A fast and accurate protein-small molecule ligand docking method based on deep learning

Abstract

The present invention discloses a method and system for fast and accurate docking of protein-small molecule ligands based on deep learning, which belongs to the field of computer application technology. The method innovatively adopts deep learning technology to achieve efficient and accurate prediction of protein-small molecule ligand docking. Through undirected composite graph and KNN graph representation, it deeply captures the complex interactions within molecules and between proteins, and constructs high-quality data input; dual encoder mechanism analysis features, generates deep embedding representation, and fully connected interaction graphs strengthen interaction understanding; combines conformation prediction, force field optimization and conformation alignment to improve docking posture accuracy; mixed density neural network model predicts distance probability and affinity, and enhances prediction diversity and accuracy. The system has a modular design and is easy to deploy across platforms. It is widely used in drug development and life science research, accelerates new drug discovery, and reveals molecular mechanisms, with broad prospects.

Description

Technical Field

[0001] The present invention belongs to the field of computer application technology, and in particular relates to a method and system for rapid and accurate protein-small molecule ligand docking based on deep learning. Background Art

[0002] Molecular docking technology is an important computational biology method that predicts the binding posture and affinity between proteins and small molecule ligands by simulating the interaction between molecules. This technology has a wide range of applications in drug design, protein function research, and bioinformatics. Molecular docking methods are mainly divided into three types: rigid docking, semi-flexible docking, and flexible docking. Among them, flexible docking allows the conformation of the research system to change freely during the docking process, so it is suitable for accurately examining the recognition between molecules, but the amount of calculation is relatively large. In the field of drug design, molecular docking technology can predict the binding mode of drug candidates with target proteins, providing important guidance for drug discovery and design. By simulating the interaction between proteins and small molecules, scientists can screen out potential drug molecules with high affinity to specific target proteins, thereby accelerating the process of drug development. As an advanced computational method, the core value of virtual screening technology lies in its ability to efficiently screen out potential binding molecules with biological activity to specific target proteins from a huge database of compounds. As a core component of structure-based VS strategy, small molecule ligand docking plays a vital role. It not only involves the precise construction of protein-small molecule ligand binding sites, but is also responsible for predicting the binding affinity between the two and guiding the selection of binding sites, thus playing a bridging role in the entire virtual screening process.

[0003] Traditional docking tools such as AutoDock, AutoDock Vina, GOLD, Glide and LeDock rely on heuristic search strategies to traverse and evaluate many possible small molecule ligand conformations, combined with carefully designed scoring functions to simplify the process of ligand posture optimization and binding strength estimation. This strategy significantly improves the execution efficiency of large-scale virtual screening tasks. However, traditional docking tools and existing deep learning tools are limited in speed, posture quality and binding affinity accuracy. In particular, when receptors need to be docked with thousands of small molecule ligands, this process will consume a lot of time and may not be accurate. Deep learning algorithms such as AutoDock GPU have significantly improved the speed of virtual screening by optimizing algorithms and utilizing high-performance computing resources. They have made significant progress in predicting protein-small molecule ligand conformations, but they still face challenges in the accuracy and speed of binding affinity prediction. The current model has limited generalization ability in complex systems and high computational costs, which makes it difficult to meet the needs of large-scale virtual screening for efficient and accurate predictions.

[0004] In summary, traditional docking tools and existing deep learning tools are limited in speed, pose quality and binding affinity accuracy. A receptor often requires docking experiments with thousands of small molecule ligands, which consumes a lot of time and has low accuracy. Therefore, continuous technological innovation and optimization, especially improvements in model architecture, training strategies and data processing methods, are of vital importance to improving the application effect of deep learning in drug development. Summary of the invention

[0005] In view of the shortcomings of traditional docking and deep learning tools in docking speed, posture prediction accuracy and binding affinity calculation, there is an urgent need for a method that can improve the efficiency and accuracy of large-scale small molecule ligand and receptor docking to accelerate the drug development process. The present invention aims to provide a protein-small molecule ligand fast and accurate docking method and system based on deep learning.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The present invention provides a method for rapid and accurate protein-small molecule ligand docking based on deep learning, comprising:

[0008] S1, obtain protein raw data and small molecule ligand raw data;

[0009] S2, structurally representing the raw data of small molecule ligands and protein raw data to construct the small molecule ligand undirected composite graph and protein KNN graph respectively;

[0010] S3, inputting the small molecule ligand undirected composite graph into the small molecule ligand encoder to generate the updated embedding representation of the small molecule ligand; inputting the protein KNN graph into the protein encoder to generate the updated embedding representation of the protein, and then constructing a fully connected interaction graph containing the interaction information between the small molecule ligand and the protein;

[0011] S4, inputting the fully connected interaction graph containing the interaction information between small molecule ligand and protein into the binding conformation prediction model to predict the protein-small molecule ligand docking posture conformation;

[0012] S5, input the protein-small molecule ligand docking posture conformation into a pre-trained mixed density neural network scoring function model to obtain the affinity score and possible binding posture between the small molecule ligand and the protein.

[0013] In S1, the protein raw data and small molecule ligand raw data are all protein-small molecule raw data after removing the core set from the general set of the PDBbind2020 database.

[0014] In S2, the construction of the small molecule ligand undirected composite graph and protein KNN graph includes:

[0015] S21, the raw data of small molecule ligands are represented as undirected composite graphs using the torch_geometric package , with atoms as nodes and covalent bonds as edges, where is the node feature, is the edge feature, represents atomic coordinates;

[0016] S22, the protein raw data uses protein residues as nodes and connects the first 30 nearest neighbor nodes to construct a KNN graph , nodes are represented by residues, is the protein node feature, For protein edge features, the carbon atom coordinates in the residues are used as node positions X p , each node or edge in the graph contains a scalar feature and vector features , is the scalar feature of the protein node, is the scalar feature of the protein edge, is the vector feature of the protein node, is the vector feature of the protein edge, the scalar feature includes the relative distance and the position distance, and the vector feature includes the unit vector of the direction, the side chain direction and the edge direction.

[0017] In S3, the construction of a fully connected interaction graph containing information on the interaction between small molecule ligands and proteins includes:

[0018] S31, inputting the undirected composite graph data of the small molecule ligand into the small molecule ligand encoder to learn the atomic interactions inside the small molecule ligand, and generating an updated embedding representation of the small molecule ligand, wherein the embedding representation includes node features, edge features and node position information;

[0019] S32, inputting protein KNN graph data into a protein encoder, learning the interactions within protein residues and the topological and geometric features between them, and generating an updated embedding representation of the protein, wherein the embedding representation includes information of neighbor nodes of the residue nodes and the topological and geometric features learned by the protein encoder;

[0020] S33, using the updated embedding representation of the small molecule ligand obtained in S31 and the updated embedding representation of the protein obtained in S32, a fully connected interaction graph containing the interaction information between the small molecule ligand and the protein is constructed.

[0021] Furthermore, the protein encoder is a geometric bottleneck perceptron.

[0022] Furthermore, the small molecule ligand encoder is a graphic transformer network.

[0023] Furthermore, the edge features consist of edge types single bond, double bond, triple bond, aromatic bond and non-bonded interaction and distance between nodes.

[0024] Furthermore, in S31, the generating of the updated embedded representation of the small molecule ligand includes: initializing the characteristics of each node and each edge chemical bond in the undirected composite graph of the molecular ligand as a d-dimensional embedding vector; inputting the initialized d-dimensional embedding vector into the small molecule ligand encoder; and using a multi-head self-attention mechanism to process the interaction between nodes to update the embedding representation of each node, thereby obtaining the updated embedded representation of the small molecule ligand.

[0025] Further, in S32, the generating of the updated embedded representation of the protein includes: mapping and updating the scalar features and vector features of each residue node in the protein KNN graph, reducing and expanding the vector features, and jointly integrating the scalar features to generate the final updated vector features and scalar features; in each layer of the protein encoder, using the KNN algorithm to select the 30 nearest neighbor nodes of each node, the node and edge information is connected to form an extended feature vector, and the extended feature vector is integrated and updated through the protein encoder to obtain the updated embedded representation of the protein.

[0026] In S4, the binding conformation prediction module is a graph neural network module, which is composed of 8 graph neural network layers improved based on self-attention. The prediction model predicts the protein-small molecule ligand docking posture conformation including: taking the updated embedding representation in the interaction graph containing the interaction information of small molecule ligand and protein as input, and predicting the protein-small molecule ligand docking posture through a multi-level self-attention mechanism; introducing a gating equation to combine the updated embedding representation with the original embedding representation, optimize the prediction results, and obtain the predicted protein-small molecule ligand docking posture conformation.

[0027] Before performing the S5 operation, it also includes an operation of optimizing the protein-small molecule ligand docking posture conformation, including force field optimization or conformation alignment processing.

[0028] In S5, the pre-trained mixed density neural network scoring function model includes:

[0029] S51, inputting the protein-small molecule ligand docking posture conformation into a mixed density neural network scoring function model, wherein the protein-small molecule ligand docking posture conformation includes a node embedding representation of the protein and a node embedding of the small molecule ligand, as well as their position information;

[0030] S52, uses a protein encoder and a self-attention-based graph neural network to process the node embedding representation of proteins and the node embedding of small molecule ligands respectively, captures the structural features of proteins and the chemical features of small molecule ligands, connects the processed feature vectors sequentially, and constructs a joint feature vector containing both protein and small molecule ligand information;

[0031] S53, passing the joint feature vector constructed in S52 to a network layer of a graph neural network, wherein the network layer includes a linear layer, a batch normalization layer, an ELU activation function, and a Dropout layer, and the combined network layer extracts high-level features;

[0032] S54, three linear layers are used to predict the mean, standard deviation and mixing coefficient respectively, which together define the probability distribution of the minimum distance between the protein and the small molecule ligand; another set of linear layers is used to directly predict the affinity score;

[0033] S55, define the loss function, which includes two parts: the error between the probability distribution output by the mixed density neural network and the true distance, and the cross entropy loss between the affinity score and the standard answer; use the optimization algorithm to minimize the loss function, set early stopping until the loss value no longer changes, which means that the scoring part has fully learned the characteristics of protein-small molecules, and obtain a pre-trained mixed density neural network scoring function model.

[0034] The system for realizing the above-mentioned protein-small molecule ligand fast and accurate docking method based on deep learning includes:

[0035] Data acquisition module, used to obtain protein raw data and small molecule ligand raw data;

[0036] The structured characterization building module is used to perform structured characterization on the raw data of small molecule ligands and proteins, respectively, and to construct the small molecule ligand undirected composite graph and protein KNN graph;

[0037] The interaction graph construction module is based on the small molecule ligand encoder and the protein encoder to obtain the updated embedding representation of the small molecule ligand and the updated embedding representation of the protein, and to construct a fully connected interaction graph containing the interaction information between the small molecule ligand and the protein;

[0038] Combined with the conformation prediction module, it is used to predict the docking posture conformation of protein-small molecule ligand;

[0039] Scoring function model module, which evaluates the affinity score and possible binding pose of the predicted docking pose conformation;

[0040] Post-processing and optimization module, used to perform force field optimization or conformation alignment post-processing on the predicted protein-small molecule ligand docking posture conformation to obtain the optimized protein-small molecule ligand docking posture conformation;

[0041] The output and visualization module is used to output the predicted docking posture conformation, affinity score and possible binding posture, and provide visualization function.

[0042] The present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned method for rapid and accurate protein-small molecule ligand docking based on deep learning are implemented.

[0043] The present invention provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the above-mentioned protein-small molecule ligand rapid and accurate docking method based on deep learning are implemented.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] The protein-small molecule ligand fast and accurate docking method based on deep learning provided by the present invention can efficiently and accurately predict the docking posture of protein and small molecule ligand through deep learning technology, and significantly improves the calculation efficiency and prediction accuracy compared with the traditional docking method. The method of the present invention characterizes the small molecule ligand and protein data as an undirected composite graph and a KNN graph respectively. This structured representation method effectively captures the atomic interactions within the molecule and the spatial relationship between protein residues, and provides high-quality data input for subsequent deep learning; based on the small molecule ligand encoder and the protein encoder, the complex features of the small molecule ligand and the protein are learned respectively, and the updated embedding representation is generated. This design enables the model to have a deeper understanding of the interaction between molecules, and the fully connected interaction graph is constructed using the updated embedding representation of the small molecule ligand and the protein. The graph contains rich interaction information and helps to predict the docking posture more accurately; by combining the conformation prediction model to predict the docking posture, and introducing force field optimization and conformation alignment post-processing, the accuracy of the docking posture is further improved, and the mixed density neural network model is used as the scoring function, which can simultaneously predict the minimum distance probability distribution and affinity score between the protein and the small molecule ligand, and improve the diversity and accuracy of the prediction.

[0046] Furthermore, by using the torch_geometric library, the small molecule ligand is represented as an undirected composite graph, which effectively captures the interaction between atoms. The structural representation of the protein is constructed by the KNN graph, and the geometric characteristics of the protein are fully considered by using the residue and carbon atom coordinates. The small molecule ligand encoder and protein encoder are used to perform deep learning on their respective graph data, which can capture the complex interactions between atoms and residues as well as their topological and geometric characteristics. The geometric bottleneck perceptron is used as the protein encoder to further enhance the model's ability to understand the three-dimensional structure of the protein. Multiple features such as edge type, distance and other non-bonded interactions are integrated into the embedding representation, which improves the accuracy and robustness of the model. The improved graph neural network module based on self-attention is used to predict the docking posture. The multi-level self-attention mechanism enhances the information interaction between nodes and improves the accuracy of the docking posture prediction. The gating equation is introduced to combine the updated embedding with the original embedding, which further improves the accuracy of the prediction result. The mixed density neural network is used as the scoring model, which can simultaneously predict the affinity score and distance distribution, providing more comprehensive information for evaluating the binding between small molecule ligands and proteins.

[0047] The system provided by the present invention constructs a complete system. The system adopts a modular design, including multiple modules such as data acquisition, data preprocessing, feature learning, interaction graph construction, docking posture prediction, post-processing, scoring model and optimization training. The modules are relatively independent of each other, which is convenient for maintenance and expansion. An implementation method of computer equipment and computer-readable storage media is provided, so that the method can be deployed and applied on multiple computing platforms, enhancing the versatility and portability of the technology. The present invention can be applied to the field of drug research and development, quickly and accurately predict the binding sites and affinities of small molecule drugs and protein targets, accelerate the process of new drug discovery and development, and can also be used to study the protein-ligand interaction mechanism and reveal the molecular regulation mechanism in life processes. It has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flow chart of the method for rapid and accurate protein-small molecule ligand docking based on deep learning provided by the present invention;

[0049] Figure 2It is a predicted conformation of protein 4qac binding to a small molecule ligand generated by the protein-small molecule ligand fast and accurate docking method based on deep learning provided by the present invention; wherein, (a) is the real structure of 4qac protein-small molecule docking in reality, (b) is the structure generated by the model-predicted 4qac protein-small molecule docking without post-processing, (c) is the predicted structure of 4qac protein-small molecule docking after force field post-processing to ensure reasonable bond length, and (d) is the predicted structure of 4qac protein-small molecule docking after conformation alignment post-processing to ensure reasonable angle;

[0050] Figure 3 3 is a schematic diagram of the results of a benchmark test of the deep learning-based protein-small molecule ligand fast and accurate docking method provided in an embodiment of the present invention on the PDBbind core set, wherein (a) is the initial RMSD distribution test result of using the model to test the PDBbind core set after training, (b) is the RMSD distribution test result of using the model to test the PDBbind core set after training and performing force field post-processing, and (c) is the RMSD distribution test result of using the model to test the PDBbind core set after training and performing conformational alignment post-processing. DETAILED DESCRIPTION

[0051] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0052] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0053] See attached Figure 1 The method for rapid and accurate protein-small molecule ligand docking based on deep learning provided by the embodiment of the present invention comprises the following steps:

[0054] S1, obtain protein raw data and small molecule ligand raw data.

[0055] The protein raw data and small molecule ligand raw data are all protein-small molecule raw data after removing the core set from the general set of the PDBbind2020 database;

[0056] S2, the small molecule ligand raw data and protein raw data are structured to construct the small molecule ligand undirected composite graph and protein KNN graph respectively.

[0057] S21, use torch_geometric to represent the small molecule ligand feature data as an undirected composite graph , with atoms as nodes and covalent bonds as edges, where is the node feature, is the edge feature, represents atomic coordinates.

[0058] S22, in order to reduce calculations and to better update the geometric information of residues and capture the interactions between residue pairs and protein-small molecule ligands, protein residues are used as nodes in the original protein data, and the first 30 nearest neighbor nodes are connected to construct a KNN graph. , nodes are represented by residues, is the protein node feature, For protein edge features, the carbon atom coordinates in the residues are used as node positions X p , each node or edge in the graph contains a scalar feature and vector features , is the scalar feature of the protein node, is the scalar feature of the protein edge, is the vector feature of the protein node, is the vector feature of the protein edge, the scalar feature includes the relative distance and the position distance, and the vector feature includes the unit vector of the direction, the side chain direction and the edge direction.

[0059] S3, inputting the small molecule ligand undirected composite graph into the small molecule ligand encoder to generate the updated embedding representation of the small molecule ligand; inputting the protein KNN graph into the protein encoder to generate the updated embedding representation of the protein, and then constructing a fully connected interaction graph containing the interaction information between the small molecule ligand and the protein;

[0060] S31, using the graph transformer network as the small molecule ligand encoder, learns the interaction between small molecule ligands. The specific process is as follows:

[0061] (1) Node features of the i-th node , the edge feature between node i and node j , initialized to d dimensions through two linear layers and :

[0062]

[0063]

[0064] in, , , , , is a learnable matrix for linear transformation;

[0065] (2) Input the initialized node and edge embeddings into the graph transformer network in The following operations are performed in the layer: k For the k A self-attention head, j Representation Node i The adjacent nodes of Representative layers:

[0066]

[0067]

[0068]

[0069]

[0070] in , , , It is the learnable parameter obtained by the linear layer, and Norm is the layer normalization. Representative The node updated when the layer i The node characteristics of Representative The node updated when the layer j The node characteristics of Representative The node updated when the layer i and nodes j The edge features between yes Through linear layers and learnable parameters The new query value for the update, is through and learnable parameters The updated key value, is through and learnable parameters The actual value of the update, represent Through linear layers and learnable parameters Updated Node i and nodes j The new edge features between them.

[0071] For adjacent nodes j Perform SoftMax operation:

[0072]

[0073] In the above formula is the dimension of each attention head, is the new query value updated according to the previous formula, is the key value updated according to the previous formula, is the new edge feature between node i and node j updated according to the previous formula, is the output after multi-head self-attention.

[0074] Update the node features according to the newly learned information after Softmax:

[0075]

[0076]

[0077] In the above formula, Representative The node updated when the layer i The node characteristics of Representative Nodes at the layer i The node characteristics of Represents the node i Its neighboring nodes j The information of the connected edges is summed up, Concat means connection, represents the number of self-attention heads, are the learnable parameters obtained by the linear layer, is the output after multi-head self-attention, In the previous formula, and learnable parameters The actual value of the update, Represents the number of self-attention heads.

[0078] The edge information is connected according to the learned new molecular interaction information and then updated by normalization and activation function:

[0079]

[0080]

[0081]

[0082] In the above formula, represents the number of self-attention heads, Representative Nodes at the layer i and nodes j The edge features between Represents the updated Nodes at the layer i and nodes j The edge features between is the output after multi-head self-attention, Representative Nodes at the layer i The node characteristics of Represents the updated Nodes at the layer i The node characteristics of Represents the last output Nodes at the layer i and nodes j The edge features between, Norm represents batch normalization, SiLU represents an activation function, where ; , , , are the learnable parameters obtained by the linear layer.

[0083] Complete The graph transformer network layer uses message passing and an improved multi-head self-attention mechanism to update the edge information embedding operation of each node as shown above. The above operation is superimposed 6 times in total, and the final node and edge information is output for embedding.

[0084] S32 uses the improved geometric bottleneck perceptron as the protein encoder to learn the interactions within protein residues. The specific process is as follows:

[0085] (1) Update node embeddings based on topological connections and geometric features within and between nodes.

[0086] The expressions of scalar features and vector features are represented by two mapping functions in the geometric bottleneck perceptron module: Fv:(V) →(z, V') and Fs:(s,z)→(s') Co-generation.

[0087] in s,V are scalar and vector features, and s’ and V' are the updated scalar and vector features. The key to the vector representation path is to aggregate the vector features into a potential permutation and rotation equivariant representation and pass it to the scalar path. For each node or edge vector feature, the geometric bottleneck perceptron generates a latent representation z as shown below. The input feature vector with representation dimension r is Downscaling, are learnable parameters:

[0088]

[0089] To generate the vector representation V’ , the geometric bottleneck perceptron will z Convert to Depth r' The expression is V’ The depth of the intermediate representation By learnable parameters calculate:

[0090]

[0091] (2) The geometric bottleneck perceptron updates the vector representation through the gating operation defined in the following formula to generate an updated vector representation V’ ,in, Expressed as norm, For application The nonlinearity of the norm, is the element-by-element multiplication. When it is a Sigmoid function, this operation acts as a gating mechanism for the intermediate representation. The intermediate representation obtained from the previous formula is:

[0092]

[0093] (3) In geometric bottleneck perceptron learning, it is necessary to jointly integrate vector and scalar features to aggregate information flow. The scalar features of each node or edge are updated by fusing the potential vector features on the vector representation path, thereby fully modeling the relationship and geometric information in the graph. The aggregation operation of scalar and transformed vector information in the scalar channel can be written as: Expressed as Norm:

[0094]

[0095] in , t is the depth of s. Then the updated s' is obtained by the following formula:

[0096]

[0097] in, The depth of , projected onto s' , s' The depth is expressed as t' );

[0098] in, represents nonlinearity, The representations are learnable parameters that perform a linear transformation.

[0099] The above is a complete message passing process of the geometric bottleneck perceptron module.

[0100] S33, after initialization, the node information and edge information will be sent to the repeated geometric bottleneck perceptron Conv module for message transmission, as shown below:

[0101] For nodes ,we will N(i) Represented as a node The neighbor nodes of , use KNN to select the 30 nearest nodes of the attachment, and the single layer of the geometric bottleneck perceptron can be written as:

[0102]

[0103] in, k represents the depth of the network, Representative k Nodes of the layer network i Related information, Representative Layer Node i Related information, Representative Layer Node j Related information, is a differentiable function, express k The differentiable function that the layer network passes through, F represents a permutation invariant aggregate function, Representative Aggregation The information of all neighbor nodes, Indicated by The message generation function generated by the geometric bottleneck sensor layer is k Layer network until Iterations are generated, and the output is , Indicate point i With Node j The side information between .

[0104] Specifically, it can be further divided into the following four steps. ∪ represents the connection operation. First, the vector and scalar representations and their side information are connected, and then integrated by the geometric bottleneck perceptron layer. The message generated by the layer, GBP stands for Geometric Bottleneck Perceptron layer, The message generation function is recursively generated by the geometric bottleneck perceptron until the Iteration generation, output , reuse F Function performs message aggregation, where FIt is a permutation-invariant function including MEAN, SUM or MAX, which is used to aggregate the generated messages and finally update the node features:

[0105]

[0106]

[0107]

[0108]

[0109] Indicates Layer Points i With Node j The side information between Indicated by k The message generation function generated by the layer geometry bottleneck perceptron layer, Representative Layer Node i Related information, Representative Layer Node j Related information, Indicates k Layer Geometry Bottleneck Perceptron Layer The message generation function generated by the iteration, Represents a collection of nodes, The set of edges connecting the nodes, G represents the graph representation, Representative Node i Node features ， Indicates k Layer The node message generation function generated by the iteration, express k The differentiable function that the layer network passes through, Representative k-1 The node coordinates of the layer, No. k The node coordinates for layer update.

[0110] Then apply p The geometric bottleneck perceptron of the layer gets updated vector and scalar representations as follows:

[0111]

[0112] A vector and scalar representing the updates of a p-layer geometric bottleneck perceptron with a network depth of k layers, represent Vector and scalar information of the geometric bottleneck perceptron in the previous layer.

[0113] The final output is through Function , express k The differentiable function that the layer network passes through, The final output is:

[0114]

[0115] S34, constructing interaction maps based on protein maps and small molecule ligand maps , H represents node features, E represents edge features, X represents node coordinate positions, subscript p represents proteins, and c represents small molecules. To represent the interaction graph of proteins and small molecules, we fully connect the protein and small molecule ligand nodes to consider protein-protein interactions, small molecule ligand-small molecule ligand interactions, and protein-small molecule ligand interactions.

[0116] S4, inputting the fully connected interaction graph containing the interaction information between small molecule ligand and protein into the binding conformation prediction model to predict the protein-small molecule ligand docking posture conformation;

[0117] The binding conformation prediction module used is a graph neural network improved based on self-attention.

[0118] In the binding pose generation process, the small molecule ligand conformation is initially generated by RDkit, represented as the coordinates of the small molecule ligand node, and randomly rotated and moved around the center of the pocket; the initial node embedding and edge features are initialized by graph normalization and multi-layer perceptron layer, as shown below:

[0119]

[0120] in, Represents graph normalization, H Represents node characteristics, c , p , s and v Representing compounds, proteins, scalars and vectors respectively, the node features of proteins and small molecule ligands updated by geometric bottleneck perceptrons and graph transformer networks are initialized by graph normalization, while the edge features of proteins and small molecule ligands updated by geometric bottleneck perceptrons and graph neural networks improved based on self-attention are initialized by linear layers; node embeddings, edge embeddings and small molecule ligand atomic positions are updated by applying graph transformer network blocks.

[0121] The self-attention-based improved graph neural network block consists of 8 self-attention-based improved graph neural network layers with self-attention, taking into account both protein-small molecule ligand interactions and small molecule ligand-small molecule ligand interactions. The message passing process of the lth layer of the self-attention-based improved graph neural network is as follows, enabling the self-attention-based improved graph neural network block to learn to robustly improve the binding pose:

[0122]

[0123] In the above formula, H and E denote the embedding of nodes and edges respectively, X represents the node location, p、c , r as well as l Represent protein, compound, cycle index and graph neural network layer index based on self-attention improvement, respectively. , , Represents protein-small molecule ligand Layer Improved graph neural network layer based on self-attention r After the cycle, the node information, edge information and node coordinates are updated. , , It is the node information, edge information and node coordinates of the previous layer. EGNN_Layer is the graph neural network layer information update method based on self-attention improvement.

[0124]

[0125] At the beginning of each cycle, the updated embedding and the original embedding will be well combined through the gating equation. The gating equation is as shown above. Gate_Block is the basic block of the EGNN layer and the residual connection used at the beginning of each recycling. F represents a general embedding, p、c as well as r Represent protein, compound, and circulation index, respectively. The updated insert for the next cycle, Represents the embedding information after eight self-attention updates in the previous cycle. Represents the initial embedding information.

[0126] The predicted protein-small molecule ligand docking posture conformation is post-processed in two ways: force field optimization and conformation alignment to ensure that the generated conformation is reasonable in terms of bond length and angle. Figure 2 .

[0127] By the attached Figure 2 The data shows that Figure 2 In the figure, (a) is the real structure of 4qac protein-small molecule docking in reality, (b) is the structure generated by the model-predicted 4qac protein-small molecule docking without post-processing, (c) is the predicted structure of 4qac protein-small molecule docking after force field post-processing to ensure reasonable bond length, and (d) is the predicted structure of 4qac protein-small molecule docking after conformational alignment post-processing to ensure reasonable angle. Taking the 4qac protein in the pdbbind core set as an example, the RMSD of the three predicted poses is less than 2Å, indicating a successful prediction. By comparing the real structure, the structure without post-processing, the structure after force field post-processing, and the structure after conformational alignment post-processing, the importance of post-processing steps in improving the predicted structure is clearly demonstrated. These post-processing steps help ensure that the generated conformations are reasonable in terms of bond length and angle, thereby improving the accuracy and reliability of the model prediction, and the processed protein-small molecule docking structure is closer to the real structure.

[0128] S5, input the protein-small molecule ligand docking posture conformation into a pre-trained mixed density neural network scoring function model to obtain the affinity score and possible binding posture between the small molecule ligand and the protein.

[0129] The scoring function module is composed of a mixed density neural network (MDN). Given the embedding and position information of the node, affinity estimation is performed through the MDN block. The probability distribution of the minimum distance between each protein and small molecule ligand node that MDN can learn can introduce distance inductive bias to the shared encoder:

[0130]

[0131]

[0132]

[0133]

[0134] In the formula, , , Represent the mean, standard deviation, and mixing coefficient respectively. , are the coding nodes of proteins and small molecule ligands. In the MDN module, the coding nodes of proteins and small molecule ligands are embedded respectively through the geometric bottleneck perceptron and the self-attention-based graph neural network. Sequential connection to obtain , and then passed through a linear layer, a batch normalization layer, an activation function (ELU), and a Dropout layer for protein-small molecule ligand interaction capture. Then, the mean, standard deviation, mixing coefficient, and node distance were determined using 3 linear layers. , , are the learnable parameters of the linear layer.

[0135] Thus, the score of the predicted docking strength is obtained:

[0136]

[0137] represents the score of predicted docking strength, represents a set of protein nodes, represents the small molecule ligand node set, and For the predicted atom and bond types, It represents the node distance. The formula calculates the distance between the protein node and each small molecule ligand node in a specific conformation, then calculates the log-likelihood value of each target ligand node pair, and sums up all possible values. This value is used to score the three-dimensional structure of the protein-small molecule ligand. The higher the score, the more likely it is to be a protein-small molecule ligand binding conformation.

[0138] Among them, the network output results after scoring network judgment and the standard answer together constitute the loss function. The loss function is the cross entropy loss function, which is used to judge the quality of the network output results until the loss function drops below the set threshold. It is considered that the learning and training of the protein encoder and small molecule ligand encoder are completed, and the predicted small molecule ligand binding pose is output.

[0139] See attached Figure 3 , is a schematic diagram of the results of the benchmark test of the method of the present invention on the pdbbind core set.

[0140] By the attached Figure 3 The data shows that Figure 3In the figure, (a) is the initial RMSD distribution test result of the pdbbind core set tested by the model after training, (b) is the RMSD distribution test result of the pdbbind core set tested by the model after training and post-processing with force field, and (c) is the RMSD distribution test result of the pdbbind core set tested by the model after training and post-processing with conformation alignment. In docking, when the RMSD is less than 2Å, it can be considered as a successful docking example. The results show that the success rate of molecular docking and the reduction of RMSD are optimized after force field post-processing, while there is no obvious optimization effect on the success rate of docking and the reduction of RMSD after alignment processing, but the molecules after the two treatments are more consistent with the bond length and angle of the real structure.

[0141] The advantages of the present invention are: it can generate binding postures and predict binding strengths quickly and accurately, characterize proteins based on residues, thereby encoding the geometric features of 3D protein structures, reducing computational costs, and especially using an improved geometric bottleneck perceptron as a protein encoder, which can simultaneously learn the scalar features and vector features of protein graph data structures, capture the three-dimensional structural features of proteins, enhance sharing by reducing the parameter space in the module, jointly integrate vector and scalar features to aggregate information flows, and update the scalar features of each node or edge by fusing the potential vector features on the vector representation path, thereby fully modeling the relationship and geometric information in the graph; it is combined with a scoring module for training, thereby helping to guide the learning of posture generation; two encoders are used, respectively for learning proteins and small molecule ligands, to achieve a combination of a fully connected graph and an improved graph neural network module based on self-attention, and to achieve rapid docking.

[0142] The above contents are only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A method for rapid and accurate protein-small molecule ligand docking based on deep learning, characterized in that: include: S1, obtain protein raw data and small molecule ligand raw data; S2, structurally representing the raw data of small molecule ligands and protein raw data to construct the small molecule ligand undirected composite graph and protein KNN graph respectively; The construction of the small molecule ligand undirected composite graph and protein KNN graph comprises: S21, the raw data of small molecule ligands are represented as undirected composite graphs using the torch_geometric package [G c =(H c ,E c ,X c )], with atoms as nodes and covalent bonds as edges, where H c is the node feature, E c is the edge feature, X c represents atomic coordinates; S22, the protein raw data uses protein residues as nodes and connects the first 30 nearest neighbor nodes to construct a KNN graph [G p =(H p ,E p ,X p )], nodes are represented by residues, H p is the protein node feature, E p For protein edge features, the carbon atom coordinates in the residue are used as node positions X p , each node or edge in the graph contains a scalar feature (H p_s ,E p_s ) and vector features (H p_v ,E p_v ), H p_s is the scalar feature of the protein node, E p_s is the scalar feature of the protein edge, H p_v is the vector feature of the protein node, E p_v is a vector feature of a protein edge, wherein the scalar feature includes a relative distance and a position distance, and the vector feature includes a unit vector of a direction, a side chain direction, and an edge direction; S3, inputting the small molecule ligand undirected composite graph into the small molecule ligand encoder to generate the updated embedding representation of the small molecule ligand; inputting the protein KNN graph into the protein encoder to generate the updated embedding representation of the protein, and then constructing a fully connected interaction graph containing the interaction information between the small molecule ligand and the protein; The construction of a fully connected interaction graph containing information on the interaction between small molecule ligands and proteins includes: S31, inputting the undirected composite graph data of the small molecule ligand into the small molecule ligand encoder to learn the atomic interactions inside the small molecule ligand, and generating an updated embedding representation of the small molecule ligand, wherein the embedding representation includes node features, edge features and node position information; S32, inputting protein KNN graph data into a protein encoder, learning the interactions within protein residues and the topological and geometric features between them, and generating an updated embedding representation of the protein, wherein the embedding representation includes information of neighbor nodes of the residue nodes and the topological and geometric features learned by the protein encoder; S33, using the updated embedding representation of the small molecule ligand obtained in S31 and the updated embedding representation of the protein obtained in S32, constructing a fully connected interaction graph containing the interaction information between the small molecule ligand and the protein; S4, inputting the fully connected interaction graph containing the interaction information between small molecule ligand and protein into the binding conformation prediction model to predict the protein-small molecule ligand docking posture conformation; S5, input the protein-small molecule ligand docking posture conformation into a pre-trained mixed density neural network scoring function model to obtain the affinity score and possible binding posture between the small molecule ligand and the protein.

2. The method for rapid and accurate protein-small molecule ligand docking based on deep learning according to claim 1, characterized in that: In S1, the protein raw data and small molecule ligand raw data are all protein-small molecule raw data after removing the core set from the general set of the PDBbind2020 database.

3. The method for rapid and accurate protein-small molecule ligand docking based on deep learning according to claim 1, characterized in that: In S4, the binding conformation prediction model is a graph neural network module, which is composed of 8 graph neural network layers improved based on self-attention. The prediction model predicts the protein-small molecule ligand docking posture conformation including: taking the updated embedding representation in the interaction graph containing the interaction information of small molecule ligand and protein as input, and predicting the protein-small molecule ligand docking posture through a multi-level self-attention mechanism; introducing a gating equation to combine the updated embedding representation with the original embedding representation, optimize the prediction results, and obtain the predicted protein-small molecule ligand docking posture conformation.

4. The method for rapid and accurate protein-small molecule ligand docking based on deep learning according to claim 1, characterized in that: Before performing the S5 operation, it also includes an operation of optimizing the protein-small molecule ligand docking posture conformation, including force field optimization or conformation alignment processing.

5. The method for rapid and accurate protein-small molecule ligand docking based on deep learning according to claim 1, characterized in that: In S5, the pre-trained mixed density neural network scoring function model includes: S51, inputting the protein-small molecule ligand docking posture conformation into a mixed density neural network scoring function model, wherein the protein-small molecule ligand docking posture conformation includes a node embedding representation of the protein and a node embedding of the small molecule ligand, as well as their position information; S52, uses a protein encoder and a self-attention-based graph neural network to process the node embedding representation of proteins and the node embedding of small molecule ligands respectively, captures the structural features of proteins and the chemical features of small molecule ligands, connects the processed feature vectors sequentially, and constructs a joint feature vector containing both protein and small molecule ligand information; S53, passing the joint feature vector constructed in S52 to a network layer of a graph neural network, wherein the network layer includes a linear layer, a batch normalization layer, an ELU activation function, and a Dropout layer, and the combined network layer extracts high-level features; S54, three linear layers are used to predict the mean, standard deviation and mixing coefficient respectively, which together define the probability distribution of the minimum distance between the protein and the small molecule ligand; another set of linear layers is used to directly predict the affinity score; S55, define the loss function, which includes two parts: the error between the probability distribution output by the mixed density neural network and the true distance, and the cross entropy loss between the affinity score and the standard answer; use the optimization algorithm to minimize the loss function, set early stopping until the loss value no longer changes, which means that the scoring part has fully learned the characteristics of protein-small molecules, and obtain a pre-trained mixed density neural network scoring function model.

6. A system for implementing the method for rapid and accurate protein-small molecule ligand docking based on deep learning as described in any one of claims 1 to 5, characterized in that: include: Data acquisition module, used to obtain protein raw data and small molecule ligand raw data; The structured characterization building module is used to perform structured characterization on the raw data of small molecule ligands and proteins, respectively, and to construct the small molecule ligand undirected composite graph and protein KNN graph; The interaction graph construction module is based on the small molecule ligand encoder and the protein encoder to obtain the updated embedding representation of the small molecule ligand and the updated embedding representation of the protein, and to construct a fully connected interaction graph containing the interaction information between the small molecule ligand and the protein; Combined with the conformation prediction module, it is used to predict the docking posture conformation of protein-small molecule ligand; Scoring function model module, which evaluates the affinity score and possible binding pose of the predicted docking pose conformation; The output and visualization module is used to output the predicted docking posture conformation, affinity score and possible binding posture, and provide visualization function.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method for rapid and accurate protein-small molecule ligand docking based on deep learning described in any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for rapid and accurate protein-small molecule ligand docking based on deep learning described in any one of claims 1 to 5 are implemented.

Images