound. Convert the SMILES string into a graphical representation of the chemical structure (molecular diagram), and define the diagram. ,in It is a set of nodes (atoms). It is an edge set (chemical bond). Molecular diagram Input message passing neural network, via method Update the feature information of each node to obtain the molecular graph feature vector. Molecular diagram The input graph isomorphic network can be represented as Obtain the feature vector of the molecular diagram The combined features are calculated using a weighted summation method. , to be represented as: The diagram The global information of the entire graph is obtained by summing the feature vectors of all nodes using the summation formula, and can be represented as follows: .

[0034] 4. To bring drug molecules The input embedding layer obtains a fixed-length vector, which can be represented as: Embedded vector Input a bidirectional RNN and compute the forward feeds respectively. and backward The hidden states are then concatenated with the forward and backward hidden states to form the bidirectional RNN output at each time step. Then will Input convolutional layer, set the kernel size to In position The convolution operation on can be represented as Output of the convolutional layer The input maximal layer reduces the feature dimensionality and captures the most important features, which can be represented as: Then, the drug sequence features are obtained by inputting them into a fully connected layer. .

[0035] 5. Target sequence feature vector Drug molecule map feature vector The fused features are obtained by inputting the drug sequence feature vector Z into a multi-head attention mechanism. , will integrate features Drug target binding affinity results were obtained through some fully connected layers.

[0036] 6. Evaluation Indicators We will use MSE (mean squared error) and CI (the concordance index) as evaluation metrics for model performance. MSE is a commonly used evaluation metric for regression tasks, which measures the performance of a sample by calculating the average of the sum of squared differences between the predicted and actual values. Predicted value and actual sample value The smaller the MSE, the better. The MSE is calculated as follows:

[0037] n It is the quantity of this book. y i These are actual values. p i These are model predictions. i Indicates the nth or nth. i =1, 2, 3, 4, 5... n ; CI is a metric that evaluates the consistency between the ranking of predicted values ​​and their actual accuracy. It is calculated as follows:

[0038] in yes The predicted value, yes The predicted value, h ( x ) is a step function. Z These are normalized hyperparameters. Step function. h ( x The definition is as follows:

[0039] 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 drug bimodal characteristics, characterized in that: The method includes the following steps: S1: Construct a three-channel deep learning network, with target sequence, drug molecule sequence and drug molecule map as inputs respectively; S2: First, the target sequence is preprocessed. The sequence is padded or split into equal-length sequences and then input into the embedding layer to obtain the sequence. This maps each amino acid to a unique vector and feeds it into the XLNet-DSC joint module for feature extraction; first, the sequence... High-dimensional feature representations are obtained by inputting into the XLNet model. ;Will As input to depthwise separable convolution, for each feature vector A large number of independent convolutional filters are applied, and then pointwise convolution is performed to fuse the features from different channels to obtain the target features. ; S3: Use the cheminformatics tool RDKit to parse the SMILES sequence of the compound; convert the SMILES string into a graphical representation of the chemical structure, i.e., a molecular diagram. Molecular diagram Simultaneously, feature vectors are obtained from graph attention networks (GIN) and message passing neural networks (MPNNs). and Then and Weighted summation, followed by pooling layers, aggregates node-level features into graph-level feature representations, yielding drug molecule graph features. ; S4: Input the SMILES sequence of the drug molecule into the embedding layer to obtain a fixed-length vector. , convert the sequence Input the data into a TextRCNN network for feature extraction to obtain drug sequence features. ; S5: Features , and Feature fusion is performed in the input multi-head attention, and the fused features are passed through a fully connected layer to obtain the drug-target binding affinity prediction results.

2. The drug target binding affinity prediction method based on drug bimodal characteristics according to claim 1, characterized in that: Specifically, S2 is: S21: Preprocess the target data, converting it into a format suitable for the XLNet model; perform sequence labeling for a given protein sequence. Use one-hot encoding for each Mapped to a unique tag ; S22: For the labeled sequence Convert to embedded representation: in It is the interrelationship of each amino acid in sequence T. It is the corresponding position embedding; S23: Feed the embedded representation into the self-attention layer; for each location Calculate its query ,key Sum : ;in There are three learnable weight matrices; a self-attention mechanism is used to obtain a globally sensitive representation for each position. ,in It is the dimension of the key K vector; The layer output is obtained through two inter-layer interactions and a feedforward network; it is then processed using LayerNorm, and then LayerNorm is applied again through a feedforward neural network (FFN). This process can be represented as follows: in The specific calculation method is as follows: , It is input. and They are The mean and variance, Values arrive between, and These are learnable parameters used to recover the proportions and positions of normalized data; S24: Will Depth-separable convolutions ,in This represents pointwise convolution. This represents depthwise convolution.

3. The drug target binding affinity prediction method based on drug bimodal characteristics according to claim 2, characterized in that: Specifically, S3 is: S3: Constructing a drug molecule graph feature extraction network The specific steps are as follows: S31: Use the cheminformatics tool RDKit to parse the SMILES sequence of compound; convert the SMILES string into a graphical representation of the chemical structure, i.e., a molecular diagram. ,in It is a set of nodes, that is, atoms. It is an edge set, i.e., a chemical bond; S32: Molecular diagram Input message passing neural network, each node of the molecular graph Features Updated in each round of message passing as follows: Among them It is a node The set of neighboring nodes, It is a node neighboring nodes The current characteristics, It is an average aggregation function. Update for fully connected layer; S33: Molecular diagram The input graph isomorphic network is represented as: in It is a multilayer perceptron. These are learnable parameters; S34: The features updated via MPNN are represented as follows The features updated via GIN are represented as The combined features are calculated using a weighted summation method. , to be represented as: , It is a weight coefficient that can learn between 0 and 1; the graph The global information of the entire graph is obtained by summing the feature vectors of all nodes using the formula, and is represented as follows: .

4. The drug target binding affinity prediction method based on drug dual-modal characteristics according to claim 1, characterized in that: Specifically, S4 is: S41: Drug molecules The input embedding layer obtains a fixed-length vector, represented as ; S42: Embed the vector Input a bidirectional RNN and compute the forward feeds respectively. and backward Hidden states; then the forward and backward hidden states are concatenated to form the bidirectional RNN output at each time step. ; then Input convolutional layer, set the kernel size to In position The convolution operation on is represented as , It is the weight matrix of the convolution kernel. It is a bias term. It is an activation function; S43: Output the convolutional layer The input max-layer reduces the feature dimensionality and captures the most important features, represented as follows: Subsequently, the drug sequence features are obtained by inputting them into a fully connected layer. .

Images