Construction method and system of protein-ligand docking scoring model based on graph Transform

Abstract

The invention discloses a method for constructing a protein-ligand docking scoring model based on a graph Transform, which comprises a graph neural network, a multilayer network based on an attention mechanism and a decoding module, and is characterized in that the decoding module comprises an MND decoding module and a pKd decoding module; the MND decoding module is used for calculating the distance probability distribution between the protein and the ligand; the pKd decoding module is used for predicting an affinity value when the protein is combined with the ligand; the loss function comprises MDN loss and pKd loss, a first layer of the multi-layer network has mask vectors which are updated along with training iteration and are used for indicating that features are reserved or shielded, and docking scoring prediction is carried out by taking the protein graph and the ligand graph as input. According to the scheme, the model simultaneously uses the MDN decoding module and the pKd decoding module, the MDN module optimizes model parameters from the perspective of probability distribution, the pKd module optimizes the model parameters from the perspective of binding affinity, the MDN decoding module and the pKd decoding module carry out docking scoring from the spatial relation and the interaction relation, and the comprehensive evaluation ability of the model can be effectively improved through cooperation of the MDN decoding module and the pKd decoding module.

Description

Technical Field

[0001] The present invention belongs to the field of computer biology, and in particular relates to a method for constructing a protein-ligand docking scoring model based on graph Transformer and a system thereof. Background Art

[0002] As a key step in molecular docking, scoring function is crucial in virtual screening for new drug discovery. The main goal of the scoring function is to quantify the interaction energy between ligands and receptors, including factors such as electrostatic interactions, van der Waals forces, and hydrogen bonds. Traditional scoring functions are mainly based on force fields. This method is based on physical and chemical principles, such as molecular mechanics models, and predicts the binding affinity between ligands and receptors by calculating the interaction energy between molecules. Although the force field-based scoring function is accurate, it has deficiencies in efficiency, flexibility, and modeling of complex interactions and solvent effects. It also lacks the ability to utilize existing biomacromolecule data, resulting in insufficient adaptability and predictive performance.

[0003] In order to solve the above problems, researchers have tried to optimize the protein-ligand docking scoring function by combining data-driven methods. With the advancement of high-throughput screening technology and structural biology methods, relevant protein-ligand complex databases have been established and improved on a large scale. The richness of these databases provides a large amount of high-quality training and verification data for experience-based and knowledge-based scoring functions, significantly improving the accuracy and reliability of molecular docking predictions. For example, ViTScore draws on the vision transformer in the field of vision, uses a 3D grid to voxelize the interaction area, uses a 3D convolutional neural network to learn features, and finally scores the conformation in combination with the distance. However, the voxelization representation method of this method may lose the local details of the protein-ligand docking conformation. Moreover, the currently disclosed methods use all features to learn when learning protein features, while proteins and ligands are actually concentrated on some points rather than all points during the docking process. In general, the current methods can only improve some capabilities. For example, ViTScore only improves the docking ability of the scoring function, and does not improve other capabilities. Because it cannot comprehensively improve the evaluation capabilities of multiple aspects, the model cannot comprehensively evaluate the conformation score. Summary of the invention

[0004] The purpose of the present invention is to propose a method and system for constructing a protein-ligand docking scoring model based on graph Transformer in response to the problems existing in the prior art.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: A method for constructing a protein-ligand docking scoring model based on graph transformer, the model includes a graph neural network connected in sequence, a multi-layer network based on attention mechanism, and a decoding module; The decoding module includes an MND decoding module and a pKd decoding module; The MND decoding module is used to calculate the distance probability distribution between protein and ligand; The pKd decoding module is used to predict the affinity value of protein and ligand binding; The loss function of the model includes MDN loss and pKd loss : , α and β are the weights of MDN loss and pKd loss, respectively; The first layer of the multi-layer network has a mask vector that is updated with each training iteration to indicate whether a feature is retained or masked; The model takes a protein graph and a ligand graph as input, and the graph neural network converts the input into an embedding vector; The embedding vector is input into the multi-layer network for mask processing and then enters the attention mechanism to calculate the node feature representation of the protein and ligand graph; The decoding module performs docking score prediction based on the node features; The model updates model parameters based on the loss function.

[0006] In the above-mentioned method for constructing a protein-ligand docking scoring model based on graph transformer, the three-dimensional spatial relationship between protein and ligand is represented as a two-dimensional graph structure to obtain the protein graph and the ligand graph respectively; The protein graph includes protein node features and protein edge features, and the ligand graph includes ligand node features and ligand edge features.

[0007] In the above-mentioned method for constructing a protein-ligand docking scoring model based on graph transformer, the ligand node features include atom type, hybridization state, formal charge, aromaticity, and the presence of free radical electrons; The ligand edge characteristics are defined by the properties of the bonds between ligand atoms, including bond type, conjugation, and ring participation.

[0008] In the above-mentioned method for constructing a protein-ligand docking scoring model based on graph transformer, the protein node features include residue type, spatial distance between key atoms in the residue, and dihedral angle description of intra-bond rotation; The protein edge features include whether two residues are connected and the distance between some atoms in the two residues.

[0009] In the above-mentioned method for constructing a protein-ligand docking scoring model based on graph transformer, the graph neural network converts the input into an embedding vector in the following way: (1) (2) in and Represent the input node and edge features respectively, i and j are index variables, representing different nodes in the same graph; Representation Node The initial characteristics of Represents the edge features connecting nodes i and j; and is the corresponding weight matrix, and are the bias terms for nodes and edges.

[0010] In the above-mentioned method for constructing a protein-ligand docking scoring model based on graph Transformer, for the input embedding vector , the features after masking It is expressed as: (3) Among them, m is a mask vector, which is used to indicate the retention or shielding of features. Through the element-by-element multiplication operation, the masked features Only those features corresponding to the value 1 in the mask vector are retained.

[0011] In the above-mentioned method for constructing a protein-ligand docking scoring model based on graph transformer, the attention mechanism calculates the node feature representation of the protein and ligand graph in the following way: (11) represents the feature representation of node i in the l+1th layer; RELU represents a nonlinear activation function; Norm Represents a normalization operation; , are two weight matrices; The feature representation of node i after updating at layer l+1 is obtained as follows: (9) i, j are index variables, node i and node j represent different nodes in the same graph; is the feature representation of node i in layer l; represents the linear transformation matrix; ∥ represents connection, k=1…K represents the number of attention heads; is the attention weight, which is related to the query vector of node i, the key vector of node j, and the edge feature vector between nodes ij; is the value vector of feature node j.

[0012] In the above-mentioned method for constructing a protein-ligand docking scoring model based on graph transformer, the MDN loss and pKd loss are: (17) (twenty one) N represents the total number of samples; represents the weight of the relationship between protein node i and ligand node j in the nth sample; Represents a given protein node i and ligand nodes j Features and When the protein node i and ligand nodes j Distance d ( i , j )’s conditional probability distribution; Indicates n The standard deviation of the Gaussian distribution of the relationship between protein node i and ligand node j in samples; Indicates i Prediction of samples pK value; Indicates i The true pK value.

[0013] In the above-mentioned method for constructing a protein-ligand docking scoring model based on graph transformer, the training data used for training the model includes the protein graph and ligand graph, and corresponding label data; The label data include the docking score of protein-ligand and the actual affinity value when the protein and ligand are bound.

[0014] A protein-ligand docking scoring system, comprising a scoring model constructed by the method; The input of the scoring model is the protein map of the target protein to be predicted and the ligand map of the target ligand, and the output is the docking score prediction of the target protein and the target ligand. The advantages of the present invention are: 1. This scheme constructs two-dimensional graphs for the three-dimensional structures of proteins and ligands respectively, and uses graph transformers to build a scoring model. At the same time, feature masks are introduced when using graph transformers to improve the understanding of specific points, which can more accurately capture the interaction between proteins and ligands, thereby improving the docking scoring ability of the model; 2. This model uses two decoding modules, MDN and pKd. MDN optimizes model parameters from the perspective of probability distribution and trains the model to find the most likely binding position and spatial relationship between ligand and protein. The pKd module optimizes model parameters from the perspective of binding affinity and trains the model to evaluate their binding ability by quantifying the affinity of protein-ligand interaction. Both modules score docking from the perspective of spatial relationship and interaction relationship respectively. The cooperation of the two can effectively improve the comprehensive evaluation ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A method for constructing a protein-ligand docking scoring model based on graph Transformer proposed in the present invention; Among them, A is the construction process of the protein-ligand docking scoring model, B is the processing flow of the Transformer part in Figure A; C is the processing flow of the decoding module part in A; Figure 2 yes Figure 1 Schematic diagram of the MDN decoding module; Figure 3 yes Figure 1 Schematic diagram of the pKd decoding module in Figure 2. DETAILED DESCRIPTION

[0016] The present invention provides a method for constructing a protein-ligand docking scoring model based on graph transformer, specifically as follows Figure 1-Figure 3 As shown: First, a two-dimensional graph is used to represent the three-dimensional structure of the protein and ligand. By mapping the three-dimensional spatial arrangement into a two-dimensional graph structure, key features such as atom type, bond type, spatial relationship, and chemical properties are encoded. The key node features of the graph structure of the protein binding pocket include residue types (such as 32 types of amino acids present in the LYS, THR, etc. dataset), spatial distances between key atoms within the residue (such as the distance between N and C atoms), and dihedral angles that describe intrabond rotations. The edge features of the graph structure include whether the two residues are connected, the distance between some atoms within the two residues (such as the distance between Cα atoms within the two residues), etc. This protein graph is represented as ,in, V prot represents the node set in the protein graph ， Each of these nodes v i represents the node feature vector associated with protein residue i; E prot represents the set of edges in the protein graph, where each edge e ij Represents the edge features describing the relationship between protein residues i and j.

[0017] Similarly, when constructing the graph structure of the ligand, atoms are represented as nodes and the bonds between them are represented as edges. The characteristics of each node in the ligand graph include atom type, hybridization state, formal charge, aromaticity, and the presence of free radical electrons. The edges in the graph are defined by the properties of the bonds between atoms, including bond type, conjugation, and ring participation. The ligand graph is formally represented as ， in V lig represents a set of nodes in the ligand graph, each of which v i Represents ligand atom The node feature vector of E lig represents the edge set in the ligand graph, where each edge e ij Represents the edge features that describe the bond properties between ligand atoms i and j.

[0018] The model of this scheme uses a multi-head self-attention mechanism that integrates graph neural networks and masks to calculate node and edge features. The input to the model is the protein graph and ligand graph obtained above. The initial features of nodes and edges are generated by the graph neural network, and the node features and edge features of ligands and proteins are converted into d-dimensional embedding vectors, as shown in the following formula: (1) (2) in and Represent the input node and edge features respectively, that is, represents the ligand node i features and protein node i features; Represents the ligand edge features connecting ligand node ij and the protein edge features connecting protein node ij. When the conversion object is a protein graph, represents the protein node i feature, Indicates the protein edge features connecting protein nodes ij, when the conversion object is a ligand graph represents the characteristics of ligand node i and protein node i, Represents the features of the ligand edges connecting ligand nodes ij.

[0019] Representation Node The initial characteristics of Represents a connected node and edge features.

[0020] and is the corresponding weight matrix, and are the bias terms for nodes and edges.

[0021] After the initial features are obtained, they are input into a multi-layer network based on the attention mechanism. The first layer of the multi-layer network has a mask vector that is updated with the training iteration to indicate whether the features are retained or masked. Certain structural elements of proteins and ligands are selectively masked by feature masks. , the features after masking It is expressed as: (3) Among them, m is a mask vector that indicates which features are retained (value 1) or masked (value 0). Through the element-by-element multiplication operation, the masked features Only the parts corresponding to the features whose value in the mask vector is 1 are retained. In a multi-layer model, this scheme only uses the masking mechanism in the first layer and does not use masks in subsequent layers.

[0022] After the initial feature representation after masking is established in the above way, the attention mechanism calculation phase begins. The attention score of the edge connecting adjacent nodes is calculated by the following formula: In the above formula, k represents the attention head, l represents the current layer number, and formulas (4), (5), (6), and (7) are used to calculate the feature representation of the query vector, key vector, value vector, and edge, respectively. , , , They represent the query weight matrix, key weight matrix, value weight matrix and edge weight matrix of the k-th attention at the l-th level respectively.

[0023] The function Norm represents the normalization operation of node features to stabilize and improve the training process. Figure 1 In the example, node i is Processing is used to generate query vectors (Query) to ask other nodes for relevant information; node j is and The processing is used to generate the key vector (Key) and the value vector (Value). The key vector determines the degree of match with the query.

[0024] Then, the attention weight is calculated by the following formula: (8) The update of node i in subsequent layers is expressed by the following formula: Indicates the dimension of the key vector.

[0025] Indicated in l Layer k In an attention head, the node i The query vector of Indicated in l Layer k In an attention head, the node j The key vector of ; Indicated in l Layer k In an attention head, the node i and nodes j The edge feature vector between .

[0026] in, , represents the linear transformation matrix, ∥ represents the connection, k=1…K represents the number of attention heads, The feature representation of node i in layer l is the node feature of the current layer and will be used to update the next layer; The updated feature representation of node i at the l+1th layer takes into account the information of node i itself and its neighboring nodes.

[0027] The final node and edge features are: It represents the feature representation of node i in the l+1th layer, which is the result of updating the node features; It represents the updated feature representation of node i at layer l+1, which is calculated by aggregating the features of neighboring nodes and combining them with its own features.

[0028] After obtaining the final node feature representation of protein and ligand and After that, they are input into the decoding module designed in this scheme. The decoding module consists of two parts: MND module and pKd decoding module. The MND module is used to calculate the probability of the distance distribution between the protein and the ligand, which is then further used to calculate the score; the pKd decoding module is used to predict the affinity value when the protein and the ligand bind, which is approximately regarded as the logKa value. The loss of the model includes MDN loss and pKd loss.

[0029] Use the following formula to update the node representation after the protein and ligand are connected: (13) in, Represents the feature vector of protein node i; Represents the feature vector of ligand node j. Concat( hiprot , hjlig ) means concatenating the feature vectors of protein node i and ligand node j to form a longer feature vector.

[0030] ELU represents the exponential linear unit activation function, BatchNorm represents batch normalization, and Dropout represents the dropout operation.

[0031] Mean used for Gaussian mixture models in the MDN module , Standard Deviation and the mixing coefficient Calculated by the following formulas: in, , , is a learnable weight matrix.

[0032] Finally, the MDN loss and score calculation are expressed as: Formula (17) is used to measure the difference between the probability distribution predicted by the model and the actual data. It captures the spatial relationship between the protein and the ligand by predicting the distance distribution between them, thereby evaluating whether docking is likely to occur. During the training process, the network parameters are adjusted to make the predicted distribution as close to the actual distribution as possible.

[0033] Where N represents the total number of samples; It represents the weight or mixing coefficient of the relationship between nodes i and j in the nth sample, and determines the contribution of each Gaussian distribution to the overall distribution.

[0034] Indicates n In the samples, the nodes i and j The standard deviation of the Gaussian distribution between the two describes the uncertainty or dispersion of the distribution.

[0035] Indicates ρ ( i , j , n ) is the mean value, σ ( i , j , n ) is the probability density function of the Gaussian distribution with standard deviation, which is used to describe the node i and j Distance d ( i , j )’s probability distribution.

[0036] Formula (18) calculates the score based on the distance probability of MDN loss. The score reflects the possibility of binding between protein and ligand. It is calculated based on the spatial relationship of formula (17). This score represents the probability of successful docking.

[0037] Where I and J represent the total number of protein nodes and ligand nodes, respectively; Represents a given protein node i and ligand nodes j Features and When the node i and j Distance d ( i , j ), which reflects the prediction accuracy of the model for the distance of a specific protein-ligand pair.

[0038] , Represent protein nodes i and ligand nodes j The eigenvector of Represents a protein node i and ligand nodes j The actual distance between.

[0039] pKd loss is related to the predicted binding affinity, derived from the dissociation constant The negative logarithm of the binding free energy ∆G and the dissociation constant The relationship is as follows: (19) At standard room temperature (T=298.15K), the dissociation constant The relationship between the negative logarithm of pKd and the binding free energy ∆G can be expressed as: (20) The pKd loss is calculated by the mean squared error between the predicted pKd value and the true pKd value: (twenty one) Formula (21) is used to measure the difference between the predicted pKd value and the true pKd value and optimize the model parameters by minimizing the sum of squares of the errors. Where N represents the total number of samples; Indicates i Prediction of samples pK value; Indicates i The true pK value.

[0040] Therefore, the overall loss function of the feature decoding module is the weighted sum of the MDN loss and the pKd loss: (twenty two) Among them, α and β are the weights of MDN loss and pKd loss respectively, It is used to train the model's docking ability. The goal of docking is to find the most likely binding position and spatial relationship between the ligand and the protein. The scoring ability used to train the model. The goal of scoring is to evaluate the binding ability of protein-ligand interactions by quantifying their affinity or stability. In general, this solution improves the docking and screening capabilities of the model through MDN, and improves the scoring and ranking capabilities of the model through pKd, and finally achieves the improvement of the comprehensive evaluation ability of the model. In the above formulas (2) to (12), nodes i and j represent two nodes in the same graph. For example, for the representation of proteins, i and j represent protein node i and protein node j respectively, and the ligands are similar.

[0041] Formulas (13) to (18) are all based on the feature representation after cascading in Formula 13, so i and j represent two nodes in the cascade feature, i represents the protein node, and j represents the ligand node.

[0042] The protein-ligand docking scoring model is constructed and trained using the corresponding training data in the above manner, and can be applied to score the conformations generated by protein-ligand molecular docking. The training data includes protein graphs and ligand graphs and corresponding label data. The protein graphs and ligand graphs contain the node features and edge features of the corresponding proteins and ligands. The label data includes the scores of the conformations generated by protein-ligand docking and the actual affinity values ​​of the proteins and ligands when they bind.

[0043] This example tests the effects of this method and other existing methods on the CASF-2016 standard scoring function test data set, using the Top1 docking success rate of docking ability and the Pearson of scoring ability as indicators (the larger the better).

[0044] Table 1 Comparison of the results of this method and other methods in CASF-2016 GlideScore-SP is a scoring function for predicting the binding affinity of protein-ligand complexes. It is an empirically based scoring function designed to maximize the discrimination between compounds with strong binding affinity and those with little or no binding ability.

[0045] DrugScore2018 is a new version of the knowledge-based scoring function DrugScore.

[0046] X-Score is a scoring function, which is usually used in conjunction with molecular docking programs (such as DOCK, AutoDock, FlexX or GOLD) to provide binding models of molecules to their target proteins in structure-based drug design studies. X-Score can then be used to re-rank them and more accurately estimate the binding free energy of these molecules.

[0047] RTMScore is a deep learning-based scoring function that uses the Transformer model to capture the interaction between proteins and ligands, mainly improving docking and screening capabilities.

[0048] As can be seen from Table 1, according to the experimental results, the proposed method in this scheme has achieved the best docking ability and scoring ability results on the CASF-2016 dataset. In general, compared with other methods, the proposed method in this scheme has higher result accuracy.

[0049] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A method for constructing a protein-ligand docking scoring model based on graph Transformer, characterized in that: The model includes a graph neural network connected in sequence, a multi-layer network based on the attention mechanism, and a decoding module; The decoding module includes an MND decoding module and a pKd decoding module; The MND decoding module is used to calculate the distance probability distribution between protein and ligand; The pKd decoding module is used to predict the affinity value of protein and ligand binding; The loss function of the model includes MDN loss and pKd loss : , α and β are the weights of MDN loss and pKd loss, respectively; The first layer of the multi-layer network has a mask vector that is updated with each training iteration to indicate whether a feature is retained or masked; The model takes a protein graph and a ligand graph as input, and the graph neural network converts the input into an embedding vector; The embedding vector is input into the multi-layer network for mask processing and then enters the attention mechanism to calculate the node feature representation of the protein and ligand graph; The decoding module performs docking score prediction based on the node features; The model updates model parameters based on the loss function.

2. The method for constructing a protein-ligand docking scoring model based on graph transformer according to claim 1, characterized in that: The three-dimensional spatial relationship between the protein and the ligand is represented as a two-dimensional graph structure to obtain the protein graph and the ligand graph respectively; The protein graph includes protein node features and protein edge features, and the ligand graph includes ligand node features and ligand edge features.

3. The method for constructing a protein-ligand docking scoring model based on graph transformer according to claim 2, characterized in that: The ligand node characteristics include atom type, hybridization state, formal charge, aromaticity, and the presence of free radical electrons; The ligand edge characteristics are defined by the properties of the bonds between ligand atoms, including bond type, conjugation, and ring participation.

4. The method for constructing a protein-ligand docking scoring model based on graph transformer according to claim 2, characterized in that: The protein node features include residue type, spatial distance between key atoms within the residue, and dihedral angle description of intrabond rotation; The protein edge features include whether two residues are connected and the distance between some atoms in the two residues.

5. The method for constructing a protein-ligand docking scoring model based on graph transformer according to claim 1, characterized in that: The graph neural network converts the input into an embedding vector as follows: （1） （2） in and Represent the input node and edge features respectively, i and j are index variables, representing different nodes in the same graph; Representation Node The initial characteristics of Represents the edge features connecting nodes i and j; and is the corresponding weight matrix, and are the bias terms for nodes and edges.

6. The method for constructing a protein-ligand docking scoring model based on graph transformer according to claim 5, characterized in that: For the input embedding vector , the features after masking It is expressed as: （3） Among them, m is a mask vector, which is used to indicate the retention or shielding of features. Through the element-by-element multiplication operation, the masked features Only those features corresponding to the value 1 in the mask vector are retained.

7. The method for constructing a protein-ligand docking scoring model based on graph transformer according to claim 1, characterized in that: The attention mechanism calculates the node feature representation of the protein and ligand graphs in the following way: (11) represents the feature representation of node i in the l+1th layer; RELU represents a nonlinear activation function; Norm Represents a normalization operation; , are two weight matrices; The feature representation of node i after updating at layer l+1 is obtained as follows: （9） i, j are index variables, node i and node j represent different nodes in the same graph; is the feature representation of node i in layer l; represents the linear transformation matrix; ∥ represents connection, k=1…K represents the number of attention heads; is the attention weight, which is related to the query vector of node i, the key vector of node j, and the edge feature vector between nodes ij; is the value vector of feature node j.

8. The method for constructing a protein-ligand docking scoring model based on graph transformer according to claim 5, characterized in that: The MDN loss and pKd loss are: （17） （21） N represents the total number of samples; represents the weight of the relationship between protein node i and ligand node j in the nth sample; Represents a given protein node i and ligand nodes j Features and When the protein node i and ligand nodes j Distance d ( i , j )’s conditional probability distribution; Indicates n The standard deviation of the Gaussian distribution of the relationship between protein node i and ligand node j in samples; Indicates i Prediction of samples pK value; Indicates i The true pK value.

9. The method for constructing a protein-ligand docking scoring model based on graph transformer according to any one of claims 1 to 8, characterized in that: The training data used for the training model includes the protein map and ligand map, as well as the corresponding label data; The label data include the docking score of protein-ligand and the actual affinity value when the protein and ligand are bound.

10. A protein-ligand docking scoring system, characterized in that: A scoring model constructed by the method according to any one of claims 1 to 9; The input of the scoring model is the protein map of the target protein to be predicted and the ligand map of the target ligand, and the output is the docking score prediction of the target protein and the target ligand.

Images