Method and device for determining molecular binding conformation

Abstract

The present application provides a method and device for determining molecular binding conformation. It includes: obtaining a molecular graph corresponding to a chemical molecule and a residue graph corresponding to a protein; inputting the molecular graph and the residue graph into a molecular docking model, wherein the molecular docking model includes: a molecular graph encoding network layer, a residue graph encoding network layer, and a conformation prediction network layer; calling the molecular graph encoding network layer to process the molecular graph, updating the molecular node features of the molecular graph, and calling the residue graph encoding network layer to process the residue graph, updating the protein node features of the residue graph; calling an interaction construction module to perform graph node pairing and new edge processing on the updated residue graph and molecular graph to obtain an interaction graph; calling the conformation prediction network layer to process the interaction graph to obtain the molecular binding conformation between the chemical molecule and the protein. The present application can improve the prediction accuracy and prediction speed of binding conformation, as well as the screening capability and hit rate of virtual screening.

Description

Technical Field

[0001] The present application relates to the field of chemical technology, and in particular to a method and device for determining molecular binding conformation. Background Art

[0002] Molecular docking is a core technology in structure-based virtual screening. It can generate binding conformations between proteins and small molecules, predict binding strength based on these conformations, and select the optimal binding conformation. Traditional docking software (such as Autodock 4, AutoDock Vina, LeDock, rDock, Glide, GOLD, and PLANTS) typically employs search algorithms to search for a range of possible conformations of the small molecule ligand within a protein pocket and uses scoring functions to predict the binding strength of these conformations and select the optimal one. To cope with large-scale virtual screening, docking software often employs heuristic algorithms, such as genetic algorithms, to replace greedy algorithms for conformational search and improve efficiency. However, heuristic algorithms can result in the optimal conformation not being found. Similarly, to improve computational efficiency, the scoring function often uses simplified molecular force fields or energy terms based on expert experience. These energy terms are faster to calculate but less accurate. While simplified search algorithms and scoring functions improve the efficiency of molecular docking in virtual screening, they also lead to lower hit rates. Furthermore, the continued emergence of numerous methods to accelerate molecular docking (such as QuickVin, FWAVina, Vina GPU, and AutoDock GPU) suggests that simplified molecular docking models remain unsatisfactory in terms of efficiency when used for large-scale virtual screening. To address these challenges, deep learning techniques have been introduced to improve the accuracy and speed of predicting binding affinities and docking conformations between protein ligands.

[0003] From a deep learning perspective, the prediction of binding affinity (drug-target interaction, DTI) can be simply viewed as a regression or classification task. Therefore, existing deep learning algorithms for regression / classification can be quickly tested for their effectiveness. Furthermore, because the prediction accuracy of binding affinity / DTI models is often less accurate than that of docking conformations, significant effort has been invested in the development of deep learning-based DTI models. Convolutional neural networks (CNNs), having been widely applied and validated in the field of vision, have also been used in the development of DTI models. For example, KDEEP and RosENet voxelize protein-ligand complexes into a three-dimensional grid, where each grid point contains channels representing several atomic features (analogous to the RGB channels in an image). Classic CNN algorithms (such as ResNet101) are then used to accurately predict binding affinity. Compared to CNNs, graph neural networks (GNNs) offer higher computational efficiency and can naturally capture the topological connectivity between atoms. Consequently, GNNs are increasingly being used in the development of DTI models. Among GNN-based DTI models, IGN, SIGN, PIGNet, DeepDock, and RTMScore are representative high-accuracy models. IGN and SIGN construct interaction graphs based on protein and ligand nodes and directly predict protein-ligand binding affinities. PIGNet does not directly predict binding affinity, but instead predicts parameters of physics-inspired energy terms and uses the sum of these energy terms as the binding affinity. DeepDock and RTMScore employ mixture density networks (MDNs) to learn the optimal distance distribution between protein-ligand nodes in experimentally resolved crystal conformations. They then calculate probabilities for each pair of protein-ligand node distances in a given binding conformation and sum them to form a binding strength score. Although these models outperform traditional scoring functions in predicting binding affinity, their accuracy is highly dependent on the accuracy of the binding conformation. Therefore, a class of DTI models, also known as sequence-based models (SBMs), has been proposed that predict protein-ligand binding based on SMILES of protein sequences and ligands rather than the binding conformation itself. TransformerCPI, a representative model in SBM, uses the protein sequence and SMILES to assign a binding probability and a binding label. Because SBM models do not require binding conformations, they offer larger training sets, lower computational costs, and can be used when the protein-ligand binding conformation is unknown. However, due to the lack of structural information during training, SBM cannot rank multiple binding conformations of the same protein-ligand complex, making it unsuitable for binding conformation prediction.

[0004] AlphaFold2 can predict the folding conformation of proteins based on sequences with atomic-level errors, making a major breakthrough in the field of protein conformation prediction. Since then, more and more researchers have begun to study how to use deep learning technology to predict the natural conformation of molecules and the ligand-bound conformation of proteins. Unlike traditional molecular docking algorithms, using deep learning algorithms to predict protein-ligand-bound conformations is expected to achieve higher accuracy and faster speed. As the first model to use deep learning algorithms to predict binding conformations, EquiBind uses a GNN architecture, which updates node features and node coordinates simultaneously during each message passing process. EquiBind has achieved a speed increase of thousands of times compared to traditional docking algorithms, but its docking accuracy has also been greatly reduced. Similarly, E3Bind uses an E(n) equivariant graph neural network (EGNN), which updates the coordinates of molecular nodes based on the current environment during each message passing and achieves accuracy that is superior to traditional docking algorithms. Unlike the two aforementioned models, TankBind adds a conformation-based binding affinity prediction module to its conformational generation. Its binding conformation and affinity prediction accuracy surpasses traditional docking algorithms, yet its computational efficiency is comparable to that of traditional docking algorithms. The denoised diffusion probability model (DDPM) has achieved remarkable performance in image generation. Qian et al. combined DDPM with a GNN architecture to develop NeuralPLexer, which predicts protein-ligand binding conformations based on protein flexibility. Its docking accuracy is comparable to that of E3Bind and TankBind. Similarly, DiffDock mimics traditional docking algorithms, treating the molecule as a rigid body. DDPM uses non-Euclidean manifolds to predict its rotation, translation, and torsion within the pocket to generate the optimal binding conformation. DiffDock achieves docking accuracy far exceeding that of the aforementioned models, but because it requires multiple runs of the model during the generation process, its computational speed is even slower than that of some fast traditional docking algorithms. Although the aforementioned deep learning-based models have achieved significant improvements in docking speed and accuracy compared to traditional docking tools, these models were tested and applied in blind docking scenarios, that is, docking when the protein pocket is unknown. In actual virtual screening, protein pockets are often known or discovered in advance through experiments. Therefore, docking in a given pocket is a more common application scenario, and traditional docking algorithms are often designed for this scenario. LigPose is a model that generates and scores docking conformations based on a given pocket. Its architecture combines the self-attention mechanism of the Transformer architecture with the EGNN, and achieves docking accuracy and speed that surpass traditional docking algorithms. However, a drawback of the EGNN architecture is that it fails to consider the bond length and angle constraints of intramolecular covalent bonds when generating conformations, resulting in unreasonable bond lengths and angles in the predicted molecular conformations.Similar to LigPose, Jiang et al. proposed MedusaGraph, which also utilizes the EGNN architecture to generate docking conformations. Unlike LigPose, MedusaGraph uses the initial molecular conformations generated by the docking software rather than randomly sampled from a normal distribution. Therefore, MedusaGraph is more of a conformational optimization model than a conformational generation model. Summary of the Invention

[0005] The technical problem to be solved by the embodiments of the present application is to provide a method and device for determining a molecular binding conformation, so as to improve the prediction accuracy and prediction speed of the binding conformation and binding strength.

[0006] In a first aspect, the present invention provides a method for determining a molecular binding conformation, the method comprising:

[0007] Obtain molecular graphs corresponding to chemical molecules and residue graphs corresponding to proteins;

[0008] Inputting the molecular graph and the residue graph into a molecular docking model, wherein the molecular docking model comprises: a molecular graph encoding network layer, a residue graph encoding network layer, and a conformation prediction network layer;

[0009] calling the molecular graph encoding network layer to process the molecular graph and update the molecular node features of the molecular graph, and calling the residue graph encoding network layer to process the residue graph and update the protein node features of the residue graph;

[0010] Call the interaction building module to process the updated residue graph and molecular graph, add new edges and obtain the interaction graph;

[0011] The conformation prediction network layer is called to process the interaction graph to obtain the molecular binding conformation between the chemical molecule and the protein.

[0012] Optionally, the molecular docking model further comprises: a scoring network layer,

[0013] After calling the conformation prediction network layer to process the interaction graph to obtain the molecular binding conformation between the chemical molecule and the protein, the method further includes:

[0014] Calling the scoring network layer to score the molecular binding conformation to obtain a score value of the molecular binding conformation;

[0015] The binding affinity between the molecular binding conformation and the target protein is determined according to the score value.

[0016] Optionally, the molecular graph encoding network layer includes: a linear layer and a multi-head attention graph neural network layer,

[0017] The calling of the molecular graph encoding network layer to process the molecular graph to obtain molecular node features of the molecular graph includes:

[0018] Calling the linear layer to perform dimension unification processing on the node features and edge features in the molecular graph to obtain node features and edge features with unified dimensions;

[0019] The multi-head attention graph neural network layer is called to perform feature processing on the node features and edge features of the unified dimension to obtain the molecular node features of the molecular graph.

[0020] Optionally, calling the residue graph encoding network layer to process the residue graph to obtain protein node features of the residue graph includes:

[0021] calling the residue graph encoding network layer to perform splicing processing on the encoding features of the sequence information of the protein and the residue node features in the residue graph to obtain residue splicing features;

[0022] The residue graph encoding network layer is called to process the residue splicing features to obtain protein node features of the residue graph.

[0023] Optionally, calling the conformation prediction network layer to process the interaction graph to obtain a predicted molecular binding conformation between the chemical molecule and the protein includes:

[0024] Normalizing the node vector representations in the interaction graph to obtain normalized node vector representations, and performing dimension conversion on the edge vector representations in the interaction graph to obtain dimension converted vector representations;

[0025] The conformation prediction network layer is called to process the normalized node vector representation and the dimension conversion vector representation to obtain a molecular binding conformation between the chemical molecule and the protein.

[0026] Optionally, calling the scoring network layer to score the molecular binding conformation to obtain the score value of the molecular binding conformation includes:

[0027] Calling the scoring network layer to perform node feature splicing processing on the protein nodes and the molecule nodes in the molecular binding conformation to generate node splicing features;

[0028] Processing the node splicing features to obtain a fusion feature representation;

[0029] The fusion feature representation is processed to obtain a score value of the molecular binding conformation.

[0030] Optionally, after calling the conformation prediction network layer to process the interaction graph to obtain the molecular binding conformation between the chemical molecule and the protein, the method further includes:

[0031] The molecular binding conformation is post-processed to obtain a processed molecular binding conformation.

[0032] Optionally, before inputting the molecular graph and the residue graph into a molecular docking model, the method further comprises:

[0033] The scoring network layer is trained using real molecular binding conformations;

[0034] When the scoring network layer converges, the training samples are called to train the conformation prediction network layer until the conformation prediction network layer converges to obtain the molecular docking model.

[0035] In a second aspect, an embodiment of the present application provides a device for determining a molecular binding conformation, the device comprising:

[0036] Molecular residue map acquisition module, used to obtain molecular maps corresponding to chemical molecules and residue maps corresponding to proteins;

[0037] a molecular residue graph input module, for inputting the molecular graph and the residue graph into a molecular docking model, wherein the molecular docking model comprises: a molecular graph encoding network layer, a residue graph encoding network layer, an interaction graph construction and conformation prediction network layer;

[0038] a node feature updating module, configured to call the molecular graph encoding network layer to process the molecular graph and update the molecular node features of the molecular graph, and to call the residue graph encoding network layer to process the residue graph and update the protein node features of the residue graph;

[0039] The interaction graph acquisition module calls the interaction construction module to process the updated residue graph and molecular graph, and add new edges to obtain the interaction graph;

[0040] The molecular binding conformation acquisition module is used to call the conformation prediction network layer to process the interaction graph to obtain the molecular binding conformation between the chemical molecule and the protein.

[0041] Optionally, the molecular docking model further comprises: a scoring network layer,

[0042] The device further comprises:

[0043] A conformation score value acquisition module is used to call the scoring network layer to score the molecular binding conformation to obtain the score value of the molecular binding conformation;

[0044] The binding affinity determination module is used to determine the binding affinity between the molecular binding conformation and the target protein according to the score value.

[0045] Optionally, the molecular graph encoding network layer includes: a linear layer and a multi-head attention graph neural network layer,

[0046] The node feature acquisition module includes:

[0047] a node feature acquisition unit, configured to call the linear layer to perform dimension unification processing on the node features and edge features in the molecular graph to obtain dimensionally unified node features and edge features;

[0048] The molecular node feature acquisition unit is used to call the multi-head attention graph neural network layer to perform feature processing on the node features and edge features of the unified dimension to obtain the molecular node features of the molecular graph.

[0049] Optionally, the node feature acquisition module includes:

[0050] a residue splicing feature acquisition unit, configured to call the residue graph encoding network layer to perform splicing processing on the encoding features of the protein sequence information and the residue node features in the residue graph to obtain residue splicing features;

[0051] The protein node feature acquisition unit is used to call the residue graph encoding network layer to process the residue splicing features to obtain the protein node features of the residue graph.

[0052] Optionally, the molecular binding conformation acquisition module includes:

[0053] a dimension conversion vector representation acquisition unit, configured to perform a normalization process on the node vector representations in the interaction graph to obtain a normalized node vector representation, and perform a dimension conversion process on the edge vector representations in the interaction graph to obtain a dimension conversion vector representation;

[0054] The predicted molecular binding conformation acquisition unit is used to call the conformation prediction network layer to process the normalized node vector representation and the dimension conversion vector representation to obtain the molecular binding conformation between the chemical molecule and the protein.

[0055] Optionally, the conformation score value acquisition module includes:

[0056] a node splicing feature generating unit, configured to call the scoring network layer to perform node feature splicing processing on the protein nodes and the molecule nodes in the molecular binding conformation to generate a node splicing feature;

[0057] A fusion feature representation acquisition unit, configured to process the node splicing features to obtain a fusion feature representation;

[0058] The conformation score value acquisition unit is used to process the fusion feature representation to obtain the score value of the molecular binding conformation.

[0059] Optionally, the device further comprises:

[0060] The post-processing module is used to perform post-processing on the molecular binding conformation to obtain a processed molecular binding conformation.

[0061] Optionally, the device further comprises:

[0062] Scoring network layer training module, used to train the scoring network layer using real molecular binding conformations;

[0063] When the scoring network layer converges, the training samples are called to train the conformation prediction network layer until the conformation prediction network layer converges to obtain the molecular docking model.

[0064] In a third aspect, an embodiment of the present application provides an electronic device, including:

[0065] A processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, any one of the above-described methods for determining a molecular binding conformation is implemented.

[0066] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which, when the instructions in the storage medium are executed by a processor of an electronic device, enables the electronic device to execute any of the molecular binding conformation determination methods described above.

[0067] Compared with the prior art, the embodiments of the present application have the following advantages:

[0068] In an embodiment of the present application, a molecular graph corresponding to a chemical molecule and a residue graph corresponding to a protein are obtained, and the molecular graph and the residue graph are input into a molecular docking model. The molecular docking model includes: a molecular graph encoding network layer, a residue graph encoding network layer, and a conformation prediction network layer. The molecular graph encoding network layer is called to process the molecular graph, and the molecular node features of the molecular graph are updated. The residue graph encoding network layer is called to process the residue graph, and the protein node features of the residue graph are updated. The interaction construction module is called to perform graph node pairing and new edge processing on the updated residue graph and molecular graph to obtain an interaction graph. The conformation prediction network layer is called to process the interaction graph to obtain the molecular binding conformation between the chemical molecule and the protein. The embodiment of the present application can improve the prediction accuracy and prediction speed of the binding conformation.

[0069] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 A flowchart of the steps of a method for determining a molecular binding conformation provided in an embodiment of the present application;

[0071] Figure 2 A schematic diagram of a model reasoning process provided in an embodiment of the present application;

[0072] Figure 3 A schematic diagram of the rationality of KarmaDock-generated conformations provided in an embodiment of the present application;

[0073] Figure 4 A schematic diagram illustrating the screening capability of KarmaDock on DEKOIS 2.0 provided in an embodiment of the present application;

[0074] Figure 5 A schematic diagram of the results of KarmaDock in virtual screening for LTK provided in an embodiment of the present application;

[0075] Figure 6 A schematic structural diagram of a molecular binding conformation determination device provided in an embodiment of the present application;

[0076] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0077] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0078] The terms used in the embodiments of the present application are for the purpose of describing specific embodiments only and are not intended to limit the present application. The singular forms "a", "an", "the" and "the" used in the embodiments of the present application and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise.

[0079] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal. In the absence of further restrictions, an element defined by the sentence "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or terminal comprising the element.

[0080] Reference Figure 1 , shows a flow chart of the steps of a method for determining a molecular binding conformation provided in an embodiment of the present application, such as Figure 1 As shown, the method for determining the molecular binding conformation may include the following steps:

[0081] Step 101: Obtain a molecular graph corresponding to a chemical molecule and a residue graph corresponding to a protein.

[0082] In an embodiment of the present application, a molecular graph corresponding to a chemical molecule can be obtained. Specifically, the chemical molecule can be represented as a two-dimensional molecular graph. The molecular graph can use atoms as nodes and covalent bonds as edges. The node features and edge features can be as shown in Table 1 below:

[0083] Table 1:

[0084]

[0085]

[0086] In Table 1 above, bool2value refers to converting a Boolean variable into a number, where True corresponds to 1 and False corresponds to 0.

[0087] In practical applications, molecular graph representation can be implemented using open source Python libraries such as rdkit or torchdrug, and the number of more or fewer node features and edge features represented has little impact.

[0088] For proteins, we can use amino acid residues as nodes and the Cα coordinates within the residues as node coordinates. We can use the K-nearest neighbor algorithm to connect the 30 nearest neighbor nodes to construct a residue graph. In addition to conventional two-dimensional information, node representation and edge representation also contain three-dimensional geometric information, as shown in Table 2 below:

[0089] Table 2:

[0090]

[0091]

[0092] After obtaining the molecular graph corresponding to the chemical molecule and the residue graph corresponding to the protein, step 102 is performed.

[0093] Step 102: Input the molecular graph and the residue graph into a molecular docking model, wherein the molecular docking model includes: a molecular graph encoding network layer, a residue graph encoding network layer, and a conformation prediction network layer.

[0094] After obtaining the molecular graph corresponding to the chemical molecule and the residue graph corresponding to the protein, the molecular graph and the residue graph can be input into the molecular docking model for processing.

[0095] In this example, the molecular docking model may include: a molecular graph encoding network layer, a residue graph encoding network layer, and a conformation prediction network layer. The molecular graph encoding network layer may be used to process the molecular graph, and the residue graph encoding network layer may be used to process the residue graph.

[0096] After the molecular graph and the residue graph are input into the molecular docking model, step 103 is performed.

[0097] Step 103: calling the molecular graph encoding network layer to process the molecular graph and update the molecular node features of the molecular graph, and calling the residue graph encoding network layer to process the residue graph and update the protein node features of the residue graph.

[0098] After the molecular graph and residue graph are input into the molecular docking model, the molecular graph encoding network layer can be called to process the molecular graph and update the molecular node features of the molecular graph. The residue graph encoding network layer can also be called to process the residue graph and update the protein node features of the residue graph.

[0099] In the specific implementation, the molecular graph and the residue graph can be input into two different encoders (i.e., the molecular graph encoding network layer and the residue graph encoding network layer), respectively, as Figure 2 The Graph Transformer (GT) and Geometric Vector Perceptrons (GVP) shown in the figure are used to learn the topological connectivity and intramolecular interactions within the graph and update the node representation.

[0100] The structure of GT is as follows Figure 2As shown in Section C, two linear layers are first used to unify the node and edge features into 128 dimensions. These features are then fed into a multi-head attention graph transformer layer. This implementation process is described in detail in conjunction with the following specific implementation methods.

[0101] In a specific implementation of the present application, step 103 may include:

[0102] Sub-step A1: calling the linear layer to perform dimension unification processing on the node features and edge features in the molecular graph to obtain dimension-unified node features and edge features.

[0103] In this embodiment, the molecular graph encoding network layer may include: a linear layer and a multi-head attention graph neural network layer.

[0104] After the molecular graph is input into the molecular docking model, the linear layer can be called to unify the dimensions of the node features and edge features in the molecular graph to obtain the node features and edge features with unified dimensions. The implementation process can be shown as the following formula (1):

[0105]

[0106] After calling the linear layer to perform dimension unification processing on the node features and edge features in the molecular graph to obtain dimensionally unified node features and edge features, sub-step A2 is executed.

[0107] Sub-step A2: Call the multi-head attention graph neural network layer to perform feature processing on the node features and edge features of the unified dimension to obtain the molecular node features of the molecular graph.

[0108] After calling the linear layer to perform dimension-unified processing on the node features and edge features in the molecular graph to obtain dimension-unified node features and edge features, the multi-head attention graph neural network layer can be called to perform feature processing on the dimension-unified node features and edge features to obtain the molecular node features of the molecular graph.

[0109] In this example, node features and edge features are input into the multi-head attention graph neural network layer (graphtransformer layer) for processing. The message passing formula of the lth layer is shown in the following formula (2-10):

[0110]

[0111]

[0112]

[0113]

[0114]

[0115]

[0116]

[0117]

[0118]

[0119] In the above formula, is the learnable parameter in the linear layer, k∈1,...,H represents the number of attention heads of the self-attention mechanism; d k is the feature dimension under each attention head; j∈N(i) is the set of neighbor nodes j of node i; Norm represents batch normalization; Concat represents feature concatenation; Dropout represents the dropout layer; SiLU is an activation function; Aggregation_Sum j∈N(i) Add the information of all neighbor nodes of i to update the node information of i; andSoftmax j∈N(i) Indicates that the SoftMax operation is performed on neighbor node j.

[0120] In the specific implementation, the GT in the encoder can be replaced by many different small molecule representation extraction models such as SchNet, SphereNet, GemNet and other geometric equivariant graph neural networks.

[0121] The network architecture of GVP can be Figure 2 As shown in part D of FIG, the gvp layer can receive scalar features and vector features and perform message transmission to update node features. The implementation process can be described in detail in conjunction with the following specific implementation methods.

[0122] In another specific implementation of the present application, the above step 103 may include:

[0123] Sub-step B1: calling the residue graph encoding network layer to perform splicing processing on the encoding features of the protein sequence information and the residue node features in the residue graph to obtain residue splicing features.

[0124] In this embodiment, the residue graph encoding network layer may be called to perform splicing processing on the encoding features of the protein sequence information and the residue node features in the residue graph to obtain the residue splicing features.

[0125] Sub-step B2: Processing the residue splicing features to obtain protein node features of the residue graph.

[0126] After obtaining the residue splicing features, the residue splicing features can be processed to obtain protein node features of the residue graph.

[0127] In this example, the message delivery update process can be shown as follows:

[0128]

[0129]

[0130]

[0131]

[0132]

[0133]

[0134]

[0135]

[0136] In the above formula, are the learnable parameters in the linear layer; and is the returned variable; σ s and σ v Denotes the activation function. Before entering the GVP layer, the protein sequence information will be encoded by the embedding layer and then combined with the protein residue node features, as shown in the following formulas (19) and (20):

[0137] h seq =Embedding(sequence) (19)

[0138] h s =Concat(h s0 , h seq ) (20)

[0139] Subsequently, the residue node features and edge features are initialized using formulas (21-22):

[0140] (h s1 , h v1 )=gvp(LayerNorm(h s , h v )) (twenty one)

[0141] (e s1 , e v1 )=gvp(LayerNorm(e s , e v )) (twenty two)

[0142] After initialization, the node features and edge features are fed into the repeated gvpconv module for message passing, as shown in the following formula (23-32):

[0143]

[0144]

[0145]

[0146]

[0147]

[0148]

[0149]

[0150]

[0151]

[0152]

[0153] In the above formulas (15), (16), and (20), the activation functions of the GVP layer for scalars and vectors are ReLU and Sigmoid respectively; the GVP layer in other formulas does not use an activation function; and and and Aggregation_Mean i∈N(j) It means that the messages on the edge formed by the set of neighbors i of node j are aggregated by taking the average.

[0154] After the molecular graph encoding network layer is called to process the molecular graph to obtain the molecular node features of the molecular graph, and the residue graph encoding network layer is called to process the residue graph to obtain the protein node features of the residue graph, step 104 is executed.

[0155] Step 104: Call the interaction building module to perform graph node pairing and edge addition processing on the updated residue graph and molecular graph to obtain an interaction graph.

[0156] After obtaining the molecular node features and protein node features, the interaction graph construction module can be called to perform graph node pairing and edge addition processing on the updated residue graph and molecular graph to obtain the interaction graph. In the specific implementation, the protein residue nodes and molecular atom nodes can be placed in one graph. In addition to the original edges, all residue graph node-molecule graph node pairs are traversed and edges are added. Similarly, all nodes within the molecular graph form edges in pairs. The molecular graph is divided into several fragments according to the rotatable bonds of the molecule. Each fragment can be regarded as a rigid body during the docking process, and the distance between its nodes is unchanged. The characteristics of the edge are the type of edge (single bond, double bond, triple bond, aromatic bond and non-bonded interaction) and the distance between nodes. The distance between protein nodes and the distance between nodes within the molecular graph are calculated based on the real coordinates of the nodes. The distance between protein nodes and molecular nodes and the distance between nodes of different fragments in the molecular graph are encoded as -1.

[0157] After calling the interaction graph construction module to perform graph node pairing and newly added edge processing on the updated residue graph and molecule graph, step 105 is executed.

[0158] Step 105: calling the conformation prediction network layer to process the interaction graph to obtain the molecular binding conformation between the chemical molecule and the protein.

[0159] After invoking the interaction graph construction module to perform node pairing and edge addition processing on the updated residue and molecular graphs, the conformation prediction network layer can be invoked to process the interaction graph to obtain the molecular binding conformation between the chemical molecule and the protein. This implementation process is described in detail in conjunction with the following specific implementation method.

[0160] In a specific implementation of the present application, the above step 105 may include:

[0161] Sub-step C1: normalizing the node vector representations in the interaction graph to obtain normalized node vector representations, and performing dimension conversion on the edge vector representations in the interaction graph to obtain dimension conversion vector representations.

[0162] In this embodiment, the node vector representations in the interaction graph may be normalized to obtain normalized node vector representations, and the edge vector representations in the interaction graph may be dimensionally transformed to obtain dimensionally transformed vector representations.

[0163] After obtaining the dimension-converted vector representation, sub-step C2 is performed.

[0164] Sub-step C2: calling the conformation prediction network layer to process the normalized node vector representation and the dimension conversion vector representation to obtain the molecular binding conformation between the chemical molecule and the protein.

[0165] After obtaining the dimension-converted vector representation, the conformation prediction network layer can be called to process the normalized node vector representation and the dimension-converted vector representation to obtain the molecular binding conformation between the chemical molecule and the protein.

[0166] like Figure 2 As shown in Section F, the conformation prediction network layer uses an 8-layer stacked E(n)equivariant graph neural network (EGNN layer) to predict the stable binding conformation formed between the protein and the ligand. The node representation of the interaction graph is first graph normalized, and the edge representation is mapped to 128 dimensions through a linear layer, as shown in the following formulas (33) and (34):

[0167] h1=GraphNorm(h0) (33)

[0168] e1=W e_init e0+b e_init (34)

[0169] Then the EGN layer is input for message passing and coordinate prediction. The message passing process of the lth layer can be shown as follows:

[0170]

[0171]

[0172]

[0173]

[0174]

[0175]

[0176]

[0177]

[0178]

[0179]

[0180]

[0181]

[0182] In the above formula, and and and is a learnable parameter in the linear layer; k∈1,...,H represents the number of attention heads in the self-attention mechanism; d k The dimension of each attention head; j∈N(i) is the set of neighbor nodes j of node i; Concat represents the concatenation operation; Dropout represents the dropout layer; LeakyReLU is an activation function; Aggregation_Sum j∈N(i) Aggregate the messages on the edge set consisting of node i and its neighbor node j by adding; and Softmax j∈N(i) Indicates that the SoftMax operation is performed on the neighbor node j; Coords_Update_Block is the module responsible for predicting coordinate changes, as shown in the following formulas (47) to (52):

[0183]

[0184]

[0185]

[0186]

[0187]

[0188]

[0189] In this embodiment, after obtaining the molecular binding conformation between the chemical molecule and the protein, the molecular binding conformation can also be post-processed to ensure the rationality of the bond lengths and bond angles of the molecular binding conformation. This implementation process can be described in detail in conjunction with the following specific implementation methods.

[0190] In a specific implementation of the present application, after the above step 105, the following steps may also be included:

[0191] Step D1: post-processing the molecular binding conformation to obtain a processed molecular binding conformation.

[0192] After the molecular binding conformation is generated, it can be post-processed to ensure the rationality of the bond lengths and angles of the conformation. There are three post-processing methods: (1) Force field optimization (molecule): Use the MMFF94 force field of RDkit to perform 10 steps of minimization on the optimized conformation to optimize the conformation; (2) Force field optimization (protein-ligand complex): Use the OPLS2005 or OPLS3 force field to perform energy optimization on the protein-ligand complex, in which the protein conformation is fixed and only the molecular conformation is changed; (3) RDkit conformation superposition: Calculate the torsion angles of the rotatable bonds of the molecular binding conformation predicted by the model, assign the same torsion angles to the reasonable conformation generated by RDkit, and then superimpose the predicted point cloud conformation and the RDkit conformation.

[0193] In this embodiment, after generating the molecular binding conformation, the molecular binding conformation can also be scored, and the binding affinity between the generated molecular binding conformation and the target protein can be determined based on the score. This implementation process can be described in detail in conjunction with the following specific implementation methods.

[0194] In a specific implementation of the present application, after the above step 105, the following steps may also be included:

[0195] Step E1: calling the scoring network layer to score the molecular binding conformation to obtain a score value of the molecular binding conformation.

[0196] In this embodiment, the molecular docking model may further include: a scoring network layer, such as Figure 1 As shown in part E.

[0197] After obtaining the molecular binding conformation, the scoring network layer can be invoked to score the molecular binding conformation and obtain a score for the molecular binding conformation. Specifically, node feature concatenation can be performed on the protein nodes and molecule nodes in the molecular binding conformation to generate a node concatenation feature. The node concatenation feature is then processed to obtain a score for the molecular binding conformation. This implementation process is described in detail in conjunction with the specific implementation described below.

[0198] In another specific implementation of the present application, the above step E1 may include:

[0199] Sub-step F1: calling the scoring network layer to perform node feature splicing processing on the protein nodes and the molecule nodes in the molecular binding conformation to generate node splicing features.

[0200] After obtaining the molecular binding conformation, the scoring network layer can be called to perform node feature splicing processing on the protein nodes and molecular nodes in the molecular binding conformation to generate node splicing features.

[0201] After calling the scoring network layer to perform node feature splicing processing on the protein nodes and the molecule nodes in the molecular binding conformation and generating the node splicing features, sub-step F2 is executed.

[0202] Sub-step F2: Process the node splicing features to obtain a fusion feature representation.

[0203] After calling the scoring network layer to perform node feature splicing processing on the protein nodes and the molecule nodes in the molecular binding conformation and generating the node splicing features, the node splicing features can be processed to obtain a fusion feature representation.

[0204] After processing the node splicing features to obtain the fused feature representation, sub-step F3 is executed.

[0205] Sub-step F3: Processing the fusion feature representation to obtain a score value of the molecular binding conformation.

[0206] After processing the node splicing features to obtain a fused feature representation, the fused feature representation can be processed to obtain a score value of the molecular binding conformation.

[0207] The scoring network layer can be Figure 2 As shown in part E, the scoring network layer can first combine protein nodes and molecule nodes in pairs, merge the node features of the protein-molecule node pairs, and obtain the fusion feature representation between the protein-molecule node pairs through the linear layer, batch normalization, ELU and dropout. Then, the mean μ of the distribution of the minimum distance between protein ligand nodes is predicted through three different (linear layer + ELU) methods. p，c , standard deviation σ p，c and the combination coefficient π p，c The specific formula is as follows:

[0208] h p，c =Dropout(ELU(BatchNorm(W p，c Concat(h p , h c )+b p，c ))) (55)

[0209] μ p，c =ELU(W μ h p，c +b μ )+1 (56)

[0210] σ p，c =ELU(W σ h p，c +b σ )+1.1 (57)

[0211] π p，c =Softmax(W π h p，c +b π ) (58)

[0212] In the above formula, W μ , W σ , b μ , b σ and b π ∈R n is a learnable parameter in the linear layer; h p and Concat, Softmax, and Dropout represent concatenation, softmax, and dropout operations respectively.

[0213] After calling the scoring network layer to score the molecular binding conformation to obtain the score value of the molecular binding conformation, step E2 is executed.

[0214] Step E2: Determine the binding affinity between the molecular binding conformation and the target protein according to the score value.

[0215] After the scoring network layer is called to score the molecular binding conformation and obtain a score value for the molecular binding conformation, the binding affinity between the molecular binding conformation and the target protein (i.e., the protein in step 101) can be determined based on the score value. In virtual drug screening, the model is used to calculate the binding conformations and binding strengths of different molecules and proteins, and they are ranked according to the binding strength scores. Based on the ranking, the top N compounds are selected for biological experimental verification (N is an integer greater than zero), thereby discovering potential drugs.

[0216] In a specific implementation, in the molecular docking model provided in the embodiments of this application, the coordinate prediction module can be replaced with a similar diffusion architecture from a step-by-step prediction EGNN module, but the GNN layer with self-attention integration is retained. During training, random noise is added to the correct binding conformation each time, and the model is asked to predict the noise. During testing, the binding conformation of the complex is randomly sampled from a Gaussian distribution, and the model is asked to gradually predict the noise and remove the noise to restore the correct binding conformation.

[0217] The model training process can be described in detail in combination with the following specific implementation methods.

[0218] In another specific implementation of the present application, before step 101, the following steps may be further included:

[0219] Step G1: Use real molecular binding conformations to train the scoring network layer.

[0220] In this embodiment, when performing model training, the scoring network layer can be trained using real molecular binding conformations.

[0221] Step G2: When the scoring network layer converges, calling training samples to train the conformation prediction network layer until the conformation prediction network layer converges to obtain the molecular docking model.

[0222] After the scoring network layer converges, the training samples can be called to train the conformation prediction network layer until the conformation prediction network layer converges to obtain a molecular docking model.

[0223] In the specific implementation, when training the model, the MDN module (scoring module) is trained first, and after the MDN module converges, the EGNN module (ensemble conformation prediction module) is trained. Because the MDN module learns the optimal distribution of distances between protein residues and ligand atom pairs, and the encoder is shared by the MDN module and the EGNN module, when learning the MDN module, an inductive bias such as the optimal distribution of residue-atom distances can be introduced to the encoder module, which helps the model reduce the search space for binding conformations, thereby helping the model learn how to quickly and accurately find stable binding conformations of proteins and ligands. In actual applications, the EGNN module is first used to predict the binding conformation of the protein and the small molecule ligand, and then input into the MDN module to predict the binding strength.

[0224] The advantages of the molecular docking model provided in the embodiments of the present application are described below in combination with several aspects:

[0225] 1. Combining Conformation Prediction Accuracy and Speed

[0226] The examples of this application compare the accuracy and speed of conformation prediction by KarmaDock (i.e., molecular docking model) with those of traditional docking software and a deep learning-based binding conformation prediction model. As shown in Table 3, KarmaDock achieved the second highest docking accuracy and the fastest docking speed on the refined set; and achieved the highest accuracy and the fastest speed on the remaining test sets.

[0227] Table 3:

[0228]

[0229]

[0230] a) Speed: The number of seconds it takes to predict a single binding conformation (LigPose only discloses the acceleration rate, not the specific time value). b) Accuracy: The number of predicted conformations that are successfully predicted (RMSD between the predicted conformation and the true conformation is less than or equal to ) conformation ratio; cDL: represents a deep learning model. d) DK: represents a traditional docking software. e) LigPose: The best-performing model for the current test set is displayed in bold. f) KarmaDock is trained on the PDBBind general excluding refined set dataset. g) KarmaDock Raw is trained on the PDBBind general set. h) KarmaDock FF and iKarmaDock Aligned add post-processing to KarmaDock Raw. j) TankBind: This example assigns the correct TankBind pocket.

[0231] 2. Prediction Accuracy of Bond Strength

[0232] The examples of this application compare the docking and screening capabilities of KarmaDock and traditional docking software / deep learning scoring functions on the CASF dataset. The docking capability is to see whether the model can select the optimal binding conformation from multiple binding conformations of the same protein-ligand complex, and the screening capability is to see whether the model can enrich the active molecules according to the scores. These two capabilities have a great impact on the hit rate of virtual screening. Therefore, this example focuses on these two capabilities of the model. As shown in Table 4, KarmaDock achieved a good result of second place in both scoring and screening capabilities, indicating that KarmaDock can well predict the binding strength between proteins and ligands.

[0233] Table 4:

[0234]

[0235]

[0236]

[0237] Success rate is an indicator used to measure the docking ability of the model, the number of successfully predicted conformations The ratio of the number of tested conformations to the total number of conformations. EF 1% (enrichment factor) is a metric used to measure the screening power of a model, measuring the ratio of the proportion of active molecules in the top 1% of scoring molecules to the proportion of active molecules in the entire compound library. DL represents a deep learning model. HB represents a hybrid model combining deep learning and traditional docking. DK represents traditional docking software.

[0238] 3. Rationality of KarmaDock-generated conformations

[0239] The embodiment of the present application uses the maximum / average error of the bond length / bond angle of the conformation to define whether a conformation is reasonable or not. Figure 3 As shown, the rationality of the conformation directly generated by KarmaDock is already very high (90%>), but the rationality of the conformation generated by traditional docking software is 100%. Therefore, the embodiment of the present application adds three post-processing methods, one is to optimize the molecular conformation a limited number of times through the MMFF94 force field, the second is to optimize the conformation of the protein-ligand complex through the OPLS series force field (fix the protein conformation and only optimize the molecular conformation) until convergence, and the third is to use the reasonable conformation generated by RDkit and the directly generated conformation to perform torsion angle superposition and spatial orientation superposition. The three post-processing methods proposed in the embodiment of the present application can greatly improve the rationality of the conformation (especially the OPLS force field optimization and RDkit conformation superposition can generate 100% conformations).

[0240] 4. Testing the model’s performance as docking software on the virtual screening dataset DEKOIS 2.0

[0241] In virtual screening, the accuracy of the docking conformation and the accuracy of the binding affinity have a crucial impact on the hit rate. Without an accurate binding conformation, even with a high-precision affinity prediction model, it is impossible to accurately predict the binding affinity; without a high-precision binding affinity prediction model, without an accurate binding conformation, it is impossible to give the true binding affinity. In the embodiment of this application, KarmaDock was used to perform a virtual screening test on DEKOIS2.0. As shown in the attached figure, Figure 4 The experimental results show that KarmaDock can achieve the best enrichment ability among a number of traditional docking models and deep learning models.

[0242] 5. Virtual Screening and Experimental Validation of CLIP1-LTK

[0243] Testing on classic datasets doesn't fully reflect KarmaDock's performance in real-world applications. Therefore, KarmaDock was used to perform virtual screening against the Clip1-LTK target on Specs (208,780 compounds) and ChemDiv (1,561,007 compounds). The testing was performed on a 48-core CPU (Intel(R) Xeon(R) Gold 6240R CPU @ 2.40GHz) and a single GPU (Tesla V100S-PCIE-32GB). Map generation took 5,489 seconds (approximately 1.5 hours) on the 48-core CPU, and docking and scoring took 24,862 seconds (approximately 6.9 hours) on the single GPU, for a total of approximately 8.4 hours. By comparison, the fastest traditional docking software, LeDock, takes approximately 1,091 hours (45 days) with the same computing resources. Therefore, KarmaDock significantly outperforms traditional docking software (by at least 130 times). In addition, biological experiments have verified that 22 of the compounds with the highest KarmaDock scores are active against the LTK target, with the highest activity reaching IC50=765nm, demonstrating the practicality of KarmaDock in actual virtual screening. It also proves that KarmaDock has greatly improved the accuracy and speed of binding conformation and strength prediction. Figure 5 As shown in Part A, the binding conformation and interaction pattern of the active compound lorlatinib and the ALK target crystallized from the experiment are given, which have been experimentally verified (Appendix Figure 5 Lorlatinib also has an inhibitory effect on the LTK target, which is highly homologous to ALK. The binding conformation of lorlatinib and LTK obtained by KarmaDock (see Figure 2). Figure 5 (shown in part B) and the crystal conformation (see Appendix Figure 5 The results are similar to those shown in part A of Figure 2. Since the pockets are not exactly the same, the RMSD cannot be directly compared. However, from the perspective of the interaction mode, the conformation docked by KarmaDock almost restores the interaction mode between lorlatinib and the target in the crystal conformation, further verifying the accuracy of KarmaDock docking. In addition, this example also visualizes the IC50 found by KarmaDock virtual screening to be 765.6nm (see Appendix). Figure 5 The interaction mode between the molecule (shown in part E) and LTK (see Appendix Figure 5 As shown in part C of the figure, it can be seen that the binding conformation docked by KarmaDock also forms a hydrogen bond with the 593rd amino acid MET.

[0244] The method for determining the molecular binding conformation provided in the embodiment of the present application obtains the molecular graph corresponding to the chemical molecule and the residue graph corresponding to the protein, and inputs the molecular graph and the residue graph into the molecular docking model. The molecular docking model includes: a molecular graph encoding network layer, a residue graph encoding network layer and a conformation prediction network layer. The molecular graph encoding network layer is called to process the molecular graph and update the molecular node features of the molecular graph, and the residue graph encoding network layer is called to process the residue graph and update the protein node features of the residue graph. The interaction construction module is called to perform graph node pairing and new edge processing on the updated residue graph and molecular graph to obtain an interaction graph, and the conformation prediction network layer is called to process the interaction graph to obtain the molecular binding conformation between the chemical molecule and the protein. The embodiment of the present application can improve the prediction accuracy and prediction speed of the binding conformation.

[0245] Reference Figure 6 , shows a schematic structural diagram of a molecular binding conformation determination device provided in an embodiment of the present application, such as Figure 6 As shown, the molecular binding conformation determination device 600 may include the following modules:

[0246] A molecular residue graph acquisition module 610 is used to acquire a molecular graph corresponding to a chemical molecule and a residue graph corresponding to a protein;

[0247] a molecular residue graph input module 620 for inputting the molecular graph and the residue graph into a molecular docking model, wherein the molecular docking model comprises: a molecular graph encoding network layer, a residue graph encoding network layer, an interaction graph construction and conformation prediction network layer;

[0248] a node feature updating module 630 for invoking the molecular graph encoding network layer to process the molecular graph and update the molecular node features of the molecular graph, and invoking the residue graph encoding network layer to process the residue graph and update the protein node features of the residue graph;

[0249] The interaction graph acquisition module 640 calls the interaction construction module to perform graph node pairing and edge addition processing on the updated residue graph and molecule graph to obtain an interaction graph;

[0250] The molecular binding conformation acquisition module 650 is used to call the conformation prediction network layer to process the interaction graph to obtain the molecular binding conformation between the chemical molecule and the protein.

[0251] Optionally, the molecular docking model further comprises: a scoring network layer,

[0252] The device further comprises:

[0253] A conformation score value acquisition module is used to call the scoring network layer to score the molecular binding conformation to obtain the score value of the molecular binding conformation;

[0254] The binding affinity determination module is used to determine the binding affinity between the molecular binding conformation and the target protein according to the score value.

[0255] Optionally, the molecular graph encoding network layer includes: a linear layer and a multi-head attention graph neural network layer,

[0256] The node feature acquisition module includes:

[0257] a node feature acquisition unit, configured to call the linear layer to perform dimension unification processing on the node features and edge features in the molecular graph to obtain dimensionally unified node features and edge features;

[0258] The molecular node feature acquisition unit is used to call the multi-head attention graph neural network layer to perform feature processing on the node features and edge features of the unified dimension to obtain the molecular node features of the molecular graph.

[0259] Optionally, the node feature acquisition module includes:

[0260] a residue splicing feature acquisition unit, configured to call the residue graph encoding network layer to perform splicing processing on the encoding features of the protein sequence information and the residue node features in the residue graph to obtain residue splicing features;

[0261] The protein node feature acquisition unit is used to call the residue graph encoding network layer to process the residue splicing features to obtain the protein node features of the residue graph.

[0262] Optionally, the molecular binding conformation acquisition module includes:

[0263] a dimension conversion vector representation acquisition unit, configured to perform a normalization process on the node vector representations in the interaction graph to obtain a normalized node vector representation, and perform a dimension conversion process on the edge vector representations in the interaction graph to obtain a dimension conversion vector representation;

[0264] The predicted molecular binding conformation acquisition unit is used to call the conformation prediction network layer to process the normalized node vector representation and the dimension conversion vector representation to obtain the molecular binding conformation between the chemical molecule and the protein.

[0265] Optionally, the conformation score value acquisition module includes:

[0266] a node splicing feature generating unit, configured to call the scoring network layer to perform node feature splicing processing on the protein nodes and the molecule nodes in the molecular binding conformation to generate a node splicing feature;

[0267] A fusion feature representation acquisition unit, configured to process the node splicing features to obtain a fusion feature representation;

[0268] The conformation score value acquisition unit is used to process the fusion feature representation to obtain the score value of the molecular binding conformation.

[0269] Optionally, the device further comprises:

[0270] The post-processing module is used to perform post-processing on the molecular binding conformation to obtain a processed molecular binding conformation.

[0271] Optionally, the device further comprises:

[0272] Scoring network layer training module, used to train the scoring network layer using real molecular binding conformations;

[0273] When the scoring network layer converges, the training samples are called to train the conformation prediction network layer until the conformation prediction network layer converges to obtain the molecular docking model.

[0274] The molecular binding conformation determination device provided in the embodiment of the present application obtains the molecular graph corresponding to the chemical molecule and the residue graph corresponding to the protein, and inputs the molecular graph and the residue graph into the molecular docking model. The molecular docking model includes: a molecular graph encoding network layer, a residue graph encoding network layer and a conformation prediction network layer. The molecular graph encoding network layer is called to process the molecular graph and update the molecular node features of the molecular graph, and the residue graph encoding network layer is called to process the residue graph and update the protein node features of the residue graph. The interaction construction module is called to perform graph node pairing and new edge processing on the updated residue graph and molecular graph to obtain an interaction graph, and the conformation prediction network layer is called to process the interaction graph to obtain the molecular binding conformation between the chemical molecule and the protein. The embodiment of the present application can improve the prediction accuracy and prediction speed of the binding conformation.

[0275] An embodiment of the present application provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the above-mentioned molecular binding conformation determination method when executed by the processor.

[0276] Figure 7 FIG. 7 is a schematic diagram showing the structure of an electronic device 700 according to an embodiment of the present invention. Figure 7As shown, electronic device 700 includes a central processing unit (CPU) 701, which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) 702 or computer program instructions loaded from a storage unit 708 into a random access memory (RAM) 703. In RAM 703, various programs and data required for the operation of electronic device 700 can also be stored. CPU 701, ROM 702 and RAM 703 are connected to each other via a bus 704. An input / output (I / O) interface 705 is also connected to bus 704.

[0277] Multiple components in the electronic device 700 are connected to the I / O interface 705, including an input unit 706, such as a keyboard, a mouse, a microphone, etc.; an output unit 707, such as various types of displays, speakers, etc.; a storage unit 708, such as a magnetic disk, an optical disk, etc.; and a communication unit 709, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 709 allows the electronic device 700 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0278] The various processes and procedures described above may be executed by the processing unit 701. For example, the method of any of the above embodiments may be implemented as a computer software program, which is tangibly contained in a computer-readable medium, such as the storage unit 708. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 700 via the ROM 702 and / or the communication unit 709. When the computer program is loaded into the RAM 703 and executed by the CPU 701, one or more actions in the method described above may be performed.

[0279] An embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which implements the above-mentioned method for determining the molecular binding conformation when executed by a processor.

[0280] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0281] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, devices, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0282] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, terminals (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal generate instructions for implementing the steps in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0283] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing terminal to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0284] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal so that a series of operational steps are executed on the computer or other programmable terminal to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable terminal for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0285] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they become aware of the basic inventive concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present invention.

[0286] The above is a detailed introduction to a molecular binding conformation determination method, a molecular binding conformation determination device, an electronic device and a computer-readable storage medium provided by the present application. Specific examples are used herein to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, based on the ideas of the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A method for determining a molecular binding conformation, characterized in that: The method comprises: Obtain molecular graphs corresponding to chemical molecules and residue graphs corresponding to proteins; Inputting the molecular graph and the residue graph into a molecular docking model, wherein the molecular docking model comprises: a molecular graph encoding network layer, a residue graph encoding network layer, and a conformation prediction network layer; calling the molecular graph encoding network layer to process the molecular graph and update the molecular node features of the molecular graph, and calling the residue graph encoding network layer to process the residue graph and update the protein node features of the residue graph; Call the interaction building module to process the updated residue graph and molecular graph, add new edges and obtain the interaction graph; calling the conformation prediction network layer to process the interaction graph to obtain a molecular binding conformation between the chemical molecule and the protein; The molecular graph encoding network layer includes: a linear layer and a multi-head attention graph neural network layer. The calling of the molecular graph encoding network layer to process the molecular graph to obtain the molecular node features of the molecular graph includes: calling the linear layer to perform dimension-unified processing on the node features and edge features in the molecular graph to obtain dimension-unified node features and edge features; calling the multi-head attention graph neural network layer to perform feature processing on the dimension-unified node features and edge features to obtain the molecular node features of the molecular graph; The calling of the residue graph encoding network layer to process the residue graph to obtain protein node features of the residue graph includes: calling the residue graph encoding network layer to perform splicing processing on the encoding features of the protein sequence information and the residue node features in the residue graph to obtain residue splicing features; calling the residue graph encoding network layer to process the residue splicing features to obtain protein node features of the residue graph; The calling of the conformation prediction network layer to process the interaction graph to obtain the predicted molecular binding conformation between the chemical molecule and the protein includes: normalizing the node vector representation in the interaction graph to obtain a normalized node vector representation, and performing dimension conversion on the edge vector representation in the interaction graph to obtain a dimension conversion vector representation; calling the conformation prediction network layer to process the normalized node vector representation and the dimension conversion vector representation to obtain the molecular binding conformation between the chemical molecule and the protein.

2. The method according to claim 1, characterized in that The molecular docking model further includes: a scoring network layer, After calling the conformation prediction network layer to process the interaction graph to obtain the molecular binding conformation between the chemical molecule and the protein, the method further includes: Calling the scoring network layer to score the molecular binding conformation to obtain a score value of the molecular binding conformation; The binding affinity between the molecular binding conformation and the target protein is determined according to the score value.

3. The method according to claim 2, characterized in that The calling of the scoring network layer to score the molecular binding conformation to obtain the score value of the molecular binding conformation includes: Calling the scoring network layer to perform node feature splicing processing on the protein nodes and the molecule nodes in the molecular binding conformation to generate node splicing features; Processing the node splicing features to obtain a fusion feature representation; The fusion feature representation is processed to obtain a score value of the molecular binding conformation.

4. The method according to claim 1, wherein After calling the conformation prediction network layer to process the interaction graph to obtain the molecular binding conformation between the chemical molecule and the protein, the method further includes: The molecular binding conformation is post-processed to obtain a processed molecular binding conformation.

5. The method according to claim 2, characterized in that Before inputting the molecular graph and the residue graph into a molecular docking model, the method further comprises: The scoring network layer is trained using real molecular binding conformations; When the scoring network layer converges, the training samples are called to train the conformation prediction network layer until the conformation prediction network layer converges to obtain the molecular docking model.

6. A molecular binding conformation determination device, characterized in that: The device comprises: Molecular residue map acquisition module, used to obtain molecular maps corresponding to chemical molecules and residue maps corresponding to proteins; a molecular residue graph input module, for inputting the molecular graph and the residue graph into a molecular docking model, wherein the molecular docking model comprises: a molecular graph encoding network layer, a residue graph encoding network layer, an interaction graph construction and conformation prediction network layer; a node feature updating module, configured to call the molecular graph encoding network layer to process the molecular graph and update the molecular node features of the molecular graph, and to call the residue graph encoding network layer to process the residue graph and update the protein node features of the residue graph; The interaction graph acquisition module calls the interaction construction module to process the updated residue graph and molecular graph, and add new edges to obtain the interaction graph; a molecular binding conformation acquisition module, configured to call the conformation prediction network layer to process the interaction graph and obtain a molecular binding conformation between the chemical molecule and the protein; The molecular graph encoding network layer includes: a linear layer and a multi-head attention graph neural network layer, and the node feature acquisition module includes: a node feature acquisition unit for calling the linear layer to perform dimension-unified processing on the node features and edge features in the molecular graph to obtain dimension-unified node features and edge features; a molecular node feature acquisition unit for calling the multi-head attention graph neural network layer to perform feature processing on the dimension-unified node features and edge features to obtain molecular node features of the molecular graph; The node feature acquisition module includes: a residue splicing feature acquisition unit, configured to call the residue graph encoding network layer to perform splicing processing on the encoding features of the protein sequence information and the residue node features in the residue graph to obtain residue splicing features; and a protein node feature acquisition unit, configured to call the residue graph encoding network layer to process the residue splicing features to obtain protein node features of the residue graph; The molecular binding conformation acquisition module includes: a dimension conversion vector representation acquisition unit, which is used to normalize the node vector representation in the interaction graph to obtain a normalized node vector representation, and to perform dimension conversion on the edge vector representation in the interaction graph to obtain a dimension conversion vector representation; and a predicted molecular binding conformation acquisition unit, which is used to call the conformation prediction network layer to process the normalized node vector representation and the dimension conversion vector representation to obtain the molecular binding conformation between the chemical molecule and the protein.

7. The device according to claim 6, characterized in that The molecular docking model further comprises: a scoring network layer, The device further comprises: A conformation score value acquisition module is used to call the scoring network layer to score the molecular binding conformation to obtain the score value of the molecular binding conformation; The binding affinity determination module is used to determine the binding affinity between the molecular binding conformation and the target protein according to the score value.

Images