A drug target binding affinity prediction method and system based on graph virtual nodes

Abstract

The application discloses a drug target affinity prediction method based on a graph virtual node, which introduces virtual nodes and virtual edges in a drug molecule structure graph, takes a graph transformer as a graph feature extractor, takes virtual node features as drug representation, then inputs protein and drug features into a feature fusion module, adds an attention mechanism and a gated skip connection mechanism in feature fusion, captures potential interaction while fusing feature information of different hidden layers, and realizes higher-precision affinity prediction. The application can solve the technical problems of long time consumption and high cost of a drug target affinity prediction method in traditional machine learning, and the technical problems of other existing deep learning that can only transmit neighbor node features and cannot consider global information of drug features, and the technical problem of directly splicing drug and protein features and failing to capture the interaction relationship between the two.

Description

Technical Field

[0001] This invention belongs to the field of biomedical technology, and more specifically, relates to a method and system for predicting drug target affinity using graph virtual nodes. Background Technology

[0002] Drug-target protein interactions play a crucial role in numerous biological processes, including immune responses and gene regulation. Quantitatively determining the binding affinity of drugs to target proteins is essential for effective disease treatment. Traditional high-throughput screening experiments for affinity measurement are time-consuming, labor-intensive, and expensive. To save costs, computational methods for predicting drug target affinity have emerged.

[0003] With the advancement of bioinformatics computing technology, there are currently two methods for predicting drug target affinity: one is the traditional machine learning method, which requires manually creating features and using models such as random forests and support vector machines to predict affinity; the other is a deep learning-based method, which first inputs the sequences of proteins and drugs, and then uses neural networks to model the sequences of drugs and targets to predict affinity.

[0004] However, both of the above methods have some drawbacks that cannot be ignored:

[0005] First, traditional machine learning methods require complex and time-consuming feature engineering, and due to the limited and uneven nature of the dataset, the prediction accuracy is low.

[0006] Second, deep learning-based methods either ignore the role of molecular structural information in predicting drug target affinity or require a large amount of protein structural information, which is difficult to obtain. Therefore, they cannot accurately predict the affinity between proteins lacking structural information and targets.

[0007] Third, because both methods ignore the global characteristics of the drug and lack an effective feature fusion scheme to capture the interaction between the drug and the target, the prediction accuracy is low. Summary of the Invention

[0008] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method and system for predicting drug target affinity based on graph virtual nodes. Its purpose is to solve the technical problems of low prediction accuracy and time-consuming and labor-intensive feature creation in traditional machine learning, the technical problem that proteins lacking structural information cannot accurately predict their affinity for targets due to the use of protein structural information, and the technical problem that deep learning neglects global information about drug features and lacks effective feature fusion schemes to capture the interaction between drugs and targets.

[0009] To achieve the above objectives, according to one aspect of the present invention, a method for predicting drug target affinity based on graph virtual nodes is provided, comprising the following steps:

[0010] (1) Obtain the drug-protein data pair to be predicted, and preprocess the drug-protein data pair to be predicted to obtain the preprocessed drug-protein data pair, which includes the simplified molecular linear input canonical SMILES sequence of the drug, the amino acid sequence of the protein, and the affinity between the two.

[0011] (2) For each protein amino acid sequence in the preprocessed drug-protein data pair obtained in step (1), the amino acid sequence is subjected to length normalization and feature encoding processing in sequence to obtain the protein feature matrix corresponding to the amino acid sequence.

[0012] (3) For each drug SMILES sequence in the preprocessed drug-protein data pair obtained in step (1), the SMILES sequence is converted into a drug molecular structure map, and virtual nodes and feature codes are added to the drug molecular structure map to obtain the drug molecular structure map corresponding to the SMILES sequence.

[0013] (4) Input the protein feature matrix corresponding to the amino acid sequence of the protein in the drug-protein data pair obtained in step (2) and the drug molecular structure diagram corresponding to the SMILES sequence of the drug in the drug-protein data pair obtained in step (3) into the pre-trained drug target binding affinity prediction model to obtain the predicted value of the binding affinity between the drug and the protein in the drug-protein data pair.

[0014] Preferably, in step (1), the drug-protein data pair to be predicted is preprocessed to obtain a preprocessed drug-protein data pair. Specifically, this process involves: first, screening the drug-protein data pair to be predicted and deleting data with missing information to obtain the SMILES sequence of the drug, the amino acid sequence of the protein, and the inhibition and dissociation constants of the drug-protein interaction; then, obtaining the affinity between the drug and protein based on the obtained inhibition and dissociation constants; the affinity calculation process is as follows:

[0015]

[0016] Where K i With K d pK represents the inhibition constant and dissociation constant, respectively. i With PK d These represent the affinity obtained using the inhibition constant and the affinity obtained using the dissociation constant, respectively, and are used as labels for subsequent model training and prediction.

[0017] Finally, the obtained drug SMILES sequences, protein amino acid sequences, and affinity are combined to form preprocessed drug-protein data pairs.

[0018] Preferably, step (2) specifically involves first normalizing the length of the amino acid sequence, representing each amino acid in the normalized amino acid sequence with a unique integer to obtain a 1000-dimensional protein embedding vector, and mapping each integer in the protein embedding vector to a 128-dimensional vector. The vectors corresponding to all amino acids in the amino acid sequence eventually form a protein feature matrix of size 1000×128 corresponding to the amino acid sequence.

[0019] Step (3) specifically involves first converting the SMILES sequence into a drug molecule structure diagram G = (V, E) using the open-source chemical library RDKit, where V represents the set of all atoms in the SMILES sequence and E represents the set of chemical bonds between atoms in the SMILES sequence. Then, a virtual node vn is added to the drug molecule structure diagram G, connecting this virtual node to all atoms in G to form virtual edges. Subsequently, feature encoding is performed on the drug molecule structure diagram after adding the virtual node to obtain an M-dimensional vector and an N-dimensional vector. These two vectors constitute the drug molecule structure diagram corresponding to the SMILES sequence; where M represents the dimension of the initial feature vector corresponding to the virtual nodes included in all atoms in the drug molecule structure diagram, and N represents the dimension of the initial feature vector corresponding to the virtual edges included in all chemical bonds in the drug molecule structure diagram.

[0020] Preferably, the drug target binding affinity prediction model includes a protein encoder based on a one-dimensional convolutional neural network (1DCNN), a drug encoder based on a graph transformer, a feature fusion module, and a multilayer perceptron module.

[0021] The 1DCNN-based protein encoder consists of three convolutional layers and one pooling layer, with the following specific structure:

[0022] The input to the first convolutional layer is a 1×1000×128 dimensional feature vector, which is convolved and activated, and the output is a 1×979×256 dimensional intermediate feature vector.

[0023] The input to the second convolutional layer is the 1×979×256-dimensional intermediate feature vector output by the first convolutional layer. It performs convolution and activation processing on this feature vector, and the output is a 1×979×256-dimensional intermediate feature vector.

[0024] The input to the third convolutional layer is the 1×979×256-dimensional intermediate feature vector output by the second convolutional layer. It performs convolution and activation processing on this feature vector, and the output is a 1×979×128-dimensional intermediate feature vector.

[0025] The input to the pooling layer is the 1×979×128-dimensional intermediate feature vector output from the third convolutional layer. It performs dimensionality reduction on this feature vector and outputs a 128-dimensional protein feature vector.

[0026] The graph-transformer-based drug encoder comprises a line preprocessing layer, ten transformer layers, and a virtual node extraction layer. Its specific structure is as follows:

[0027] The input to the preprocessing layer is the M-dimensional atomic feature vector and the N-dimensional chemical bond feature vector from the drug molecule structure diagram. The atomic feature vector and the chemical bond feature vector are then subjected to mapping processing and Laplace feature position embedding processing, respectively. The output is a 128-dimensional atomic feature vector and a 128-dimensional chemical bond feature vector.

[0028] The input to the first transformer layer is a 128-dimensional atomic feature vector and a 128-dimensional chemical bond feature vector output from the preprocessing layer. It uses attention heads to perform attention calculation (i.e., there are 8 independent attention heads inside the transformer, and the attention score between two atoms is calculated through the attention mechanism in each attention head), message passing processing (i.e., by aggregating the current features of adjacent atoms and updating the atomic features by weighting according to the attention score to obtain a 16-dimensional intermediate atomic feature vector, and multiplying it with the attention score to obtain a 16-dimensional intermediate chemical bond feature vector), and residual summation processing. The output is a 128-dimensional atomic feature vector and a 128-dimensional chemical bond feature vector.

[0029] The input to the second transformer layer is the 128-dimensional atomic feature vector and the 128-dimensional chemical bond feature vector output by the first transformer layer. It uses an attention head to perform attention calculation, message passing and residual summation on the atomic feature vector and the chemical bond feature vector respectively, and outputs the 128-dimensional atomic feature vector and the 128-dimensional chemical bond feature vector respectively.

[0030] …

[0031] Similarly, the output of the tenth transformer layer is a 128-dimensional atomic feature vector and a 128-dimensional chemical bond feature vector.

[0032] The input to the virtual node extraction layer is a 128-dimensional atomic feature vector output from the tenth transformer layer. It performs lookup processing on this atomic feature vector and outputs a 128-dimensional drug feature vector.

[0033] Preferably, the feature fusion module includes two linear fusion layers and a gated skip mechanism layer, and its specific structure is as follows:

[0034] The input to the first linear fusion layer is a 128-dimensional protein feature vector output from the pooling layer of the protein encoder and a 128-dimensional drug feature vector output from the virtual node extraction layer of the drug encoder. Attention calculation and weighted summation are performed on the protein feature vector and the drug feature vector, and the output is a 128-dimensional intermediate fusion feature vector.

[0035] The input to the second linear fusion layer is the 128-dimensional intermediate fusion feature vector output by the first linear fusion layer. It performs attention calculation and weighted summation on the intermediate fusion feature vector, and the output is a 128-dimensional intermediate fusion feature vector.

[0036] The input to the gated skip mechanism layer is the 128-dimensional intermediate fusion feature vector output by the second linear fusion layer. It performs attention calculation and weighted summation on the intermediate fusion feature vector, and the output is the 128-dimensional final fusion feature vector.

[0037] The multilayer perceptron module consists of four fully connected layers, and its specific structure is as follows:

[0038] The input to the first linear layer is the 128-dimensional final fused feature vector output by the gated skip mechanism layer. It performs linear regression and activation processing on the final fused feature vector, and the output is a 1024-dimensional intermediate feature vector.

[0039] The input to the second linear layer is the 128-dimensional intermediate feature vector output by the first linear layer. It performs linear regression and activation processing on the intermediate feature vector, and the output is a 1024-dimensional intermediate feature vector.

[0040] The input to the third linear layer is the 1024-dimensional intermediate feature vector output by the second linear layer. It performs linear regression and activation processing on the intermediate feature vector, and the output is a 256-dimensional intermediate feature vector.

[0041] The input to the fourth linear layer is the 256-dimensional intermediate feature vector output by the third linear layer. It performs linear regression and activation processing on the intermediate feature vector, and the output is a 1-dimensional affinity prediction value.

[0042] Preferably, the drug target binding affinity prediction model is obtained through the following steps:

[0043] (4-1) Obtain multiple drug-protein data pairs to be predicted, preprocess all drug-protein data pairs to be predicted to obtain preprocessed drug-protein data pairs (which include the SMILES sequence of the drug, the amino acid sequence of the protein and the affinity between the two), and divide all preprocessed drug-protein data pairs into training set and test set in a 4:1 ratio.

[0044] (4-2) For each drug-protein data pair in the training set obtained in step (4-1), each amino acid sequence is subjected to length normalization and feature encoding in sequence to obtain the protein feature matrix corresponding to the amino acid sequence.

[0045] (4-3) For each drug-protein data pair in the training set obtained in step (4-1), each SMILES sequence is converted into a drug molecule structure diagram, and virtual nodes and feature codes are added to the drug molecule structure diagram to obtain the drug molecule structure diagram corresponding to the SMILES sequence.

[0046] (4-4) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the protein feature matrix corresponding to the amino acid sequence obtained in step (4-2) is input into the first convolutional layer in the protein encoder of the drug target affinity prediction model for convolution processing to obtain a 1×979×256 dimensional feature vector. The ReLU activation function is then used to activate the feature vector to obtain an intermediate feature vector with a dimension of 1×979×256 corresponding to the amino acid sequence.

[0047] (4-5) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate feature vector corresponding to the amino acid sequence obtained in step (4-4) is input into the second convolutional layer in the protein encoder of the drug target affinity prediction model for convolution processing to obtain a 1×979×256 dimensional feature vector. The ReLU activation function is then used to activate the feature vector to obtain the intermediate feature vector corresponding to the amino acid sequence with a dimension of 1×979×256.

[0048] (4-6) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate feature vector corresponding to the amino acid sequence obtained in step (4-5) is input into the third convolutional layer of the protein encoder in the drug target affinity prediction model for convolution processing to obtain a 1×979×128 dimensional feature vector. The ReLU activation function is then used to activate the feature vector to obtain an intermediate feature vector with a dimension of 1×979×256 corresponding to the amino acid sequence.

[0049] (4-7) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate feature vector corresponding to the amino acid sequence obtained in step (4-6) is input into the pooling layer of the protein encoder for dimensionality reduction to obtain a 128-dimensional protein feature vector e. t .

[0050] (4-8) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the M-dimensional atomic feature vector and N-dimensional chemical bond feature vector of the drug molecule structure diagram corresponding to the SMILES sequence obtained in step (4-3) are input into the preprocessing layer of the drug encoder for mapping processing to obtain the 128-dimensional intermediate atom feature vector of the i-th atom. And the 128-dimensional chemical bond feature vector e between the i-th atom and the j-th atom ij (0) , where i∈[1, the total number of atoms in the drug molecule structure diagram corresponding to the SMILES sequence], j∈[1, the total number of atoms in the drug molecule structure diagram corresponding to the SMILES sequence].

[0051] (4-9) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate atom feature vector of the i-th atom obtained in step (4-8) is... The preprocessing layer of the input drug encoder is subjected to Laplacian feature position embedding to obtain the processed 128-dimensional atomic feature vector corresponding to the i-th atom.

[0052] (4-10) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic feature vector corresponding to the i-th atom obtained in step (4-9) is... The chemical bond feature vector e between the i-th atom and the j-th atom obtained in steps (4-8) ij (0)The first transformer layer of the input drug encoder is used to calculate the attention score to obtain the attention score between the i-th atom and the j-th atom.

[0053] (4-11) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the attention scores corresponding to the i-th atom and the j-th atom obtained in step (4-10) are... The first transformer layer of the input drug encoder undergoes message passing processing to obtain the updated intermediate atom feature vector corresponding to the i-th atom. and the intermediate chemical bond feature vectors corresponding to the i-th atom and the j-th atom

[0054] (4-12) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate atom feature vector corresponding to the i-th atom obtained in step (4-11) is... and the intermediate chemical bond feature vectors corresponding to the i-th atom and the j-th atom The first transformer layer of the drug encoder is input to concatenate the intermediate atom feature vectors and intermediate chemical bond feature vectors obtained from all attention heads in the first transformer layer, and then the concatenation result is compared with the original atom feature vectors input to the transformer layer. and chemical bond eigenvectors Add them together to obtain the intermediate atom feature vector corresponding to the i-th atom after processing. and intermediate chemical bond eigenvectors

[0055] (4-13) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate atom feature vector corresponding to the i-th atom obtained in step (4-12) is... and intermediate chemical bond eigenvectors Perform residual summation separately to obtain the atomic feature vector corresponding to the i-th atom after processing. and the chemical bond feature vectors corresponding to the i-th atom and the j-th atom.

[0056] (4-14) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic feature vector corresponding to the i-th atom obtained in step (4-13) is... and the chemical bond feature vectors corresponding to the i-th atom and the j-th atom. The second transformer layer of the input drug encoder undergoes the same processing as steps (4-10) to (4-14) above to obtain the feature vector corresponding to the i-th atom in the final output of the i-th transformer layer. and the chemical bond feature vectors corresponding to the i-th atom and the j-th atom. ...and so on, until finally, after processing by the tenth transformer layer, we obtain the atomic feature vector corresponding to the i-th atom in the final output of the tenth transformer layer. and the chemical bond feature vectors corresponding to the i-th atom and the j-th atom.

[0057] (4-15) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic feature vector corresponding to the i-th atom obtained in step (4-14) is... The virtual node extraction layer of the input drug encoder performs a lookup operation (by using the atom index, finding the feature vector of the virtual node represented by the last atom, and treating it as the feature representation of the entire drug molecule structure map) to obtain a 128-dimensional drug feature vector e. d .

[0058] (4-16) The protein feature vector e obtained in step (4-7) t The drug feature vector e obtained in step (4-15) d Attention calculation is performed in the first linear fusion layer of the input feature fusion module to obtain an attention score W1, and the attention score W1 is then compared with the protein feature vector e obtained in step (4-7). t The drug feature vector e obtained in step (4-15) d A weighted summation process is performed to obtain the first fusion feature.

[0059] (4-17) The first fusion feature obtained in step (4-16) Attention calculation is performed in the second linear fusion layer of the input feature fusion module to obtain an attention score W2, and this attention score W2 is then compared with the protein feature vector e obtained in step (4-5). t The drug feature vector e obtained in step (4-12) d A weighted summation process is performed to obtain the second fusion feature.

[0060] (4-18) The second fusion feature obtained in step (4-17) In the gated skipping mechanism layer of the input feature fusion module, an attention score W3 is obtained, and the attention score W3 and the drug feature vector e are compared. dand protein feature vector e t and fusion features A weighted summation process is performed to obtain the final fused feature vector e of the drug and protein target. dt .

[0061] (4-19) The fusion feature e obtained in step (4-20) dt Input to the multilayer perceptron module to obtain the predicted value of affinity.

[0062] (4-20) Based on the drug-protein data obtained in step (4-1), the affinity between the drug and protein is compared with the predicted value of the affinity obtained in step (4-19). Calculate the loss function and use it to train the drug target binding affinity prediction model until the model converges, thus obtaining a preliminarily trained drug target binding affinity prediction model.

[0063] (4-21) Use the test set obtained in step (4-1) to test the drug target binding affinity prediction model initially trained in step (4-19) until the value of the loss function reaches the optimal value, thereby obtaining the trained drug target binding affinity prediction model.

[0064] Preferably, step (4-9) specifically involves calculating the symmetric normalized Laplace matrix L using the identity matrix I, degree matrix D, and adjacency matrix A of the drug molecule structure diagram. The calculation formula is as follows: Then, the Laplacian position eigenvector is calculated using the following formula: Where λ i are the eigenvectors of the Laplacian matrix L, C are the learnable parameters, and b C Bias term. Next, will With atomic characteristics Add them together to obtain the processed 128-dimensional atomic feature vector corresponding to the i-th atom. The calculation formula is as follows:

[0065] The specific calculation formula in step (4-10) is as follows:

[0066]

[0067] in Let be the linear transformation matrices of the eigenvectors of the i-th and j-th atoms respectively during the processing of the first transformer. d is the linear transformation matrix of the eigenvectors of the chemical bond between the i-th and j-th atoms in the first transformer layer. h The linear transformation matrix The dimension of the eigenvectors.

[0068] Preferably, in step (4-11), for message passing of an atom, the current features of adjacent atoms (considered as the j-th atom) are aggregated and sorted according to attention scores. The eigenvector of the i-th atom is updated using weighted methods. The specific calculation process is as follows:

[0069]

[0070] Where N i Let represent the set of adjacent atoms of the i-th atom in the SMILES sequence. Let ω be the linear transformation matrix of the eigenvectors of the adjacent atoms of the i-th atom in the first transformer layer. ij yes Attention score after activation by the Softmax function.

[0071] For message passing in chemical bonds, directly Assigned as the intermediate chemical bond feature vector Right now

[0072] The calculation process in step (4-13) is as follows:

[0073]

[0074] Where Norm represents normalization. and Represents the learnable parameters, and ReLU represents the activation function.

[0075] Preferably, the calculation process in step (4-16) is as follows:

[0076] W1 = Sigmoid(Linear(Linear(e t +e d )))

[0077]

[0078] Here, Linear represents linear regression and Sigmoid is the activation function.

[0079] The calculation process in step (4-17) is as follows:

[0080]

[0081] The specific steps of the calculation process in step (4-18) are as follows:

[0082]

[0083] The calculation process for step (4-19) is as follows:

[0084]

[0085] The formula for calculating the loss function in step (4-20) is as follows:

[0086]

[0087] Where y represents the affinity between the drug and protein in the drug-protein data pair obtained in step (4-1).

[0088] According to another aspect of the present invention, a drug target binding affinity prediction system based on graph virtual nodes is provided, comprising:

[0089] The first module is used to acquire the drug-protein data pair to be predicted, and to preprocess the drug-protein data pair to be predicted to obtain the preprocessed drug-protein data pair, which includes the simplified molecular linear input canonical (SMILES) sequence of the drug, the amino acid sequence of the protein, and the affinity between the two.

[0090] The second module is used to perform length normalization and feature encoding on the amino acid sequence of each protein in the preprocessed drug-protein data pair obtained by the first module, so as to obtain the protein feature matrix corresponding to the amino acid sequence.

[0091] The third module is used to convert the SMILES sequence of each drug in the preprocessed drug-protein data pair obtained by the first module into a drug molecular structure map, and to add virtual nodes and feature codes to the drug molecular structure map to obtain the drug molecular structure map corresponding to the SMILES sequence.

[0092] The fourth module is used to input the protein feature matrix corresponding to the amino acid sequence of the protein in the drug-protein data pair obtained in the second module, and the drug molecular structure map corresponding to the SMILES sequence of the drug in the drug-protein data pair obtained in the third module into a pre-trained drug target binding affinity prediction model, so as to obtain the predicted value of the binding affinity between the drug and the protein in the drug-protein data pair.

[0093] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0094] (1) Since the present invention adopts steps (2) to (4), it extracts drug and protein features through deep learning. Deep learning can automatically learn potential features and drug target correlation during the training process. Therefore, it can solve the technical problems of existing traditional machine learning methods that require complex and time-consuming feature engineering and have low prediction accuracy due to limited and uneven datasets.

[0095] (2) The present invention employs steps (3) and (4), which convert the SMILES sequence of the drug into a molecular structure diagram and use a graph transformer to extract drug features. It takes into account the structural information of the molecule and provides a global receptive field to extract drug molecular structure diagram features more effectively. Therefore, it can solve the technical problem of ignoring the structural information and global information of the drug in deep learning.

[0096] (3) Since the present invention adopts steps (3) to (4), it proposes the idea of ​​virtual nodes and uses the output virtual node features as the final drug representation features. The virtual nodes are connected to the real atomic nodes through virtual edges. They can automatically learn all the high-dimensional atomic feature information in graph convolution. The virtual node features as graph representation have stronger robustness and are suitable for various downstream tasks. Therefore, it can solve the technical problems of existing methods failing to extract the global features of drugs well and having poor robustness for downstream tasks.

[0097] (4) Since the present invention uses steps (4-4) to (4-7), it extracts protein sequence features through a one-dimensional convolutional neural network 1DCNN. Only the amino acid sequence of the protein is needed to effectively extract sequence features. Therefore, it can solve the technical problem that a large amount of protein structural information is required, but such protein structural information is difficult to obtain, which leads to the inability to accurately predict the affinity between proteins lacking structural information and targets.

[0098] (5) By adopting steps (4-16) to (4-18), this invention proposes an attention-based linear feature fusion module and adds a gated skip connection mechanism to the feature fusion. While capturing the interaction, it fuses the feature information of different hidden layers, which can extract the potential association information between the drug and the target protein and improve the affinity prediction accuracy. Therefore, it can solve the technical problem that existing methods lack an effective feature fusion scheme to capture the interaction between the drug and the target, resulting in low prediction accuracy. Attached Figure Description

[0099] Figure 1 This is an overall flowchart of the drug target binding affinity prediction method based on graph virtual nodes of the present invention;

[0100] Figure 2This is a schematic diagram of the network structure of the drug target binding affinity prediction model used in the method of this invention. Detailed Implementation

[0101] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0102] The basic idea of ​​this invention is to address the above-mentioned deficiencies or improvement needs of existing technologies by providing a drug target affinity prediction method and system based on graph virtual nodes. This invention proposes a method for constructing initial features of drugs and target proteins, proposes using 1DCNN to extract drug sequence features, and designs a drug feature extraction method based on graph virtual nodes. Virtual nodes and virtual edges are introduced into the drug molecular structure graph. A Graph Transformer is used as the graph feature extractor, and virtual node features are used as drug representations. Then, protein and drug features are input into a feature fusion module. The feature fusion incorporates attention and gated skip connection mechanisms to capture potential interactions while fusing feature information from different hidden layers, achieving higher accuracy affinity prediction.

[0103] The purpose of this invention is to solve the technical problems of long time consumption and high cost in traditional machine learning methods for predicting drug target affinity, the technical problems of existing deep learning methods that can only transmit the features of neighboring nodes and cannot consider the global information of drug features, and the technical problems of directly splicing drug and protein features and failing to capture the interaction between them.

[0104] The advantage of this invention is that it improves the performance of predicting drug target binding affinity and can be used for large-scale drug screening.

[0105] like Figure 1 As shown, this invention provides a method for predicting drug target binding affinity based on graph virtual nodes, comprising the following steps:

[0106] (1) Obtain the drug-protein data pair to be predicted, and preprocess the drug-protein data pair to be predicted to obtain the preprocessed drug-protein data pair, which includes the simplified molecular input line entry system (SMILES) sequence of the drug, the amino acid sequence of the protein, and the affinity between the two.

[0107] In this step, the drug-protein data pair to be predicted is preprocessed to obtain a preprocessed drug-protein data pair. Specifically, this process involves first screening the drug-protein data pair to be predicted, removing data with missing information to obtain the drug's SMILES sequence, the protein's amino acid sequence, and the inhibition and dissociation constants of the drug-protein interaction. Then, the affinity between the drug and protein is obtained based on the obtained inhibition and dissociation constants. The affinity calculation process is as follows:

[0108]

[0109] Where K i With K d pK represents the inhibition constant and dissociation constant, respectively. i With PK d These represent the affinity obtained using the inhibition constant and the affinity obtained using the dissociation constant, respectively, and are used as labels for subsequent model training and prediction.

[0110] Finally, the obtained drug SMILES sequences, protein amino acid sequences, and affinity are combined to form preprocessed drug-protein data pairs.

[0111] The advantage of this step (1) is that it can map the excessively large or small inhibition constant and dissociation constant to a moderately large affinity value, which makes it easier for the model to be trained.

[0112] (2) For each protein amino acid sequence in the preprocessed drug-protein data pair obtained in step (1), the amino acid sequence is subjected to length normalization and feature encoding processing in sequence to obtain the protein feature matrix corresponding to the amino acid sequence.

[0113] Specifically, this step involves first normalizing the length of the amino acid sequence (i.e., truncating amino acid sequences longer than 1000 and padding with zeros for sequences shorter than 1000). Then, each amino acid in the normalized sequence is represented by a unique integer (i.e., each amino acid is abbreviated as a letter, and the 26 letters correspond to integers from 1 to 26 in sequence; for example, the abbreviation "A" for alanine is represented by "1", the abbreviation "C" for cysteine ​​is represented by "3", and so on). This yields a 1000-dimensional protein embedding vector. Each integer in this protein embedding vector is then mapped to a 128-dimensional vector. The vectors corresponding to all amino acids in the sequence ultimately form a 1000×128-dimensional protein feature matrix corresponding to the amino acid sequence.

[0114] (3) For each drug SMILES sequence in the preprocessed drug-protein data pair obtained in step (1), the SMILES sequence is converted into a drug molecular structure map, and virtual nodes and feature codes are added to the drug molecular structure map to obtain the drug molecular structure map corresponding to the SMILES sequence.

[0115] Specifically, this step involves first using the open-source chemical library RDKit to convert the SMILES sequence into a drug molecule structure diagram G = (V, E), where V represents the set of all atoms in the SMILES sequence and E represents the set of chemical bonds between atoms in the SMILES sequence. Then, a virtual node vn is added to the drug molecule structure diagram G, connecting the virtual node to all atoms in the drug molecule structure diagram G to form virtual edges. Subsequently, feature encoding is performed on the drug molecule structure diagram after adding the virtual node (the initial feature vector of each atom in the drug molecule structure diagram is determined according to its sign, formal charge, atomic hybridization, atomic chirality, etc., and the initial feature vector of each chemical bond comes from the properties of the chemical bond, such as bond type, aromaticity, or conjugation), to obtain an M-dimensional vector and an N-dimensional vector. The two vectors constitute the drug molecule structure diagram corresponding to the SMILES sequence; where M represents the dimension of the initial feature vector corresponding to the virtual nodes included by all atoms in the drug molecule structure diagram (which is equal to 44 in this invention), and N represents the dimension of the initial feature vector corresponding to the virtual edges included by all chemical bonds in the drug molecule structure diagram (which is equal to 10 in this invention).

[0116] The advantage of this step (3) is that it converts the drug's SMILES sequence into a drug molecular structure diagram to fully consider the structural information of the molecule, and proposes the idea of ​​virtual nodes so that the model can extract the global features of the drug and improve the model's performance.

[0117] (4) Input the protein feature matrix corresponding to the amino acid sequence of the protein in the drug-protein data pair obtained in step (2) and the drug molecular structure diagram corresponding to the SMILES sequence of the drug in the drug-protein data pair obtained in step (3) into the pre-trained drug target binding affinity prediction model to obtain the predicted value of the binding affinity between the drug and the protein in the drug-protein data pair.

[0118] Specifically, the drug target binding affinity prediction model of the present invention is obtained through training the following steps:

[0119] (4-1) Obtain multiple drug-protein data pairs to be predicted, preprocess all drug-protein data pairs to be predicted to obtain preprocessed drug-protein data pairs (which include the SMILES sequence of the drug, the amino acid sequence of the protein and the affinity between them), and then divide all preprocessed drug-protein data pairs into training set and test set in a 4:1 ratio.

[0120] (4-2) For each drug-protein data pair in the training set obtained in step (4-1), each amino acid sequence is subjected to length normalization and feature encoding in sequence to obtain the protein feature matrix corresponding to the amino acid sequence.

[0121] The process of this step is exactly the same as step (2) above, and will not be repeated here;

[0122] (4-3) For each drug-protein data pair in the training set obtained in step (4-1), each SMILES sequence is converted into a drug molecule structure diagram, and virtual nodes and feature codes are added to the drug molecule structure diagram to obtain the drug molecule structure diagram corresponding to the SMILES sequence.

[0123] The process of this step is exactly the same as step (3) above, and will not be repeated here;

[0124] (4-4) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the protein feature matrix corresponding to the amino acid sequence obtained in step (4-2) is input into the first convolutional layer in the protein encoder of the drug target affinity prediction model for convolution processing to obtain a 1×979×256 dimensional feature vector. The ReLU activation function is then used to activate the feature vector to obtain an intermediate feature vector with a dimension of 1×979×256 corresponding to the amino acid sequence.

[0125] (4-5) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate feature vector corresponding to the amino acid sequence obtained in step (4-4) is input into the second convolutional layer in the protein encoder of the drug target affinity prediction model for convolution processing to obtain a 1×979×256 dimensional feature vector. The ReLU activation function is then used to activate the feature vector to obtain the intermediate feature vector corresponding to the amino acid sequence with a dimension of 1×979×256.

[0126] (4-6) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate feature vector corresponding to the amino acid sequence obtained in step (4-5) is input into the third convolutional layer of the protein encoder in the drug target affinity prediction model for convolution processing to obtain a 1×979×128 dimensional feature vector. The ReLU activation function is then used to activate the feature vector to obtain an intermediate feature vector with a dimension of 1×979×256 corresponding to the amino acid sequence.

[0127] (4-7) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate feature vector corresponding to the amino acid sequence obtained in step (4-6) is input into the pooling layer of the protein encoder for dimensionality reduction to obtain a 128-dimensional protein feature vector e. t .

[0128] The advantage of the above sub-steps (4-4) to (4-7) is that the convolutional neural network can learn the features of the protein sequence, thereby capturing high-level semantic information and ensuring that the model can learn more effective information.

[0129] (4-8) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the M-dimensional atomic feature vector and N-dimensional chemical bond feature vector of the drug molecule structure diagram corresponding to the SMILES sequence obtained in step (4-3) are input into the preprocessing layer of the drug encoder for mapping processing to obtain the 128-dimensional intermediate atom feature vector of the i-th atom. And the 128-dimensional chemical bond feature vector e between the i-th atom and the j-th atom ij (0) , where i∈[1, the total number of atoms in the drug molecule structure diagram corresponding to the SMILES sequence], j∈[1, the total number of atoms in the drug molecule structure diagram corresponding to the SMILES sequence].

[0130] (4-9) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate atom feature vector of the i-th atom obtained in step (4-8) is... The preprocessing layer of the input drug encoder is subjected to Laplacian feature position embedding to obtain the processed 128-dimensional atomic feature vector corresponding to the i-th atom.

[0131] Specifically, this step involves calculating the symmetric normalized Laplacian matrix L using the identity matrix I, degree matrix D, and adjacency matrix A of the drug molecule structure diagram. The calculation formula is as follows: Then, the Laplacian position eigenvector is calculated using the following formula: Where λ i are the eigenvectors of the Laplacian matrix L, C are the learnable parameters, and b C Bias term. Next, will With atomic characteristics Add them together to obtain the processed 128-dimensional atomic feature vector corresponding to the i-th atom. The calculation formula is as follows:

[0132] (4-10) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic feature vector corresponding to the i-th atom obtained in step (4-9) is... The chemical bond feature vector e between the i-th atom and the j-th atom obtained in steps (4-8) ij (0) The first transformer layer of the input drug encoder is used to calculate the attention score to obtain the attention score between the i-th atom and the j-th atom.

[0133] The specific calculation formula in this step is as follows:

[0134]

[0135] in Let be the linear transformation matrices of the eigenvectors of the i-th and j-th atoms respectively during the processing of the first transformer. d is the linear transformation matrix of the eigenvectors of the chemical bond between the i-th and j-th atoms in the first transformer layer. h The linear transformation matrix The dimension of the eigenvectors.

[0136] (4-11) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the attention scores corresponding to the i-th atom and the j-th atom obtained in step (4-10) are... The first transformer layer of the input drug encoder undergoes message passing processing to obtain the updated intermediate atom feature vector corresponding to the i-th atom. and the intermediate chemical bond feature vectors corresponding to the i-th atom and the j-th atom

[0137] Specifically, for message passing of atoms, the current features of adjacent atoms (considered as the j-th atom) are aggregated and sorted by attention score. The eigenvector of the i-th atom is updated using weighted methods. The specific calculation process for this step is as follows:

[0138]

[0139] Where N i Let represent the set of adjacent atoms of the i-th atom in the SMILES sequence. Let ω be the linear transformation matrix of the eigenvectors of the adjacent atoms of the i-th atom in the first transformer layer. ij yes Attention score after activation by the Softmax function.

[0140] For message passing in chemical bonds, directly Assigned as the intermediate chemical bond feature vector Right now

[0141] (4-12) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate atom feature vector corresponding to the i-th atom obtained in step (4-11) is... and the intermediate chemical bond feature vectors corresponding to the i-th atom and the j-th atom The first transformer layer of the drug encoder is input to concatenate the intermediate atom feature vectors and intermediate chemical bond feature vectors obtained from all attention heads in the first transformer layer, and then the concatenation result is compared with the original atom feature vectors input to the transformer layer. and chemical bond eigenvectors Add them together to obtain the intermediate atom feature vector corresponding to the i-th atom after processing. and intermediate chemical bond eigenvectors

[0142] (4-13) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate atom feature vector corresponding to the i-th atom obtained in step (4-12) is... and intermediate chemical bond eigenvectors Perform residual summation separately to obtain the atomic feature vector corresponding to the i-th atom after processing. and the chemical bond feature vectors corresponding to the i-th atom and the j-th atom.

[0143] The specific calculation process for this step is as follows:

[0144]

[0145] Where Norm represents normalization. and Represents the learnable parameters, and ReLU represents the activation function.

[0146] (4-14) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic feature vector corresponding to the i-th atom obtained in step (4-13) is... and the chemical bond feature vectors corresponding to the i-th atom and the j-th atom. The second transformer layer of the input drug encoder undergoes the same processing as steps (4-10) to (4-14) above to obtain the feature vector corresponding to the i-th atom in the final output of the i-th transformer layer. and the chemical bond feature vectors corresponding to the i-th atom and the j-th atom. ...and so on, until finally, after processing by the tenth transformer layer, we obtain the atomic feature vector corresponding to the i-th atom in the final output of the tenth transformer layer. and the chemical bond feature vectors corresponding to the i-th atom and the j-th atom.

[0147] (4-15) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic feature vector corresponding to the i-th atom obtained in step (4-14) is... The virtual node extraction layer of the input drug encoder performs a lookup operation (by using the atom index, finding the feature vector of the virtual node represented by the last atom, and treating it as the feature representation of the entire drug molecule structure map) to obtain a 128-dimensional drug feature vector e. d .

[0148] The advantage of the above sub-steps (4-8) to (4-15) is that by using Graph Transformer to extract drug features, not only are neighbor features considered, but a wider global receptive field is provided, which can more effectively extract features of drug molecular structure graphs.

[0149] (4-16) The protein feature vector e obtained in step (4-7) t The drug feature vector e obtained in step (4-15) d Attention calculation is performed in the first linear fusion layer of the input feature fusion module to obtain an attention score W1, and the attention score W1 is then compared with the protein feature vector e obtained in step (4-7). t The drug feature vector e obtained in step (4-15) d A weighted summation process is performed to obtain the first fusion feature.

[0150] The calculation process for this step is as follows:

[0151] W1 = Sigmoid(Linear(Linear(e t +e d )))

[0152]

[0153] Here, Linear represents linear regression and Sigmoid is the activation function.

[0154] (4-17) The first fusion feature obtained in step (4-16) Attention calculation is performed in the second linear fusion layer of the input feature fusion module to obtain an attention score W2, and this attention score W2 is then compared with the protein feature vector e obtained in step (4-5). t The drug feature vector e obtained in step (4-12) d A weighted summation process is performed to obtain the second fusion feature.

[0155] The calculation process for this step is as follows:

[0156]

[0157] (4-18) The second fusion feature obtained in step (4-17) In the gated skipping mechanism layer of the input feature fusion module, an attention score W3 is obtained, and the attention score W3 and the drug feature vector e are compared. d and protein feature vector e t and fusion features A weighted summation process is performed to obtain the final fused feature vector e of the drug and protein target. dt .

[0158] The specific steps of the calculation process in this step are as follows:

[0159]

[0160] The advantages of the sub-steps (4-16) to (4-18) above are that, through the attention-based linear fusion mechanism, the interaction between drug features and protein features can be better captured, and the interaction information of biomolecules can be learned; moreover, the gating mechanism can retain the information of the initial features and prevent some important information of the features from being lost during the training process.

[0161] (4-19) The fusion feature e obtained in step (4-20) dt Input to the multilayer perceptron module to obtain the predicted value of affinity.

[0162] The specific steps of this calculation process are as follows:

[0163]

[0164] In the four linear layers that make up this multilayer perceptron module, the first linear layer has an input dimension of 128 and an output dimension of 1024, the second linear layer has an input dimension of 1024 and an output dimension of 1024, the third linear layer has an input dimension of 1024 and an output dimension of 256, and the fourth linear layer has an input dimension of 256 and an output dimension of 1.

[0165] (4-20) Based on the drug-protein data obtained in step (4-1), the affinity between the drug and protein is compared with the predicted value of the affinity obtained in step (4-19). Calculate the loss function and use it to train the drug target binding affinity prediction model until the model converges, thus obtaining a preliminarily trained drug target binding affinity prediction model.

[0166] Specifically, the loss function calculation formula for this step is as follows:

[0167]

[0168] Where y represents the affinity between the drug and protein in the drug-protein data pair obtained in step (4-1).

[0169] (4-21) Use the test set obtained in step (4-1) to test the drug target binding affinity prediction model initially trained in step (4-19) until the value of the loss function reaches the optimal value, thereby obtaining the trained drug target binding affinity prediction model.

[0170] Test Results

[0171] The testing environment for this invention was as follows: Ubuntu 22.04 operating system, Intel(R) Xeon(R) Gold6133 CPU, NVIDIA RTX3090 24GB GPU, and PyTorch 3.7 was used to implement the algorithm of this invention.

[0172] To ensure the reliability of our experimental results, we employed five-fold cross-validation in this study. All data were randomly divided into five parts: one part served as the test set, and the remaining four parts were used for training. This resulted in five distinct outcomes for each experiment. We report the average score as the final performance metric. This setup ensured that our results were stable and reliable. The learning rate significantly impacts model performance. To improve training efficiency, we initialized the learning rate to 0.0003 and decayed it to 0.0001 after 100 iterations. The batch size for all three benchmark datasets was uniformly set to 128, and training lasted for 1000 iterations. Training was stopped to prevent overfitting if the model's loss did not decrease after 200 epochs.

[0173] Based on the test results from three different test sets recorded in Tables 1, 2, and 3, it can be seen that the graph virtual node-based drug target binding affinity prediction method proposed in this invention outperforms existing methods in four regression prediction metrics: mean squared error (MSE), concordance index (CI), modified correlation coefficient (r2m), and Pearson correlation coefficient (PCC).

[0174] Table 1 compares the prediction results on the Metz dataset.

[0175]

[0176] Table 2 shows a comparison of prediction results on the Davis dataset.

[0177]

[0178] Table 3 Comparison of prediction results on the KIBA dataset

[0179]

[0180] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting drug target binding affinity based on graph virtual nodes, characterized in that, Includes the following steps: (1) Obtain the drug-protein data pair to be predicted, and preprocess the drug-protein data pair to be predicted to obtain the preprocessed drug-protein data pair, which includes the simplified molecular linear input canonical SMILES sequence of the drug, the amino acid sequence of the protein, and the affinity between the two. (2) For each protein amino acid sequence in the preprocessed drug-protein data pair obtained in step (1), the amino acid sequence is subjected to length normalization and feature encoding processing in sequence to obtain the protein feature matrix corresponding to the amino acid sequence. (3) For each drug SMILES sequence in the preprocessed drug-protein data pair obtained in step (1), the SMILES sequence is converted into a drug molecular structure map, and virtual nodes and feature codes are added to the drug molecular structure map to obtain the drug molecular structure map corresponding to the SMILES sequence; Specifically, step (3) firstly, the open-source chemical library RDKit is used to convert the SMILES sequence into a drug molecular structure map. Here, V represents the set of all atoms in the SMILES sequence, and E represents the set of chemical bonds between atoms in the SMILES sequence. Then, a virtual node vn is added to the drug molecule structure diagram G, and this virtual node is connected to all atoms in the drug molecule structure diagram G to form virtual edges. Subsequently, the drug molecule structure diagram after adding the virtual node is feature-encoded to obtain an M-dimensional vector and an N-dimensional vector. The two vectors constitute the drug molecule structure diagram corresponding to the SMILES sequence. Where M represents the dimension of the initial feature vector corresponding to the virtual nodes included by all atoms in the drug molecule structure diagram, and N represents the dimension of the initial feature vector corresponding to the virtual edges included by all chemical bonds in the drug molecule structure diagram. (4) Input the protein feature matrix corresponding to the amino acid sequence of the protein in the drug-protein data pair obtained in step (2) and the drug molecular structure diagram corresponding to the SMILES sequence of the drug in the drug-protein data pair obtained in step (3) into the pre-trained drug target binding affinity prediction model to obtain the predicted value of the binding affinity between the drug and the protein in the drug-protein data pair. The drug target binding affinity prediction model includes a protein encoder based on a one-dimensional convolutional neural network (1DCNN), a drug encoder based on a graph transformer, a feature fusion module, and a multilayer perceptron module. The 1DCNN-based protein encoder consists of three convolutional layers and one pooling layer; The graph-based drug encoder includes a preprocessing layer, ten transformer layers, and a virtual node extraction layer. The input to the virtual node extraction layer is a 128-dimensional atomic feature vector output from the tenth transformer layer. It performs lookup processing on this atomic feature vector and outputs a 128-dimensional drug feature vector. The feature fusion module consists of two linear fusion layers and a gated skip mechanism layer, and its specific structure is as follows: The input to the first linear fusion layer is the 128-dimensional protein feature vector output from the pooling layer of the protein encoder and the 128-dimensional drug feature vector output from the virtual node extraction layer of the drug encoder. It performs attention calculation and weighted summation on the protein feature vector and the drug feature vector, and the output is a 128-dimensional intermediate fusion feature vector. The input to the second linear fusion layer is the 128-dimensional intermediate fusion feature vector output by the first linear fusion layer. It performs attention calculation and weighted summation on the intermediate fusion feature vector, and the output is a 128-dimensional intermediate fusion feature vector. The input to the gated skip mechanism layer is the 128-dimensional intermediate fused feature vector output by the second linear fusion layer. It performs attention calculation and weighted summation on the intermediate fused feature vector, and the output is the 128-dimensional final fused feature vector.

2. The method for predicting drug target binding affinity based on graph virtual nodes according to claim 1, characterized in that, In step (1), the drug-protein data pair to be predicted is preprocessed to obtain the preprocessed drug-protein data pair. Specifically, the process involves first screening the drug-protein data pair to be predicted, deleting data with missing information, to obtain the SMILES sequence of the drug, the amino acid sequence of the protein, and the inhibition and dissociation constants of the drug-protein interaction; then, the affinity between the drug and the protein is obtained based on the obtained inhibition and dissociation constants; the affinity calculation process is as follows: ； ； in and These represent the inhibition constant and the dissociation constant, respectively. These represent the affinity obtained using the inhibition constant and the affinity obtained using the dissociation constant, respectively, and are used as labels for subsequent model training and prediction. Finally, the obtained drug SMILES sequences, protein amino acid sequences, and affinity are combined to form preprocessed drug-protein data pairs.

3. The method for predicting drug target binding affinity based on graph virtual nodes according to claim 1 or 2, characterized in that, Step (2) specifically involves first normalizing the length of the amino acid sequence, then representing each amino acid in the normalized sequence with a unique integer to obtain a 1000-dimensional protein embedding vector, and finally mapping each integer in the protein embedding vector to a 128-dimensional vector. The vectors corresponding to all amino acids in the sequence ultimately form a vector of size [missing information - likely a specific value]. Protein feature matrix.

4. The method for predicting drug target binding affinity based on graph virtual nodes according to claim 3, characterized in that, The specific structure of the protein encoder based on 1DCNN is as follows: The input to the first convolutional layer is 1 1000 A 128-dimensional feature vector is subjected to convolution and activation processing, and the output is 1. 979 256-dimensional intermediate feature vector; The input to the second convolutional layer is the output of the first convolutional layer, which is 1. 979 A 256-dimensional intermediate feature vector is used to perform convolution and activation processing on the feature vector, and the output is 1. 979 256-dimensional intermediate feature vector; The input to the third convolutional layer is the output of the second convolutional layer, which is 1. 979 A 256-dimensional intermediate feature vector is used to perform convolution and activation processing on the feature vector, and the output is 1. 979 A 128-dimensional intermediate feature vector; The input to the pooling layer is the output of the third convolutional layer. 979 The intermediate feature vector is 128-dimensional. The dimensionality of this feature vector is reduced, and the output is a 128-dimensional protein feature vector. The specific structure of the graph-transformer-based drug encoder is as follows: The input to the preprocessing layer is the M-dimensional atomic feature vector and the N-dimensional chemical bond feature vector in the drug molecule structure diagram. The preprocessing layer performs mapping processing and Laplace feature position embedding processing on the atomic feature vector and the chemical bond feature vector respectively. The output is a 128-dimensional atomic feature vector and a 128-dimensional chemical bond feature vector. The input to the first transformer layer is a 128-dimensional atomic feature vector and a 128-dimensional chemical bond feature vector output from the preprocessing layer. It uses attention heads to perform attention calculations on the atomic feature vector and chemical bond feature vector respectively. That is, there are 8 independent attention heads inside the transformer. In each attention head, attention mechanism is used to calculate the attention score between two atoms and message passing processing, that is, by aggregating the current features of adjacent atoms and updating the features of atoms according to the attention score to obtain a 16-dimensional intermediate atomic feature vector. Then, it is multiplied with the attention score to obtain a 16-dimensional intermediate chemical bond feature vector and residual summation processing. The output is a 128-dimensional atomic feature vector and a 128-dimensional chemical bond feature vector. The inputs to the second to tenth transformer layers are the 128-dimensional atomic feature vector and the 128-dimensional chemical bond feature vector output from the previous transformer layer. The attention head is used to perform attention calculation, message passing and residual summation on the atomic feature vector and the chemical bond feature vector respectively, and the outputs are the 128-dimensional atomic feature vector and the 128-dimensional chemical bond feature vector respectively.

5. The method for predicting drug target binding affinity based on graph virtual nodes according to claim 4, characterized in that, The multilayer perceptron module consists of four fully connected layers, and its specific structure is as follows: The input to the first linear layer is the 128-dimensional final fused feature vector output by the gated skip mechanism layer. It performs linear regression and activation processing on the final fused feature vector, and the output is a 1024-dimensional intermediate feature vector. The input to the second linear layer is the 1024-dimensional intermediate feature vector output by the first linear layer. It performs linear regression and activation processing on the intermediate feature vector and outputs a 1024-dimensional intermediate feature vector. The input to the third linear layer is the 1024-dimensional intermediate feature vector output by the second linear layer. It performs linear regression and activation processing on the intermediate feature vector, and the output is a 256-dimensional intermediate feature vector. The input to the fourth linear layer is the 256-dimensional intermediate feature vector output by the third linear layer. It performs linear regression and activation processing on the intermediate feature vector, and the output is a 1-dimensional affinity prediction value.

6. The method for predicting drug target binding affinity based on graph virtual nodes according to claim 5, characterized in that, The drug target binding affinity prediction model is trained using the following steps: (4-1) Obtain multiple drug-protein data pairs to be predicted, preprocess all drug-protein data pairs to be predicted to obtain preprocessed drug-protein data pairs, which include the SMILES sequence of the drug, the amino acid sequence of the protein and the affinity between them, and divide all preprocessed drug-protein data pairs into training set and test set in a 4:1 ratio. (4-2) For each drug-protein data pair in the training set obtained in step (4-1), each amino acid sequence is subjected to length normalization and feature encoding in sequence to obtain the protein feature matrix corresponding to the amino acid sequence. (4-3) For each drug-protein data pair in the training set obtained in step (4-1), each SMILES sequence is converted into a drug molecule structure diagram, and virtual nodes and feature codes are added to the drug molecule structure diagram to obtain the drug molecule structure diagram corresponding to the SMILES sequence. (4-4) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the protein feature matrix corresponding to the amino acid sequence obtained in step (4-2) is input into the first convolutional layer of the protein encoder in the drug target affinity prediction model for convolution processing to obtain 1 979 A 256-dimensional feature vector is generated, and the ReLU activation function is used to activate this feature vector to obtain a 1-dimensional feature vector corresponding to the amino acid sequence. 979 256-dimensional intermediate feature vector; (4-5) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate feature vector corresponding to the amino acid sequence obtained in step (4-4) is input into the second convolutional layer of the protein encoder in the drug target affinity prediction model for convolution processing to obtain 1 979 A 256-dimensional feature vector is generated, and the ReLU activation function is used to activate this feature vector to obtain a 1-dimensional feature vector corresponding to the amino acid sequence. 979 256-dimensional intermediate feature vector; (4-6) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate feature vector corresponding to the amino acid sequence obtained in step (4-5) is input into the third convolutional layer of the protein encoder in the drug target affinity prediction model for convolution processing to obtain 1 979 A 128-dimensional feature vector is generated, and the ReLU activation function is used to activate this feature vector to obtain a 1-dimensional feature vector corresponding to the amino acid sequence. 979 A 128-dimensional intermediate feature vector; (4-7) For each amino acid sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate feature vector corresponding to the amino acid sequence obtained in step (4-6) is input into the pooling layer of the protein encoder for dimensionality reduction to obtain a 128-dimensional protein feature vector. ; (4-8) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the M-dimensional atomic feature vector and N-dimensional chemical bond feature vector of the drug molecule structure diagram corresponding to the SMILES sequence obtained in step (4-3) are input into the preprocessing layer of the drug encoder for mapping processing to obtain the 128-dimensional intermediate atom feature vector of the i-th atom. And the 128-dimensional chemical bond feature vector between the i-th atom and the j-th atom. , where i∈[1, the total number of atoms in the drug molecule structure diagram corresponding to the SMILES sequence], j∈[1, the total number of atoms in the drug molecule structure diagram corresponding to the SMILES sequence]; (4-9) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate atom feature vector of the i-th atom obtained in step (4-8) is... The preprocessing layer of the input drug encoder is subjected to Laplacian feature position embedding to obtain the processed 128-dimensional atomic feature vector corresponding to the i-th atom. ; (4-10) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic feature vector corresponding to the i-th atom obtained in step (4-9) is... The chemical bond feature vector between the i-th atom and the j-th atom obtained in steps (4-8) The first transformer layer of the input drug encoder is used to calculate the attention score to obtain the attention score between the i-th atom and the j-th atom. ; (4-11) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the attention scores corresponding to the i-th atom and the j-th atom obtained in step (4-10) are... The first transformer layer of the input drug encoder undergoes message passing processing to obtain the updated intermediate atom feature vector corresponding to the i-th atom. And the feature vectors of intermediate chemical bonds corresponding to the i-th atom and the j-th atom. ; (4-12) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate atom feature vector corresponding to the i-th atom obtained in step (4-11) is... And the feature vectors of intermediate chemical bonds corresponding to the i-th atom and the j-th atom. The first transformer layer of the drug encoder is input to concatenate the intermediate atom feature vectors and intermediate chemical bond feature vectors obtained from all attention heads in the first transformer layer, and then the concatenation result is compared with the original atom feature vectors input to the transformer layer. and chemical bond eigenvectors Add them together to obtain the intermediate atom feature vector corresponding to the i-th atom after processing. and intermediate chemical bond eigenvectors ; (4-13) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the intermediate atom feature vector corresponding to the i-th atom obtained in step (4-12) is... and intermediate chemical bond eigenvectors Perform residual summation separately to obtain the atomic feature vector corresponding to the i-th atom after processing. And the chemical bond feature vectors corresponding to the i-th atom and the j-th atom. ; (4-14) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic feature vector corresponding to the i-th atom obtained in step (4-13) is... And the chemical bond feature vectors corresponding to the i-th atom and the j-th atom. The second transformer layer of the input drug encoder undergoes the same processing as steps (4-10) to (4-14) above to obtain the feature vector corresponding to the i-th atom in the final output of the second transformer layer. And the chemical bond feature vectors corresponding to the i-th atom and the j-th atom. Finally, after processing by the tenth transformer layer, the atomic feature vector corresponding to the i-th atom is obtained from the final output of the tenth transformer layer. And the chemical bond feature vectors corresponding to the i-th atom and the j-th atom. ; (4-15) For each SMILES sequence in each drug-protein data pair in the training set obtained in step (4-1), the atomic feature vector corresponding to the i-th atom obtained in step (4-14) is... The virtual node extraction layer of the input drug encoder performs a search operation. By indexing the atomic numbers, it finds the feature vector of the virtual node represented by the last atom, and treats it as the feature representation of the entire drug molecule structure diagram to obtain a 128-dimensional drug feature vector. ; (4-16) The protein feature vector obtained in step (4-7) The drug feature vector obtained in step (4-15) Attention calculation is performed in the first linear fusion layer of the input feature fusion module to obtain the attention score. And attention score Compared with the protein feature vector obtained in steps (4-7) The drug feature vector obtained in step (4-15) A weighted summation process is performed to obtain the first fusion feature. ; (4-17) The first fusion feature obtained in step (4-16) Attention is calculated in the second linear fusion layer of the input feature fusion module to obtain an attention score. And the attention score Compared with the protein feature vector obtained in steps (4-7) The drug feature vector obtained in step (4-15) A weighted summation process is performed to obtain the second fusion feature. ; (4-18) The second fusion feature obtained in step (4-17) In the gated skipping mechanism layer of the input feature fusion module, attention scores are obtained. And the attention score Drug feature vector and protein feature vectors and fusion features A weighted summation process is performed to obtain the final fused feature vector of the drug and protein target. ; (4-19) The fusion features obtained in step (4-18) Input to the multilayer perceptron module to obtain the predicted value of affinity. ; (4-20) The affinity between the drug and protein in the drug-protein data obtained in step (4-1) and the predicted value of the affinity obtained in step (4-19). Calculate the loss function and use it to train the drug target binding affinity prediction model until the model converges, thus obtaining a preliminarily trained drug target binding affinity prediction model. (4-21) Use the test set obtained in step (4-1) to test the drug target binding affinity prediction model initially trained in step (4-20) until the value of the loss function reaches the optimal value, thereby obtaining the trained drug target binding affinity prediction model.

7. The method for predicting drug target binding affinity based on graph virtual nodes according to claim 6, characterized in that, Step (4-9) specifically involves calculating the symmetric normalized Laplacian matrix L using the identity matrix I, degree matrix D, and adjacency matrix A of the drug molecule structure diagram. The calculation formula is as follows: ; Then, the Laplacian position eigenvector is calculated using the following formula: ,in are the eigenvectors of the Laplacian matrix L, and C are the learnable parameters. Bias term; then, With atomic characteristics Add them together to obtain the processed 128-dimensional atomic feature vector corresponding to the i-th atom. The calculation formula is as follows: ; The specific calculation formula in step (4-10) is as follows: ； in , Let be the linear transformation matrices of the eigenvectors of the i-th and j-th atoms respectively during the processing of the first transformer. Let be the linear transformation matrix of the eigenvectors of the chemical bond between the i-th and j-th atoms in the first transformer layer. The linear transformation matrix The eigenvector dimension.

8. The method for predicting drug target binding affinity based on graph virtual nodes according to claim 7, characterized in that, In step (4-11), for message passing of an atom, the current features of adjacent atoms, i.e., the j-th atom, are aggregated and sorted according to attention scores. The eigenvector of the i-th atom is updated using weighted methods. The specific calculation process is as follows: ； in Let represent the set of adjacent atoms of the i-th atom in the SMILES sequence. Let be the linear transformation matrix of the eigenvectors of the adjacent atoms of the i-th atom in the first transformer layer. yes Attention score after activation by the Softmax function; For message passing in chemical bonds, directly... Assigned as the intermediate chemical bond feature vector ,Right now ; The calculation process in step (4-13) is as follows: ； ； Where Norm represents normalization. and Represents the learnable parameters, and ReLU represents the activation function.

9. The method for predicting drug target binding affinity based on graph virtual nodes according to claim 8, characterized in that, The calculation process for step (4-16) is as follows: ； ； Where Linear represents linear regression and Sigmoid is the activation function; The calculation process in step (4-17) is as follows: ； ； The specific steps of the calculation process in step (4-18) are as follows: ； ； The calculation process for step (4-19) is as follows: ； The formula for calculating the loss function in step (4-20) is as follows: ； Where y represents the affinity between the drug and protein in the drug-protein data pair obtained in step (4-1).

10. A drug target binding affinity prediction system based on graph virtual nodes, implemented based on the drug target binding affinity prediction method based on graph virtual nodes as described in claim 1, characterized in that, The drug target binding affinity prediction system based on graph virtual nodes includes: The first module is used to acquire the drug-protein data pair to be predicted, and to preprocess the drug-protein data pair to be predicted to obtain the preprocessed drug-protein data pair, which includes the simplified molecular linear input canonical SMILES sequence of the drug, the amino acid sequence of the protein, and the affinity between the two. The second module is used to perform length normalization and feature encoding on the amino acid sequence of each protein in the preprocessed drug-protein data pair obtained by the first module, so as to obtain the protein feature matrix corresponding to the amino acid sequence. The third module is used to convert the SMILES sequence of each drug in the preprocessed drug-protein data pair obtained by the first module into a drug molecular structure map, and add virtual nodes and feature codes to the drug molecular structure map to obtain the drug molecular structure map corresponding to the SMILES sequence. The fourth module is used to input the protein feature matrix corresponding to the amino acid sequence of the protein in the drug-protein data pair obtained in the second module, and the drug molecular structure map corresponding to the SMILES sequence of the drug in the drug-protein data pair obtained in the third module into a pre-trained drug target binding affinity prediction model, so as to obtain the predicted value of the binding affinity between the drug and the protein in the drug-protein data pair.

Images