Drug target affinity prediction method based on graph representation and attention fusion mechanism

Abstract

The invention provides a drug target affinity prediction method based on graph representation and an attention fusion mechanism. According to the method, SMILES character strings of drug molecules are converted into molecular graphs, and graph representation of drug characteristics is generated; and inferring residue contact information of the target by using an ESM-1b model, constructing a weighted residue contact graph, and obtaining sequence-level embedding through average residue embedding. Then, a deep graph isomorphic network (GIN) is used, and a graph neural network module combining graph convolution (GCN) and graph attention (GAT) is used for extracting feature information of the medicine and the target; in the feature fusion stage, an attention fusion mechanism is introduced, and the complex interaction mode between the drug and the target is captured by dynamically adjusting the importance of the features. And finally, processing the fused features by using a multi-layer sensor (MLP), and predicting the affinity of the drug and the target. According to the method, the deep learning technology and the attention fusion strategy are combined, so that the accuracy of drug target affinity prediction is effectively improved.

Description

Technical Field

[0001] The present invention relates to the fields of drug research and development and computer-aided drug design. Specifically, it relates to a method for predicting drug-target affinity based on graph representation and attention fusion mechanism. Background Art

[0002] In the process of drug discovery, drug-target affinity prediction (DTA) is one of the key steps, and its core lies in accurately evaluating the interaction strength between a drug (ligand) and a target (usually a protein). Traditional computational methods, such as molecular docking, although able to precisely simulate the binding mechanism between a drug and a target, are often limited in terms of computational efficiency and accuracy when dealing with complex and variable biological macromolecules. In recent years, with the rapid progress of information technology, computer-aided drug design (CADD) has developed rapidly. With the help of advanced computational algorithms, CADD can efficiently identify and simulate the interaction patterns between drugs and target proteins, providing a chemical basis for the structural optimization of small molecule compounds, thereby reducing the trial-and-error costs in the drug research and development process and accelerating the process of new drug research and development.

[0003] The binding strength between a drug and a target is usually studied using regression analysis methods, and the strength of the interaction is commonly measured by values such as inhibition constant (K i ), dissociation constant (K d ), and half-maximal inhibitory concentration (IC 50 ) in experiments. Higher affinity values are not only directly related to the clinical effectiveness of drugs but also allow drugs to maintain efficacy at low doses, thereby reducing the risk of potential side effects. Accurately identifying disease-specific targets is an important challenge and opportunity in the field of drug research and development.

[0004] In the patent document of CN 116665766A, a drug-target binding affinity prediction model and method based on a graph dilation convolution strategy are disclosed. This method extracts features from the topological structure of drug molecules, local chemical properties, and target protein sequences through a multi-channel aggregation network, a multi-layer residual convolution module, and a bidirectional long short-term memory network, and finally inputs them into the prediction module after integration through a fusion layer, effectively improving the accuracy of affinity prediction and the efficiency of the drug redirection process. However, when constructing the association mapping between drugs and targets, this method uses a simple feature splicing strategy, which is difficult to fully capture the complex interaction patterns between the two, may lead to the loss of key biological information, and affect the performance of the prediction model. Summary of the Invention

[0005] The present invention provides a method for predicting drug-target affinity based on graph representation and attention fusion mechanism. By combining deep learning technology and attention fusion strategy, this method effectively improves the accuracy and reliability of drug-target affinity prediction.

[0006] To achieve the above objectives, the technical solution of the present invention is specifically as follows: (1) Use the feature encoding module to perform preliminary encoding on the drug molecule and target protein sequence. Convert the SMILES string of the drug molecule into a molecular graph, extract atomic features, and generate the graph representation of the drug; for the target protein, use the ESM-1b model to infer the residue contact information of the target, construct a weighted residue contact graph, and obtain the sequence-level embedding through average residue embedding; (2) Use the deep graph isomorphism network (GIN) module to extract the feature information of the drug molecular graph, and use the graph neural network module combining graph convolution (GCN) and graph attention (GAT) to extract the feature information of the target contact map; (3) In the feature fusion stage, introduce the attention fusion mechanism, capture the complex interaction patterns between the drug and the target by dynamically adjusting the importance of features. Specifically, input the drug features and target features into the attention network, calculate the attention coefficient weights, and perform weighted summation on the features to obtain the fused features; (4) Input the above fused features into a multi-layer perceptron for regression prediction to obtain the affinity score of the drug target.

[0007] In the feature encoding module described in step (1), the encoding of the drug molecule involves converting the SMILES string into a molecular graph representation through the RDKit tool. Each atom in the drug molecule is regarded as a node, and the chemical bonds between atoms are represented as edges. Extract the chemical properties of the atoms as their feature vectors, and construct an adjacency matrix to represent the connection relationship between atoms.

[0008] The atomic features described in step (1) include five categories: atomic symbol, number of adjacent atoms, number of adjacent hydrogen atoms, valence of the atom, and whether it has an aromatic structure.

[0009] In the feature encoding module described in step (1), the ESM-1b model requires that the input sequence length of the target protein does not exceed 1024. For amino acid sequences with a length exceeding 1000, they are processed by truncation.

[0010] In the feature encoding module described in step (1), the construction of the weighted residue contact map involves setting a threshold of 0.5 based on the contact probability between residues. When the contact probability exceeds the threshold, the relationship between residues is established as a connecting edge, and the contact probability is used as the weight of the connecting edge; if it is below the threshold, it is considered that there is no contact.

[0011] As described in step (2), the deep graph isomorphism network module is used to extract the molecular graph features of the drug. GIN enhances the expression ability of node features by aggregating the features of neighboring nodes and combining its own node features. GIN flexibly adjusts the contribution degrees of the node's own features and neighboring features by introducing a multi-layer perceptron and a learnable parameter ε. The situation of using MLP to update the features of each layer of nodes in GIN is as follows:

[0012] In the formula is the feature representation of node i at the k-th layer, and N(i) is the set of nodes adjacent to node i.

[0013] As described in step (2), the number of GIN layers in the deep graph isomorphism network module is 5, and a batch normalization layer is connected after each layer of GIN. The output of the last layer of GIN is passed to a global max pooling layer, which can effectively ignore the differences in the node set and the number of nodes, thereby generating a graph-level embedding representation D:

[0014]

[0015] As described in step (2), GCN is used to process graph-structured data, and the operation of each layer of convolution is as follows:

[0016] In the formula, H (k) represents the node feature matrix of the k-th layer, is obtained by adding the adjacency matrix A and the identity matrix, is the degree matrix of, w (k) is the weight matrix of the k-th layer, and σ represents the non-linear activation function.

[0017] As described in step (2), GAT dynamically adjusts the importance of neighboring node features through the attention mechanism. The specific formula for the node features in GAT is expressed as:

[0018] In the formula, is the feature vector of node v in the k-th layer, N(v) is the set of neighboring nodes of node v, and W (k) is the weight matrix of the k-th layer. is the attention coefficient between node v and node u, which can be calculated by the following formula:

[0019]

[0020] where a (k) is the attention vector of the k-th layer, denotes the concatenation operation.

[0021] As described in step (2), a strategy combining multiple layers of GCN and GAT is used to extract the feature information of the target contact map. Among them, the number of layers of the graph convolutional network is set to 2, the number of layers of the graph attention network is 1, and a Pooling layer and a fully connected layer are connected thereafter.

[0022] As described in step (3), in the feature fusion stage, an attention fusion mechanism is introduced. With the help of a non-linear activation function and a learnable parametric mapping, the importance of each feature is dynamically adjusted according to the input features. First, we extract the graph representation feature D of the drug from the compound and the structural feature P from the target protein respectively. Next, D and P are input into the attention network to calculate the corresponding attention coefficient weights α, and its calculation formula is as follows:

[0023] α = Sigmoid(attention(D + P)),

[0024] attention(x) = w2(ReLU(w1x + b1) + b2, where w1 and w2 are trainable weights, and b1 and b2 are bias terms. Finally, the drug feature D and the target feature P are weighted and summed with the calculated attention coefficient weights, so as to obtain a feature embedding representation that fuses drug and target information:

[0025] F = D ⊙ α + P ⊙ (1 - α), where ⊙ represents element-wise multiplication.

[0026] As described in step (4), a multi-layer perceptron (MLP) is used to process the fused features. The input layer of the MLP receives the output features from the attention module and consists of 3 fully connected layers. Among them, the first two fully connected layers each use 1024 nodes, and a Dropout layer is connected after each connection layer to prevent the model from overfitting. The third fully connected layer contains 512 nodes, and an output layer without a Dropout layer is connected thereafter. In addition, the ReLU activation function is used in the fully connected layer, and the mean squared error is selected as the loss function, and its calculation formula is:

[0027] where y i represents the actual label value of the i-th sample, represents the predicted value of the model for the i-th sample, and n represents the total number of samples.

[0028] Through the above steps, by combining the graph neural network and the attention fusion mechanism, the present invention can accurately capture the structural information of drugs and targets, and can also effectively fuse the interaction patterns between them, thereby improving the accuracy and reliability of prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a model framework diagram of the drug-target affinity prediction method in the present invention;

[0030] Figure 2 It is a schematic diagram of the update of GIN node features in the present invention;

[0031] Figure 3 It is a structural diagram of the target feature extraction module in the present invention;

[0032] Figure 4 It is a structural diagram of the attention fusion module in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0033] To further elaborate on the drug-target affinity prediction method based on graph representation and attention fusion mechanism, the technical details of each link will be described in detail below in conjunction with the drawings and embodiments. The embodiments are only used to illustrate the present invention and do not limit any other form of the present invention.

[0034] Embodiment 1

[0035] This embodiment mainly includes key steps such as data preprocessing, feature extraction, feature fusion, and regression prediction. The model framework diagram of the drug-target affinity prediction method in the present invention is as Figure 1 shown.

[0036] The data preprocessing includes the construction of the drug molecular graph and the construction of the target protein graph. For drug molecules, first, use RDKit to parse the SMILES string to generate a molecular object. The feature vector of each atom (node) includes the following dimensions: atom type (44 dimensions), number of adjacent atoms (11 dimensions), number of hydrogen atoms (11 dimensions), valence (11 dimensions), and whether it is an aromatic structure (1 dimension). Each atom feature is finally represented as a 78-dimensional binary vector.

[0037] For the preprocessing of the data, for the target protein, the amino acid sequence of the input protein is used, and the ESM-1b pre-trained model (with the parameter version esm1b_t33_650M_UR50S) is used to predict the residue bond contact probability. Each amino acid residue serves as a node in the graph, and the node features generate an embedding vector of 1280 dimensions. If the contact probability between residues is not less than 0.5, an edge is established, and the edge weight is the contact probability value. For sequences longer than 1000, the first 1000 residues are intercepted, the key structural domain is retained, and the sequence-level embedding representation is obtained by averaging the residue embeddings.

[0038] For the feature extraction module, for the feature extraction of the drug molecule graph, the graph isomorphism network is used to extract the graph feature representation of the drug molecule. The schematic diagram of node update for each layer is as Figure 2 . The number of layers of the graph isomorphism network is set to 5 layers, and each layer outputs 128 dimensions. Among them, the ReLU activation function and global max pooling are used, and finally a 128-dimensional graph-level feature vector is generated.

[0039] For the feature extraction module, for the feature extraction of the target protein, a graph neural network combining GCN-GAT is adopted, and its structure is as Figure 3 . Specifically, it consists of 2 layers of GCN and 1 layer of GAT, and a 128-dimensional feature vector is generated after passing through the global pooling layer.

[0040] For the feature fusion module, its attention fusion module is as Figure 4 . Specifically, in the feature fusion stage, the attention mechanism is introduced to dynamically adjust the importance of drug and target features. The graph representation feature D∈R 128 of the input drug and the target feature P∈R 128 . D and P are input into the attention network to calculate the corresponding attention coefficient weights. Then, the features are weighted and summed with the help of the attention coefficient weights to obtain the feature embedding representation of the fused drug target.

[0041] For the regression prediction module, the fused features are input into a multi-layer perceptron to output the affinity score of the drug target. The MLP contains three fully connected layers. The first and second layers have 1024 nodes, the ReLU activation function is used, and a Dropout layer is connected afterwards, with the dropout rate set to 0.5. The third layer has 512 nodes, the ReLU activation function is used, and no Dropout layer is connected afterwards. And the mean squared error is used as the loss function.

[0042] Example 2

[0043] In this example, according to the method described in Example 1, data preparation, model training and optimization, and model evaluation of the drug target affinity prediction method are provided.

[0044] The data preparation mentioned above means selecting a suitable dataset. The dataset should contain the sequence information of drugs and targets, as well as their affinity scores. For example, the initial KIBA dataset contains a large amount of sparse interaction information, and this dataset needs to be filtered. The filtered dataset contains 229 targets and 2111 drug molecules, with a total of 118,254 drug-target interaction pairs. The dataset is divided, and the ratio of the training set to the test set is 4:1.

[0045] For the model training and optimization, according to the steps in Example 1, the drug-target affinity prediction model is trained on the KIBA dataset. During the training process, the Adam optimizer is used, and the learning rate is set to 0.001. Five-fold cross-validation is performed on the dataset, and the average value of the experimental results is taken as the final performance metric.

[0046] For the model evaluation, the concordance index (CI) and the mean squared error (MSE) are used as evaluation metrics to evaluate the accuracy and reliability of the model.

[0047] The concordance index is used to evaluate the consistency between the predicted value and the actual value, and its calculation formula is:

[0048] In the formula, δ x , δ y represents the label value, b x , b y is the corresponding predicted value, Z is the normalization constant. h(x) is the step function, which is defined as follows:

[0049]

[0050] The mean squared error is used to measure the difference between the predicted value and the actual value, and its calculation formula is:

[0051] In the formula, y i represents the label value of the i-th sample, represents the predicted value of the i-th sample, and N represents the number of samples.

[0052] The above embodiments of the present invention are only for clearly explaining the specific implementation schemes of the present invention, and do not limit the implementation manners of the present invention. Without departing from the principle of the present invention, any modifications, equivalent replacements, etc. made to the above embodiments shall be regarded as within the protection scope of the claims of the present invention.

Claims

1. A drug target affinity prediction method based on graph representation and attention fusion mechanism, characterized in that: The following steps are involved: (1) Feature encoding of drug molecules and target protein sequence information through feature encoding module; (2) By representing the input features of drugs and targets as graph data, the feature information of drugs and targets is processed using a deep graph isomorphism network (GIN), a graph neural network module combined with graph convolution (GCN), and graph attention (GAT); (3) In the feature fusion stage, an attention fusion mechanism is introduced to simulate the interaction between drug molecules and target proteins, dynamically adjust the importance of features, and capture the interaction pattern between drugs and targets; (4) Input the above fused features into a multi-layer perceptron for regression prediction.

2. The method according to claim 1, characterized in that The feature encoding module includes converting the SMILES string of the drug molecule into a molecular graph and extracting the atomic features to generate a graph representation of the drug; and using the ESM-1b model to infer the residue contact information of the target protein and construct a weighted residue contact graph.

3. The method according to claim 1, characterized in that The deep graph isomorphism network module is used to extract the molecular structure characteristics of drugs. The number of layers is set to 5, and each layer is connected to a ReLU activation function and a batch normalization layer.

4. The method according to claim 1, characterized in that: The method for constructing the weighted residue contact graph includes: setting a threshold based on the residue contact probability, and when the contact probability exceeds the threshold, establishing the relationship between the residues as a connecting edge, and using the contact probability as the weight of the connecting edge.

5. The method according to claim 1, characterized in that: The graph neural network module is used to extract the structural features of the target protein, wherein the number of layers of the graph convolutional network is set to 2, the number of layers of the graph attention network is 1, followed by a Pooling layer and a fully connected layer.

6. The method according to claim 1, characterized in that The attention mechanism is introduced in the feature fusion stage, and the importance of features is dynamically adjusted through nonlinear activation functions and learnable parameterized mappings to capture the complex interaction patterns between drugs and targets.

7. The method according to claim 1, characterized in that The multilayer perceptron includes three layers, wherein the first two layers are each equipped with 1024 nodes, and a Dropout layer is set after each layer to prevent overfitting; the third layer contains 512 nodes, followed by an output layer without Dropout.

Images