A method and system for predicting drug-target protein binding affinity

Abstract

The application provides a drug-target protein binding affinity prediction method and system, relates to the technical field of biological information technology, and comprises the following steps: learning local information of an atomic composite graph at a local level by using an isometric graph neural network; fully utilizing three-dimensional space structure information by using the isometric graph neural network, and keeping rotation and translation of coordinates unchanged during message transmission; then using a node feature extraction model and an atomic interaction matrix to learn deep-level interaction of the atomic composite graph at a global and long-distance level, improving interpretability of feature learning of the protein and the ligand, and making prediction performance of the drug-target protein binding affinity prediction model optimal. The application combines interaction of each distance level of the drug-target protein by fitting a dynamic process of drug-protein combination, simultaneously considers rotation and translation invariance of atomic mutual combination, and effectively improves accuracy of drug-target protein binding affinity prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and in particular to a method and system for predicting drug-target protein binding affinity. Background Art

[0002] Binding affinity numerically quantifies the strength of protein-ligand interactions. Generally speaking, discovering ligands with high affinity that bind to target proteins, thereby assisting in the search for new compounds that can treat specific diseases, is a major focus of pharmaceutical research. Furthermore, higher binding affinity allows for the use of lower drug doses while achieving equivalent therapeutic effects, which can effectively reduce the toxicity caused by drug ingestion. Therefore, finding a computational method that accurately predicts protein-ligand binding affinity to assist in drug design and accelerate drug screening is of great significance for new drug discovery, drug repositioning, and the prediction of drug side effects.

[0003] At present, the research methods for predicting protein-ligand binding affinity are mainly divided into energy simulation-centered methods, machine learning-based scoring function methods, and deep learning-centered methods; the energy simulation-centered methods use a combination of all-atom molecular dynamics-based simulations and free energy simulations to simulate the entire binding process. Although they are based on strict and accurate physical calculation methods, the computational cost is too high; in comparison, the machine learning-based scoring function methods have improved computational efficiency to a certain extent, but because they rely more on artificially created features and additional expert knowledge, there is still much room for improvement in the prediction accuracy of binding affinity; the deep learning-centered method uses deep learning for affinity prediction, which can directly learn the features of the original data and greatly improve the accuracy of protein-ligand binding affinity prediction.

[0004] Existing deep learning-centric methods, such as some sequence-based methods, input the amino acid sequence of proteins and the SMILE sequence of small molecules, and then use convolutional neural networks to learn their representations to predict binding affinity. However, sequence data does not have three-dimensional spatial information, and this method cannot well fit the dynamic process of drug-protein binding, affecting the accuracy of binding affinity prediction; there are also graph-based methods that use proteins and ligands to construct graph data, and then use graph neural networks to learn the topological structure of the graph to obtain a graph representation. These methods have made impressive progress, but they do not consider the interaction between proteins and ligands after binding or ignore the spatial coordinate information of atoms when composing the graph. Although a small amount of work has been attempted, such as using position coordinates directly in convolutional neural networks based on spatial graph convolutional networks, and introducing polar coordinates based on spatial graph networks, and adding angle information on this basis, they do not consider the rotation and translation invariance during message transmission, which also affects the accuracy of binding affinity prediction; and after the protein and ligand form a complex, these methods generally only consider the protein atoms within a certain range that bind to small molecules to form pockets for computational reasons, which will ignore long-range interactions. The van der Waals forces and hydrogen bonds between these large number of long-range atoms will also affect the prediction of binding affinity. Ignoring long-range interactions will further affect the accuracy of drug-target protein binding affinity prediction. Summary of the Invention

[0005] In order to solve the problem that the above-mentioned existing technologies are difficult to fit the dynamic process of drug-protein binding, do not consider the rotational and translational invariance in the message transmission process, ignore the interaction between long-distance atoms, and thus affect the accuracy of drug-target protein binding affinity prediction, the present invention proposes a drug-target protein binding affinity prediction method and system. By fitting the dynamic process of drug-protein binding, the interactions at various distance levels of the drug-target protein are combined, and at the same time the rotational and translational invariance of the mutual binding of atoms is considered, thereby effectively improving the accuracy of drug-target protein binding affinity prediction.

[0006] In order to achieve the above technical effects, the technical solutions of the present invention are as follows:

[0007] A method for predicting drug-target protein binding affinity comprises the following steps:

[0008] S1. Obtaining an atomic composite image of a drug-target protein and related characteristic data of the atomic composite image;

[0009] S2. Constructing a drug-target protein binding affinity prediction model, wherein the drug-target protein binding affinity prediction model includes a node feature extraction model and a multi-layer perceptron;

[0010] S3. The relevant feature data is input as the input data of the node feature extraction model to the node feature extraction model, and the node feature extraction model outputs local information and global information of the atomic composite graph;

[0011] S4. Preprocessing the local information and the global information to obtain a graph representation, inputting the graph representation into the multilayer perceptron, and using the multilayer perceptron to predict drug-target protein binding affinity;

[0012] S5. Construct an atomic interaction matrix for learning the long-range information of the atoms in the atomic composite graph, and use the atomic interaction matrix to optimize the performance of the drug-target protein binding affinity prediction model in predicting drug-target protein binding affinity, so as to obtain a drug-target protein binding affinity prediction model with the best prediction performance.

[0013] Preferably, the obtaining of the drug-target protein atomic composite graph and related characteristic data of the atomic composite graph includes:

[0014] S101. Obtain target protein molecule files and drug molecule files;

[0015] S102. combining the target protein molecule in the target protein molecule file and the drug molecule in the drug molecule file to obtain a three-dimensional complex of the target protein molecule and the drug molecule;

[0016] S103. Extracting an atomic composite image consisting of target protein atoms and drug atoms from the three-dimensional complex according to the pocket region where the target protein molecule and the drug molecule bind

[0017] S104. From the atomic composite diagram Extract the relevant feature data, the relevant feature data including the atomic composite graph The atomic node feature matrix h, the atomic three-dimensional coordinate matrix and the edge matrix ε.

[0018] Preferably, the node feature extraction model includes a parallel local information extraction module for extracting local information of the atomic composite graph and a global information extraction module for extracting global information of the atomic composite graph. The local information extraction module is provided with two layers of equivariant graph neural network connected in sequence, and the global information extraction module includes a self-attention layer and a feedforward neural network layer connected in sequence.

[0019] Preferably, extracting local information of the atomic composite graph using the local information extraction module includes:

[0020] S301. The atomic node feature matrix h, the atomic three-dimensional coordinate matrix and the edge matrix ε are input into the equivariant graph neural network in the local information extraction module;

[0021] S302. Determine the message m of the equivariant graph neural network using the edge matrix ε ij The transfer direction is based on the atomic node feature matrix h and the atomic three-dimensional coordinate matrix Calculate the message m ij as follows:

[0022]

[0023] Among them, σ represents the SiLU activation function, represents the first learning parameter, x i Represents an atomic composite graph The three-dimensional coordinates of the i-th atom in x j Represents an atomic composite graph The three-dimensional coordinates of the jth atom in n Represents an atomic composite graph The three-dimensional coordinates of the nth atom in , Represents an atomic composite graph The characteristic matrix of the i-th atom in , Represents an atomic composite graph The characteristic matrix of the jth atom in ;

[0024] S303. Based on the message m ij , the atomic node feature matrix is ​​updated using a two-layer equivariant graph neural network as follows:

[0025]

[0026] in, represents the updated atomic node feature matrix, represents the updated three-dimensional coordinates of the atom, and N(i) represents the composite graph with the atom The neighboring atoms of the i-th atom in , represents the second learning parameter, is the third learning parameter, || represents the concatenation operation;

[0027] S304. The atomic node feature matrix obtained after the two-layer equivariant graph neural network update is connected using residuals to obtain the local feature matrix h of the atomic composite graph. e , denote the local feature matrix h e is the local information.

[0028] Preferably, extracting global information of the atomic composite graph using the global information extraction module includes:

[0029] S311. Input the atomic feature matrix into the global information extraction module, process the atomic node feature matrix, and obtain the query Q, keyword K, and value V that need to be input into the self-attention layer;

[0030] S312. Input the query Q, keyword K, and value V into the self-attention layer, and the self-attention layer outputs an attention score, which is normalized;

[0031] S313. Combining the normalized attention score with the atomic node feature matrix through a residual connection to obtain combined information;

[0032] S314. Inputting the combined information into the feedforward neural network, and updating the feature representation of the atomic node feature matrix by the feedforward neural network;

[0033] S315. Normalizing the feature representation;

[0034] S316. Combine the normalized feature representation with the atomic node feature matrix through residual connection to obtain the global feature matrix h of the atomic composite graph T , denote the global feature matrix h T is the global information.

[0035] Preferably, the preprocessing of the local information and the global information includes:

[0036] S401. The local feature matrix h e and the global feature matrix h T By connecting them in series, we can obtain the atomic node feature h′ which contains three-dimensional structural information, local information and global information.

[0037] S402. Sum and pool the atomic node features h′ to obtain the graph representation

[0038] Preferably, the constructing of an atomic interaction matrix for learning long-range information of atoms in the atomic composite graph comprises:

[0039] S501. Calculate atomic composite graph The edge embedding E ij as follows:

[0040] E ij =W d d ij

[0041] Among them, W d represents the embedding weight matrix, d ij Represents an atomic composite graph Paired atoms (v i ,v j ) of the Euclidean distance, υ i Represents an atomic composite graph The i-th atom in j Represents an atomic composite graph The jth atom in ;

[0042] S502. Based on the edge embedding E ij , calculate the atomic composite graph The interaction between pairs of atoms in h p,l as follows:

[0043]

[0044] Among them, γ(v i ,v j ) means return (υ i ,υ j ) is a function of the atomic number, δ represents the Kronecker function, (p,l) represents (υ i ,υ j ) atomic type, W h represents the first learning weight matrix;

[0045] S503. Based on the pair-wise atomic interaction h p,l , calculate the atomic interaction matrix as follows:

[0046]

[0047] Among them, W k Represents the second learning weight matrix.

[0048] Preferably, the optimizing the performance of the drug-target protein binding affinity prediction model in predicting drug-target protein binding affinity by using the atomic interaction matrix comprises:

[0049] S511. Using the atomic interaction matrix, calculate the atomic long-range interaction loss function as follows:

[0050]

[0051] Among them, Train represents the training set, K pl represents the knowledge-based interaction matrix;

[0052] S512. Using the multi-layer perceptron to predict drug-target protein binding affinity The calculation expression is as follows:

[0053]

[0054] Wherein, MLP() represents the activation function of the multi-layer perceptron;

[0055] S513. Based on the drug-target protein binding affinity Calculate the prediction loss function of the drug-target protein binding affinity prediction model The calculation expression is as follows:

[0056]

[0057] S514. Based on the long-range interaction loss function and the drug-target protein binding affinity Calculate the total loss function for training the drug-target protein binding affinity prediction model The calculation expression is as follows:

[0058]

[0059] Among them, λ represents the weight coefficient;

[0060] S515. Using the total loss function The drug-target protein binding affinity prediction model is trained to obtain a drug-target protein binding affinity prediction model with the best prediction performance.

[0061] The present invention also proposes a drug-target protein binding affinity prediction system, comprising:

[0062] A data acquisition module, configured to acquire an atomic composite image of a drug and a target protein and related characteristic data of the atomic composite image;

[0063] A model building module is used to build a drug-target protein binding affinity prediction model, wherein the drug-target protein binding affinity prediction model includes a node feature extraction model and a multi-layer perceptron;

[0064] An input module, configured to input the relevant feature data as input data of the node feature extraction model into the node feature extraction model, and the node feature extraction model outputs local information and global information of the atomic composite graph;

[0065] a prediction module, configured to preprocess the local information and the global information to obtain a graph representation, input the graph representation into the multilayer perceptron, and use the multilayer perceptron to predict drug-target protein binding affinity;

[0066] A prediction performance optimization module is used to construct an atomic interaction matrix for learning the long-range information of atoms in the atomic composite graph, and use the atomic interaction matrix to optimize the performance of the drug-target protein binding affinity prediction model in predicting drug-target protein binding affinity, so as to obtain a drug-target protein binding affinity prediction model with the best prediction performance.

[0067] The present invention also provides a computer device, characterized in that it includes: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus;

[0068] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform the operation of the drug-target protein binding affinity prediction method.

[0069] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0070] The present invention proposes a method and system for predicting drug-target protein binding affinity. First, for the task of drug-target protein binding affinity prediction, a general framework is innovatively proposed that simultaneously considers interactions at multiple levels. This framework can automatically learn features of different scales at different levels. Second, an equivariant graph neural network is used to learn local information of the atomic composite graph at the local level. The equivariant graph neural network has two parts: message passing and aggregation. However, it can fully utilize three-dimensional spatial structural information, maintaining the rotation and translation of coordinates invariant during message passing. Then, at the global and long-range levels, a node feature extraction model and an atomic interaction matrix are innovatively used to learn deep interactions in the atomic composite graph. This improves the interpretability of protein and ligand feature learning, optimizes the predictive performance of the drug-target protein binding affinity prediction model, and effectively enhances the accuracy of drug-target protein binding affinity prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 A flowchart of a method for predicting drug-target protein binding affinity provided in an embodiment of the present invention is shown;

[0072] Figure 2 A diagram showing the structure of the drug-target protein binding affinity prediction model provided in an embodiment of the present invention;

[0073] Figure 3 A comparison diagram of the first ablation experiment provided in an embodiment of the present invention;

[0074] Figure 4 A comparison diagram of the second ablation experiment provided in an embodiment of the present invention;

[0075] Figure 5 A block diagram of a drug-target protein binding affinity prediction system provided in an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0076] The accompanying drawings are for illustrative purposes only and are not to be construed as limiting this patent;

[0077] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are some embodiments of the present invention. A person skilled in the art can also derive other drawings based on these drawings without inventive effort. It should be understood that the step numbers used herein are for convenience of description only and do not limit the order in which the steps must be executed.

[0078] It is understandable to those skilled in the art that some well-known contents may be omitted in the drawings. The technical solution of the present invention will be further described below with reference to the drawings and embodiments.

[0079] Example 1

[0080] like Figure 1 and Figure 2 As shown, this embodiment proposes a method for predicting drug-target protein binding affinity, comprising the following steps:

[0081] S1. Obtaining an atomic composite image of a drug-target protein and related characteristic data of the atomic composite image;

[0082] In step S1, before obtaining the atomic complex map of the drug-target protein and the relevant characteristic data of the atomic complex map, first prepare the data set and collect a data set with a binary relationship between the drug and the target protein. The binary relationship of the data set is "drug-protein complex-affinity". The data sets used in this step are all public PDBbind data sets. The PDBbind data set includes structural data collected from the Protein Data Bank and related experimental measurements. The PDBbind data set also provides structural information about the protein and ligand complex, including the atomic coordinates of the protein and the structure and chemical information of the ligand. Then, obtain the atomic complex map of the drug-target protein and the relevant characteristic data of the atomic complex map, including:

[0083] S101. Obtain target protein molecule files and drug molecule files;

[0084] S102. combining the target protein molecule in the target protein molecule file and the drug molecule in the drug molecule file to obtain a three-dimensional complex of the target protein molecule and the drug molecule;

[0085] S103. Extracting an atomic composite image consisting of target protein atoms and drug atoms from the three-dimensional complex according to the pocket region where the target protein molecule and the drug molecule bind

[0086] S104. From the atomic composite diagram Extract the relevant feature data, the relevant feature data including the atomic composite graph The atomic node feature matrix h, the atomic three-dimensional coordinate matrix and the edge matrix ε.

[0087] It should be noted that the target protein molecule and the drug molecule combine to form a three-dimensional complex. The target protein atoms within a certain distance around the drug atoms are considered as a pocket, and then all the target protein atoms and drug atoms in this area form an atomic composite diagram. Atomic node set n represents the number of all atoms; the feature matrix of each atomic node is h, which consists of 18-dimensional features such as atomic type and electrostatic interaction; for edge construction, it is defined as adding an edge e between two atoms in the atomic composite graph if the Euclidean distance between them is less than a threshold ij , atomic three-dimensional coordinate matrix Represents the three-dimensional coordinates of all atoms in the atomic composite diagram.

[0088] S2. Constructing a drug-target protein binding affinity prediction model, wherein the drug-target protein binding affinity prediction model includes a node feature extraction model and a multi-layer perceptron;

[0089] S3. The relevant feature data is input as the input data of the node feature extraction model to the node feature extraction model, and the node feature extraction model outputs local information and global information of the atomic composite graph;

[0090] In step S3, in order not to lose the original feature information, the feature of each atom is encoded into a 36-dimensional feature vector using unique hot encoding, and the strategy of filling the first 18 bits with 0 or the last 18 bits with 0 is used to distinguish between drug atoms and target protein atoms; the node feature extraction model includes a local information extraction module for extracting local information of the atomic composite graph in parallel and a global information extraction module for extracting global information of the atomic composite graph, the local information extraction module is provided with two layers of equivariant graph neural networks connected in sequence, the input dimension of the first layer of equivariant graph neural network is the dimension of the atomic node feature matrix, i.e. 36, and the output dimension is 128, and the ReLU activation function is selected for nonlinear activation; the input dimension of the second layer of equivariant graph neural network is 128, and the output dimension is 128. After feature extraction, the updated atomic node feature matrix is ​​obtained; the global information extraction module includes a self-attention layer and a feedforward neural network layer connected in sequence;

[0091] Extracting local information of the atomic composite graph using the local information extraction module includes:

[0092] S301. The atomic node feature matrix h, the atomic three-dimensional coordinate matrix and the edge matrix ε are input into the equivariant graph neural network in the local information extraction module;

[0093] In step S301 , in order to conveniently store the structure of the atomic composite graph, the edge matrix ε stores the structural information of the atomic composite graph.

[0094] S302. Determine the message m of the equivariant graph neural network using the edge matrix ε ij The transfer direction is based on the atomic node feature matrix h and the atomic three-dimensional coordinate matrix Calculate the message m ij as follows:

[0095]

[0096] Among them, σ represents the SiLU activation function, represents the first learning parameter, Represents an atomic composite graph The three-dimensional coordinates of the i-th atom in x j Represents an atomic composite graph The three-dimensional coordinates of the jth atom in n Represents an atomic composite graph The three-dimensional coordinates of the nth atom in , Represents an atomic composite graph The characteristic matrix of the i-th atom in , Represents an atomic composite graph The characteristic matrix of the jth atom in ;

[0097] S303. Based on the message m ij , the atomic node feature matrix h is updated using a two-layer equivariant graph neural network as follows:

[0098]

[0099] in, represents the updated atomic node feature matrix, represents the updated three-dimensional coordinates of the atom, and N(i) represents the composite graph with the atom The neighboring atoms of the i-th atom in , represents the second learning parameter, || represents the concatenation operation;

[0100] The atomic three-dimensional coordinate matrix is ​​also calculated using a two-layer equivariant graph neural network. Make the update as follows:

[0101]

[0102] in, represents the third learning parameter;

[0103] S304. The atomic node feature matrix obtained after the two-layer equivariant graph neural network update is connected using residuals to obtain the local feature matrix h of the atomic composite graph. e , denote the local feature matrix h e is the local information.

[0104] It should be noted that the equivariant graph neural network is similar to the graph neural network. The equivariant graph neural network also has two parts: message passing and aggregation. However, it can make full use of the three-dimensional spatial structure information and maintain the atomic three-dimensional coordinate matrix during message passing. The rotation and translation of the coordinates are invariant and the nodes are kept From the overall perspective, the local information extraction module is used to extract the local information of the atomic composite graph, including inputting the atomic node feature matrix h, the atomic three-dimensional coordinate matrix and the edge matrix ε and output the updated h l+1 and x l+1 By introducing the distance information between atomic nodes into the message to be transmitted, x i -x j The introduction of the coordinate update process not only fully utilizes the three-dimensional structural information of the atomic composite graph, but also maintains the equivariance of rotation and translation; due to the over-smoothing problem of graph neural networks, this embodiment uses two layers and one layer of residual connections to capture the first-order and second-order neighborhood node features.

[0105] The global information extraction module is a Transformer network, and the global information of the atomic composite graph is extracted using the global information extraction module, including:

[0106] S311. Input the atomic feature matrix into the global information extraction module, process the atomic node feature matrix, and obtain the query Q, keyword K and value V that need to be input into the self-attention layer; wherein Q, K and V are all obtained from the atomic feature matrix through the MLP network.

[0107] S312. Input the query Q, keyword K, and value V into the self-attention layer, and the self-attention layer outputs an attention score, which is normalized;

[0108] In step S312, the query Q, keyword K and value V are input into the self-attention layer and then substituted into the following attention calculation formula:

[0109]

[0110] Among them, A(Q,K,V) represents the attention score, d represents the hidden layer feature dimension 128;

[0111] S313. Combining the normalized attention score with the atomic node feature matrix through a residual connection to obtain combined information;

[0112] S314. Inputting the combined information into the feedforward neural network, and updating the feature representation of the atomic node feature matrix by the feedforward neural network;

[0113] S315. Normalizing the feature representation;

[0114] S316. Combine the normalized feature representation with the atomic node feature matrix through residual connection to obtain the global feature matrix h of the atomic composite graph T , denote the global feature matrix h T is the global information.

[0115] It should be noted that, first, in order to overcome the challenge of over-smoothing and learn the global node features in the atomic composite graph, the atomic composite graph is regarded as a fully connected graph; then the atomic feature matrix is ​​input into a Transformer network consisting of a self-attention layer and a feedforward neural network layer to extract global information. Since the Transformer network has achieved great success in various fields, in this embodiment, the atomic feature matrix is ​​input and updated using the self-attention layer and the feedforward neural network layer; the self-attention mechanism of the self-attention layer is calculated using the attention calculation formula. Unlike other attention mechanisms, the Transformer network proposed in this embodiment can effectively capture the features of the entire atomic composite graph using only one attention head; at the same time, residual connections and normalization layers are used before and after the feedforward neural network layer. In this way, an updated representation of the atomic nodes in the entire atomic composite graph is obtained.

[0116] S4. Preprocessing the local information and the global information to obtain a graph representation, inputting the graph representation into the multilayer perceptron, and using the multilayer perceptron to predict drug-target protein binding affinity;

[0117] In step S4, the preprocessing of the local information and the global information includes:

[0118] S401. The local feature matrix h e and the global feature matrix h T By connecting them in series, we can obtain the atomic node feature h′ which contains three-dimensional structural information, local information and global information. The calculation expression of h′ is as follows:

[0119]

[0120] S402. Sum and pool the atomic node features h′ to obtain the graph representation

[0121] S5. Construct an atomic interaction matrix for learning the long-range information of the atoms in the atomic composite graph, and use the atomic interaction matrix to optimize the performance of the drug-target protein binding affinity prediction model in predicting drug-target protein binding affinity, so as to obtain a drug-target protein binding affinity prediction model with the best prediction performance.

[0122] In step S5, the pairwise atomic interactions of the target protein and the drug are combined with edge embedding to capture long-range interactions; an atomic interaction matrix for learning the long-range information of the atoms in the atomic composite graph is constructed, including:

[0123] S501. Calculate atomic composite graph The edge embedding Eij as follows:

[0124] E ij =W d d ij

[0125] Among them, W d represents the embedding weight matrix, d ij Represents an atomic composite graph Paired atoms (v i ,v j ) of the Euclidean distance, υ i Represents an atomic composite graph The i-th atom in j Represents an atomic composite graph The jth atom in the target protein set P = {C, N, O, F, P, S, Cl, Br, I} and the drug set L = {C, N, O, S} are selected. Halogen atom types are selected for both the target protein set and the drug set to maintain feature sparsity when constructing each feature. The feature includes the number of occurrences of long-range specific drug-target protein atom pairs. The elements in the two sets are numbered as atomic numbers to facilitate subsequent matrix operations, such as P ∈ {6, 7, 8, ...} and L ∈ {6, 7, 8, 16}. Therefore, the knowledge-based interaction matrix K can be created. pl , which is defined as:

[0126]

[0127] Among them, ν p represents the set of atomic nodes of atom type p, and ν l The set of atomic nodes of atom type l, Indicates the first distance, d ij Represents the second distance, θ() represents the step function, used to calculate The frequency of occurrence of each drug-target protein atom pair within the distance, s(p,l) represents the intermediate parameter, and the calculation expression of s(p,l) is as follows:

[0128]

[0129] S502. Based on the edge embedding E ij , calculate the atomic composite graph The interaction between pairs of atoms in h p,l as follows:

[0130]

[0131] Among them, γ(v i ,v j ) means return (υ i ,υj ) atomic number function, such as the atomic numbers of carbon and oxygen are (6,8), δ represents the Kronecker function, and its value is only considered as the atomic γ(v i ,v j ) is 1 when it is (p, l), otherwise it is 0, (p, l) means (υ i ,υ j ) atomic type, W h represents the first learning weight matrix;

[0132] S503. Based on the pair-wise atomic interaction h p,l , calculate the atomic interaction matrix as follows:

[0133]

[0134] Among them, W k Represents the second learning weight matrix; 9 target protein atoms are defined in the protein and 4 drug atoms are defined in the drug ligand, so the atomic interaction matrix It is then flattened into a vector to conveniently calculate the interaction score.

[0135] Considering that the composite graph-based method cannot utilize the information that has been proven to be helpful for affinity prediction over a long distance, the drug-target protein binding affinity prediction method proposed in this example

[0136] An additional knowledge-based atomic interaction matrix is ​​introduced. Inspired by the knowledge-based scoring function, the occurrence frequency of atom type pairs can be used to obtain pairwise interactions. The learned edge features are used to approximate an original feature matrix. Therefore, the atomic interaction matrix is ​​used to optimize the performance of the drug-target protein binding affinity prediction model in predicting drug-target protein binding affinity, including:

[0137] S511. Using the atomic interaction matrix, calculate the atomic long-range interaction loss function as follows:

[0138]

[0139] Among them, Train represents the training set, K pl represents the knowledge-based interaction matrix;

[0140] S512. Using the multi-layer perceptron to predict drug-target protein binding affinity The calculation expression is as follows:

[0141]

[0142] Wherein, MLP() represents the activation function of the multi-layer perceptron;

[0143] S513. Based on the drug-target protein binding affinity Calculate the prediction loss function of the drug-target protein binding affinity prediction model The calculation expression is as follows:

[0144]

[0145] Among them, y represents the true label corresponding to the atomic composite graph, and the loss function is predicted Predicted value of the drug-target protein binding affinity prediction model will be close to the true label y corresponding to the atomic composite graph.

[0146] S514. Based on the long-range interaction loss function and the drug-target protein binding affinity Calculate the total loss function for training the drug-target protein binding affinity prediction model The calculation expression is as follows:

[0147]

[0148] Among them, λ represents the weight coefficient; when λ is set to 1.75, the prediction performance of the drug-target protein binding affinity prediction model reaches the best.

[0149] In step S514, considering the influence of long-range interactions on the prediction of drug-target protein binding affinity, the total loss function is defined as To train the drug-target protein binding affinity prediction model.

[0150] S505. Using the total loss function Training the drug-target protein binding affinity prediction model to obtain a drug-target protein binding affinity prediction model with the best prediction performance;

[0151] It should be noted that during the training process of the drug-target protein binding affinity prediction model, the Adam optimizer is used to update the network parameters of the drug-target protein binding affinity prediction model, and weight decay is set. The weight decay rate is set to 0.0001, the learning rate is set to 0.001, and the learning rate is adjusted by adjusting the learning rate at equal intervals to make the training of the drug-target protein binding affinity prediction model more stable; after the network parameters of the drug-target protein binding affinity prediction model are iteratively updated, the final drug-target protein binding affinity of the output node is output.

[0152] In this example, for the task of drug-target protein binding affinity prediction, a general framework that simultaneously considers interactions at multiple levels is innovatively proposed. This framework can automatically learn features of different scales at different levels. Secondly, an equivariant graph neural network is used to learn local information of the atomic composite graph at the local level. The equivariant graph neural network has two parts: message passing and aggregation. However, it can fully utilize three-dimensional spatial structural information, maintaining the rotation and translation of coordinates during message passing. Then, at the global and long-range levels, a node feature extraction model and an atomic interaction matrix are innovatively used to learn deep interactions in the atomic composite graph. This improves the interpretability of protein and ligand feature learning, optimizes the predictive performance of the drug-target protein binding affinity prediction model, and effectively enhances the accuracy of drug-target protein binding affinity prediction.

[0153] Example 2

[0154] This example further verifies and analyzes the predictive performance of the drug-target protein binding affinity prediction model proposed in Example 1. The specific process is as follows:

[0155] 1) Data Collection: This example conducts experiments on the PDBbind-2016 dataset. PDBbind-2016, from the PDBbind database, is one of the most commonly used datasets for predicting drug-target protein binding affinities because it provides detailed 3D structures of protein-ligand complexes and corresponding experimentally verified binding affinities. PDBbind-2016 has three overlapping subsets: a general set containing 16,179 protein-ligand complexes, a high-quality refined set from the general set containing 4,057 conformations, and a core set of 290 conformations selected from the refined set. The remaining 3,767 protein-ligand complexes were randomly divided into training and validation sets at a ratio of 9:1. The core set served as the test set. The CSAR-HiQ dataset, with two subsets containing 167 and 176 complexes, was used as an additional test set to evaluate the generalization ability of our model.

[0156] 2) Model building: A drug-target protein binding affinity prediction model is built according to the drug-target protein binding affinity prediction model building steps described in Example 1, and the atomic complex graph and features of the drug-target protein are input into the network model.

[0157] 3) Selecting validation methods and evaluation indicators: This step selects the training, validation, and testing methods, that is, using the root mean square error (RMSE), mean absolute error (MAE), standard deviation (SD), and correlation index (R) to evaluate the performance of the drug-target protein binding affinity prediction model.

[0158] This example treats the drug-target protein binding affinity prediction problem for drug repositioning as a regression problem. For regression problems, the predictions are a series of discrete values, and the true labels are also a series of discrete values. Based on these values, evaluation metrics such as the root mean square error (RMSE), mean absolute error (MAE), standard deviation (SD), and correlation index (R) can be calculated. The calculation formulas for these four metrics are as follows:

[0159]

[0160]

[0161]

[0162]

[0163] Among the above four evaluation indicators, the smaller the RMSE, SD and MAE values ​​are, the better the prediction performance of the drug-target protein binding affinity prediction model is, and the smaller the R value is, the worse the prediction performance of the drug-target protein binding affinity prediction model is.

[0164] 4) Parameter setting and effect evaluation: In the construction step of the atomic composite graph, set the cutoff distance to control the presence of edges between any two atoms. Furthermore, the raw node features consist of 18 atomic properties, including the number of atoms, charge, and hybridization, encoded using unique vectors. The equivariant graph neural network in the model was set to two layers on the PDBbind-2016 core set for 500 iterations; for the CSAR-HiQ test set, only one equivariant graph neural network was used for 70 iterations. Furthermore, the Adam optimizer had a learning rate of 0.001, a total hidden dimension of 128 atoms, a batch size of 32, and a hyperparameter λ of 1.75.

[0165] 5) In order to verify the effectiveness of drug-target protein binding affinity prediction, this example set up an ablation experiment, removing the local information extraction module, global information extraction module, and pairwise atomic long-range interaction module composed of the equivariant graph neural network in this example and then comparing them, using RMSE and R as evaluation indicators. The experimental results are as follows Figure 3 、 Figure 4 As shown, this indicates that the model designed by the present invention can greatly improve the performance of drug-target protein binding affinity prediction for protein-ligand affinity prediction and new dataset scenarios, and performs well in new dataset scenarios.

[0166] Parameter setting and effect evaluation. In the construction step of the composite graph, set the cutoff distance to control whether there is an edge between any two atoms. In addition, the original node features consist of 18 atomic properties such as the number of atoms, charge, hybridization, etc., which are encoded using one-hot vectors. The equivariant graph neural network in the model is set to 2 layers on the PDBbind-2016 core set for 500 iterations. For the CSAR-HiQ test set, only one layer of equivariant graph neural network and 70 iterations are set. In addition, the learning rate of the Adam optimizer is 0.001, and all hidden dimensions of atoms are 128. The batch size is set to 32, and the hyperparameter λ is set to 1.75.

[0167] Example 3

[0168] See also Figure 5 This embodiment also proposes a drug-target protein binding affinity prediction system, including:

[0169] A data acquisition module, configured to acquire an atomic composite image of a drug and a target protein and related characteristic data of the atomic composite image;

[0170] A model building module is used to build a drug-target protein binding affinity prediction model, wherein the drug-target protein binding affinity prediction model includes a node feature extraction model and a multi-layer perceptron;

[0171] An input module, configured to input the relevant feature data as input data of the node feature extraction model into the node feature extraction model, and the node feature extraction model outputs local information and global information of the atomic composite graph;

[0172] The node feature extraction model includes a parallel local information extraction module for extracting local information of the atomic composite graph and a global information extraction module for extracting global information of the atomic composite graph. The local information extraction module is provided with two layers of equivariant graph neural network connected in sequence, and the global information extraction module includes a self-attention layer and a feedforward neural network layer connected in sequence.

[0173] a prediction module, configured to preprocess the local information and the global information to obtain a graph representation, input the graph representation into the multilayer perceptron, and use the multilayer perceptron to predict drug-target protein binding affinity;

[0174] A prediction performance optimization module is used to construct an atomic interaction matrix for learning the long-range information of atoms in the atomic composite graph, and use the atomic interaction matrix to optimize the performance of the drug-target protein binding affinity prediction model in predicting drug-target protein binding affinity, so as to obtain a drug-target protein binding affinity prediction model with the best prediction performance.

[0175] In this example, for the task of drug-target protein binding affinity prediction, a general framework that simultaneously considers interactions at multiple levels is innovatively proposed. This framework can automatically learn features of different scales at different levels. Secondly, an equivariant graph neural network is used to learn local information of the atomic composite graph at the local level. The equivariant graph neural network has two parts: message passing and aggregation. However, it can fully utilize three-dimensional spatial structural information, maintaining the rotation and translation of coordinates during message passing. Then, at the global and long-range levels, a node feature extraction model and an atomic interaction matrix are innovatively used to learn deep interactions in the atomic composite graph. This improves the interpretability of protein and ligand feature learning, optimizes the predictive performance of the drug-target protein binding affinity prediction model, and effectively enhances the accuracy of drug-target protein binding affinity prediction.

[0176] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting drug-target protein binding affinity, characterized in that: The following steps are involved: S1. Obtaining an atomic composite image of a drug-target protein and related characteristic data of the atomic composite image; S2. Constructing a drug-target protein binding affinity prediction model, wherein the drug-target protein binding affinity prediction model includes a node feature extraction model and a multi-layer perceptron; S3. Using the relevant feature data as input data of the node feature extraction model, inputting the node feature extraction model into the node feature extraction model, and having the node feature extraction model output local information and global information of the atomic composite graph; the node feature extraction model includes a local information extraction module for extracting local information of the atomic composite graph and a global information extraction module for extracting global information of the atomic composite graph in parallel, the local information extraction module having two layers of equivariant graph neural network connected in sequence, and the global information extraction module including a self-attention layer and a feedforward neural network layer connected in sequence; Extracting local information of the atomic composite graph using the local information extraction module includes: S301. The atomic node feature matrix , the atomic three-dimensional coordinate matrix and edge matrix inputting the equivariant graph neural network into the local information extraction module; S302. Using the edge matrix Determine the message of the equivariant graph neural network The transfer direction is based on the atomic node feature matrix and the atomic three-dimensional coordinate matrix , calculate the message as follows: in, represents the SiLU activation function, represents the first learning parameter, , Represents an atomic composite graph The i The three-dimensional coordinates of atoms, Represents an atomic composite graph The j The three-dimensional coordinates of atoms, Represents an atomic composite graph The n The three-dimensional coordinates of atoms, Represents an atomic composite graph The i The characteristic matrix of atoms, Represents an atomic composite graph The j Characteristic matrix of atoms; S303. Based on the message , using a two-layer equivariant graph neural network to analyze the atomic node feature matrix Make the update as follows: in, represents the updated atomic node feature matrix, represents the updated three-dimensional coordinates of the atom, Representation and atomic composite graph The i The neighboring atoms of an atom are represents the second learning parameter, Indicates a concatenation operation; S304. The atomic node feature matrix obtained after the two-layer equivariant graph neural network update is connected using residuals to obtain the local feature matrix of the atomic composite graph. , the local feature matrix is the local information; S4. Preprocessing the local information and the global information to obtain a graph representation, inputting the graph representation into the multilayer perceptron, and using the multilayer perceptron to predict drug-target protein binding affinity; S5. Construct an atomic interaction matrix for learning the long-range information of the atoms in the atomic composite graph, and use the atomic interaction matrix to optimize the performance of the drug-target protein binding affinity prediction model in predicting drug-target protein binding affinity, so as to obtain a drug-target protein binding affinity prediction model with the best prediction performance.

2. The method for predicting drug-target protein binding affinity according to claim 1, wherein The step of obtaining the drug-target protein atomic composite graph and related characteristic data of the atomic composite graph includes: S101. Obtain target protein molecule files and drug molecule files; S102. combining the target protein molecule in the target protein molecule file and the drug molecule in the drug molecule file to obtain a three-dimensional complex of the target protein molecule and the drug molecule; S103. Extracting an atomic composite image consisting of target protein atoms and drug atoms from the three-dimensional complex according to the pocket region where the target protein molecule and the drug molecule bind ; S104. From the atomic composite diagram Extract the relevant feature data, the relevant feature data including the atomic composite graph The atomic node feature matrix in , atomic three-dimensional coordinate matrix and edge matrix .

3. The method for predicting drug-target protein binding affinity according to claim 2, wherein: The preprocessing of the local information and the global information includes: S401. The local feature matrix and the global feature matrix By connecting them in series, we can obtain atomic node features that contain three-dimensional structural information, local information, and global information. ; S402. The atomic node features Perform summing and pooling to obtain the graph representation .

4. The method for predicting drug-target protein binding affinity according to claim 3, wherein: The constructing of an atomic interaction matrix for learning long-range information of atoms in the atomic composite graph includes: S501. Calculate atomic composite graph Edge embedding as follows: in, represents the embedding weight matrix, Represents an atomic composite graph Paired atoms The Euclidean distance, Represents an atomic composite graph The i atoms, Represents an atomic composite graph The j atoms; S502. Based on the edge embedding , calculate the atomic composite graph Paired atomic interactions as follows: in, Indicates return ( , ) as a function of atomic number, represents the Kronecker function, express( , ) of the atoms, represents the first learning weight matrix; S503. Based on the pair-atom interaction , calculate the atomic interaction matrix as follows: in, Represents the second learning weight matrix.

5. The method for predicting drug-target protein binding affinity according to claim 4, wherein: The optimizing the performance of the drug-target protein binding affinity prediction model in predicting drug-target protein binding affinity by using the atomic interaction matrix includes: S511. Using the atomic interaction matrix, calculate the atomic long-range interaction loss function as follows: Among them, Train represents the training set, represents the knowledge-based interaction matrix; S512. Using the multi-layer perceptron to predict drug-target protein binding affinity , The calculation expression is as follows: in, MLP () represents the activation function of the multilayer perceptron; S513. Based on the drug-target protein binding affinity , calculate the prediction loss function of the drug-target protein binding affinity prediction model , The calculation expression is as follows: in, Indicates the true label corresponding to the atomic composite graph; S514. Based on the long-range interaction loss function and the drug-target protein binding affinity , calculate the total loss function for training the drug-target protein binding affinity prediction model , The calculation expression is as follows: in, represents the weight coefficient; S515. Using the total loss function The drug-target protein binding affinity prediction model is trained to obtain a drug-target protein binding affinity prediction model with the best prediction performance.

6. A drug-target protein binding affinity prediction system, the system being implemented based on the drug-target protein binding affinity prediction method according to any one of claims 1 to 5, characterized in that: include: A data acquisition module, configured to acquire an atomic composite image of a drug and a target protein and related characteristic data of the atomic composite image; A model building module is used to build a drug-target protein binding affinity prediction model, wherein the drug-target protein binding affinity prediction model includes a node feature extraction model and a multi-layer perceptron; An input module, configured to input the relevant feature data as input data of the node feature extraction model into the node feature extraction model, and the node feature extraction model outputs local information and global information of the atomic composite graph; a prediction module, configured to preprocess the local information and the global information to obtain a graph representation, input the graph representation into the multilayer perceptron, and use the multilayer perceptron to predict drug-target protein binding affinity; A prediction performance optimization module is used to construct an atomic interaction matrix for learning the long-range information of atoms in the atomic composite graph, and use the atomic interaction matrix to optimize the performance of the drug-target protein binding affinity prediction model in predicting drug-target protein binding affinity, so as to obtain a drug-target protein binding affinity prediction model with the best prediction performance.

7. The drug-target protein binding affinity prediction system according to claim 6, characterized in that: The node feature extraction model includes a parallel local information extraction module for extracting local information of the atomic composite graph and a global information extraction module for extracting global information of the atomic composite graph. The local information extraction module is provided with two layers of equivariant graph neural network connected in sequence, and the global information extraction module includes a self-attention layer and a feedforward neural network layer connected in sequence.

Images