Drug target affinity prediction method based on multi-channel information fusion

Abstract

The invention provides a drug target affinity prediction method based on multi-channel information fusion. According to the method, chemical information coding, a deep learning technology and a feature fusion strategy are comprehensively applied, a Transform module and a multi-layer graph isomorphic network module (MGIN) are respectively adopted to carry out feature extraction on chemical information and molecular structures of drugs, and sequence features of targets are learned through a convolutional neural network layer (CNNs) containing three branches. And then, fusing the extracted local chemical features, global structural features and target sequence features of the drug through a feature fusion module to form multi-channel interactive fusion drug target joint characterization. And finally, inputting the joint representation into a prediction module for regression prediction. According to the method, through bimodal feature representation of drug molecules and multi-channel information interaction of the drug targets, a more comprehensive and information-rich molecular structure coding mode is realized, and the accuracy of predicting the affinity of the drug targets is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a drug target affinity prediction method based on multi-channel information fusion. Background Art

[0002] Drug-target affinity prediction is a key step in drug design and development. Traditional methods usually rely on a single type of feature representation, which makes it difficult to fully capture the complex interactions between drug molecules and targets. In recent years, with the development of deep learning technology, multimodal feature representation and multi-channel information fusion methods have shown great potential in drug-target affinity prediction. However, existing methods still have certain limitations in feature extraction and fusion, and a more efficient and accurate prediction method is urgently needed.

[0003] The targets of drugs are biological macromolecules, including proteins, nucleic acids, etc. Among them, more than 90% of drug targets are protein compounds, and only a few are nucleic acids. Many experiments have confirmed that drug targets are closely related to diseases. As important channels for the metabolism and signal transmission of pathological substances, drug targets regulate specific diseases without causing much impact on normal physiological states or other disease states. Drugs are mainly small molecule chemical drugs, which can change the symptoms of diseases through their biological activities. According to statistics, 70% of drug small molecules target four types of proteins, including GPCRs, ion channels, protein kinases, and nuclear receptors. Drugs transmit chemical signals into cells or nuclei through these proteins to change cell activities. Traditional experimental methods no longer meet the research and development needs of modern medicine. The development of bioinformatics provides an opportunity to use computer technology to screen and discover interactions between potential drug targets, thereby accelerating the research and development of medicines. In the early stages of deep learning, researchers used hand-crafted descriptors to represent drugs and proteins and designed fully connected neural networks (FCNNs) for prediction. Subsequently, many end-to-end models were proposed. However, the black-box nature of the models limits the interpretability from a biological perspective. They ignore the fact that only specific parts of the protein or specific atoms of the drug are involved in the intermolecular interaction, rather than the entire structure. In order to improve the interpretability of the model, an attention mechanism was introduced in the prediction of drug-target binding affinity (DTA), which is used to simulate the single non-covalent molecular interaction between the drug and the protein.

[0004] A drug target binding affinity prediction model and method based on graph dilation convolution strategy is disclosed in the patent document with publication number CN 116665766 A. The multi-channel general aggregation network and multi-layer residual convolution network and bidirectional long and short cycle memory network modules of the graph dilation convolution strategy are used to extract features of drug molecular structure, local chemical information and target structure respectively. Finally, the extracted features are cascaded and passed through the DTA prediction module to predict the drug target affinity value, which effectively improves the drug target binding affinity prediction accuracy and the success rate of the drug redirection process. However, the complex mapping pattern selection between compounds and proteins may lead to the loss of key features with a simple connection modeling method, thereby reducing the accuracy of affinity prediction. Summary of the invention

[0005] The present invention provides a drug-target affinity prediction method based on multi-channel information fusion, which realizes a more comprehensive and information-rich molecular structure encoding method through multi-modal feature representation and multi-channel information fusion of drug molecules and target information, and effectively improves the prediction accuracy of the affinity between drugs and targets.

[0006] In order to achieve the above purpose, the technical solution of the present invention has the following specific steps: (1) Use the feature encoding module to preliminarily encode the drug molecule and target information. The local chemical information of the drug molecule is encoded using the MACCS descriptor, and the molecular structure information is represented by the molecular graph; the target information is represented by sequence data; (2) The Transformer module is used to perform deep feature extraction on the chemical information of drugs. A lightweight multi-layer graph isomorphism network module (MGIN) is used to extract the molecular structure characteristics of drugs. A convolutional neural network layer (CNNs) with three branches is used to detect the local residue patterns of proteins at different scales and extract the sequence characteristics of the target. (3) The feature fusion module fuses the extracted local chemical features and global structural features of the drug with the target sequence features to form a multi-channel interactive fusion of drug-target joint features; (4) The above combined features are input into the prediction module for regression prediction to obtain the affinity score of the drug target.

[0007] In the feature encoding module described in step (1), the MACCS descriptor is a descriptor that can describe the local chemical environment of a molecule and can effectively capture the chemical properties of the molecule. MACCS quantifies chemical molecules through 166 predefined chemical fragments, which are expressed as follows:

[0008] F MACCS (M) = {m1,m2,...,m 166}, where M is the drug molecule, m i Indicates whether the i-th chemical fragment is present or absent, represented by 1 or 0 respectively.

[0009] In the feature encoding module described in step (1), the encoding of the drug molecular structure involves converting the SMILES (Simplified Molecular Input Line Entry System) string of the drug molecule into a molecular graph representation through the RDKit tool. For each atom (vertex) in the molecular graph, its chemical properties are extracted as its feature vector, and these features include atom type, atomic valence, whether it is in a ring, etc.; according to the molecular structure, an adjacency matrix is ​​constructed to represent the connection relationship between atoms. Through the constructed molecular graph (vertex features and adjacency matrix), the topological information of the drug compound is extracted.

[0010] As described in step (2), the Transformer module is used to perform deep feature extraction on the chemical information of the drug. Transformer is a deep learning model based on the self-attention mechanism, which can capture the long-range dependencies within drug molecules and enhance the diversity of feature representation. The formula is expressed as:

[0011] Among them, Q, K, and V represent query, key, and value matrices respectively, and d k The dimension of the key.

[0012] As described in step (2), a lightweight multi-layer graph isomorphism network module (MGIN) is used to extract the molecular structure features of the drug. MGIN is a graph neural network that can process molecular graph structure data and extract the global structural features of molecules. The number of layers of MGIN is 4, and each layer is followed by a ReLU activation function and a batch normalization layer to improve the accuracy of feature representation. The formula is expressed as:

[0013] H (l+1) =ReLU(BatchNorm(AH (l) W (l) )), where H (l) represents the input feature matrix of the lth layer, W (l) represents the weight matrix of the lth layer, and A is the adjacency matrix.

[0014] As described in step (2), the sequence features of the target are learned through a convolutional network layer (CNNs) with three branches. Each branch is used to detect the local residue pattern of the protein at different scales. Each branch consists of a series of 3×3 convolutional layers, with the number of convolutional layers being 3, 2, and 1 respectively. In each branch, the convolutional features are pooled through the maxpool operation to extract the key information of the target.

[0015] As described in step (3), the feature fusion module fuses the extracted bimodal features of the drug molecules, namely the local chemical features and the global structural features, with the target sequence features to form a four-channel interactive fusion of the drug target joint representation. This process is implemented using two parallel multi-head cross-attention networks, and the formula is:

[0016] F fused = Cross_Attention(F D ,F P ), where F D is the characteristic representation of drug molecules, F P It is the characteristic representation of the target.

[0017] As described in step (4), the joint features are input into the prediction module for regression prediction to obtain the affinity score of the drug target. The prediction module uses a multi-layer perceptron (MLPs) model for regression prediction. The MLPs model consists of three fully connected layers (FCNLayer), each followed by a ReLU and Dropout layer. The calculation formula is:

[0018] y pred =MLP(F), where y pred is the predicted drug target affinity value, and F is the joint feature after the fusion module.

[0019] Through the above steps, the present invention fully utilizes the bimodal feature representation of drug molecules and the multi-channel information interaction of drug targets, realizes a more comprehensive and information-rich molecular structure encoding method, and effectively improves the accuracy of predicting drug target affinity. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of the method for predicting drug target affinity in the present invention;

[0021] Figure 2 The overall framework diagram of the drug target affinity prediction method of the present invention;

[0022] Figure 3 This is the architecture diagram of the Fusion Block in the present invention. DETAILED DESCRIPTION

[0023] To further illustrate the drug target affinity prediction method based on multi-channel information fusion, the specific implementation details are described in detail below in conjunction with the accompanying drawings. The embodiments are only used to illustrate the present invention and are not limited to any other forms according to the present invention.

[0024] Example 1

[0025] This embodiment provides a construction and application of a drug target affinity prediction model, including a feature encoding module, a feature extraction module, a feature fusion module and a regression prediction module. The flowchart of the drug target affinity prediction method in the present invention is as follows: Figure 1 shown.

[0026] The feature encoding module selects the affinity data of known drug molecules and target proteins. The drug molecules are represented in the form of SMILES strings, and the target protein sequences are represented in the form of amino acid sequences. The chemical information of the drug molecules is encoded using MACCS descriptors to form a fixed-length binary vector for describing the chemical structure characteristics of the molecules. The formula used is:

[0027] F MACCS (M) = {m1,m2,...,m 166}, where M is the drug molecule and m is used to describe the chemical structure characteristics of the molecule. The value of m is:

[0028]

[0029] The feature encoding module uses the RDKit library to convert the SMILES string of the drug molecule into a molecular graph representation. The vertex features and adjacency matrix of each drug molecule are obtained. Each vertex corresponds to an atom, and the vertex features include atom type, atomic valence, etc. The molecular graph representation is:

[0030] G=(V,E),v i ∈V,e ij ∈E,v i represents the i-th atom, edge e ij represents the chemical bond between the i-th atom and the j-th atom.

[0031] The feature extraction module is as follows Figure 2 , the Transformer module is used to perform deep feature extraction on the chemical information vector of drug molecules. The Transformer module relies on the self-attention mechanism and can effectively capture the complex patterns in chemical information. The formula is expressed as:

[0032] Among them, Q, K, and V represent query, key, and value matrices respectively, and d k The dimension of the key.

[0033] The feature extraction module is as follows Figure 2 , the drug molecule graph is feature extracted through a lightweight multi-layer graph isomorphism network module (MGIN) to obtain the global structural features of the drug molecule. The number of layers of the graph isomorphism network is 4. The formula is expressed as:

[0034] H (l+1) =ReLU(BatchNorm(AH (l) W (l) )), where H (l) represents the input feature matrix of the lth layer, W (l) represents the weight matrix of the lth layer, and A is the adjacency matrix.

[0035] The feature extraction module is as follows Figure 2 , a convolutional network layer with three different branches is applied to the target protein sequence. Each branch has a different number of convolutional layers. The key information of the sequence is extracted through the maxpool operation to obtain the sequence features of the target. The number of convolutional layers in different branches is 3, 2, and 1 respectively. The output is expressed as:

[0036] T s =W(m(f1(s))||m(f2(s))||m(f3(s))), where s is the embedded representation of the protein, f i is the i-th branch, m represents the maxpool operation, and W is a learnable matrix.

[0037] The feature fusion module is as follows Figure 3 , the local chemical features, global structural features and target sequence features of the extracted drug molecules are fused through two parallel multi-head cross-attention networks to form a four-channel interactive fusion of drug target joint features. The process uses two parallel multi-head cross-attention networks to achieve this, and the formula is:

[0038] F fused = Cross_Attention(F D ,F P ), where F D is the characteristic representation of drug molecules, F P It is the characteristic representation of the target.

[0039] The regression prediction module inputs the joint features into a three-layer multi-layer perceptron (MLPs) model for regression prediction. The prediction module combines the ReLU activation function and the Dropout operation to output the affinity prediction value between the drug and the target.

[0040] Example 2

[0041] This example provides experimental verification of a drug target affinity prediction method based on the method described in Example 1, including experimental preparation, model training, model evaluation and result analysis.

[0042] The experimental preparation is to select a batch of drug molecules and target protein pairs with drug target affinity data from public databases as data sets. These data sets should contain the chemical structure information of the drug and the sequence information of the target, as well as the affinity score between them. The data sets are randomly divided into training set, validation set and test set, with a ratio of 8:1:1. The training set is used for model training, the validation set is used for model tuning, and the test set is used for model evaluation.

[0043] The model training is based on the training set data and follows the steps in Example 1 to construct a drug target affinity prediction model. The model includes feature encoding, feature extraction, feature fusion and regression prediction modules. The validation set is used to adjust and optimize the model parameters. The hyperparameters of the model, such as learning rate, batch size, number of network layers, etc., are adjusted until the performance of the model on the validation set reaches a satisfactory level.

[0044] The model evaluation uses the data of the test set to evaluate the prediction performance of the model. Common evaluation indicators, such as mean square error (MSE) and consistency index (CI), are used to evaluate the accuracy and reliability of the model in predicting drug target affinity.

[0045] The formula for calculating the mean square error is:

[0046] Among them, DTA i and Label i They represent the predicted value and the basic true value respectively, and N represents the number of samples.

[0047] The calculation formula of consistency index is:

[0048] Among them, DTA max and DTA min Represent the highest affinity δ i and the minimum affinity δ j The predicted value of ζ represents the step function, and its value is:

[0049]

[0050] The results analysis compares the model's prediction results with the actual drug target affinity values, analyzes the model's prediction results, explores the model's prediction effect on different types of drugs and target proteins, evaluates the overall generalization ability of the model, and discusses possible improvement directions, such as introducing more data, optimizing the model structure, and adjusting the training strategy.

[0051] The above embodiments of the present invention are only for the purpose of clearly illustrating the specific implementation of the present invention, and are not intended to limit the implementation of the present invention. All modifications, equivalent substitutions and improvements in the above description, as long as they do not depart from the principles of the present invention, should be deemed to be within the scope of protection of the claims of the present invention.

Claims

1. A drug target affinity prediction method based on multi-channel information fusion, characterized in that: The steps include: (1) Feature encoding of drug molecules and target information through feature encoding module; (2) The Transformer module is used to extract the chemical information of the drug. The lightweight multi-layer graph isomorphism network module (MGIN) is used to extract the molecular structure characteristics of the drug. The sequence characteristics of the target are learned through a three-branch convolutional neural network layer (CNNs). (3) The extracted local chemical features, global structural features, and target sequence features of the drug are fused through a feature fusion module to form a multi-channel interactive fusion of drug-target joint features; (4) Input the above joint features into the prediction module for regression prediction.

2. The method according to claim 1, characterized in that The feature encoding module is used to preliminarily encode the drug molecules and target information so that the feature extraction module can process them. The local chemical information of the drug molecules is encoded using MACCS descriptors.

3. The method according to claim 1, characterized in that The Transformer module is used to extract local chemical information of drugs to enhance the diversity of feature representation.

4. The method according to claim 1, characterized in that: The multi-layer graph isomorphism network module (MGIN) is used to extract the global molecular structure features of drugs to improve the accuracy of feature representation. It has 4 layers, each followed by a ReLU activation function and a batch normalization layer.

5. The method according to claim 1, characterized in that The convolutional neural network layer (CNNs) has three branches, which detect local residue patterns of proteins at different scales. Each branch contains a series of 3×3 convolutional layers, and the number of convolutional layers is 3, 2 and 1 respectively. In each branch, the convolutional features are pooled through the maxpool operation to extract the key information of the target.

6. The method according to claim 1, characterized in that The feature fusion module fuses the bimodal feature representation of the drug molecule with the target feature representation through two parallel multi-head cross-attention networks to form a four-channel interactive fusion of the drug target joint representation.

7. The method according to claim 1, characterized in that The prediction module is used to perform regression prediction on the joint features, input the joint representation into a multi-layer perceptron (MLPs) model, and obtain the affinity score of the drug target to improve the accuracy of the drug target affinity prediction.

Images