A method and device for training an in-pocket molecule generation model

Abstract

The application provides a pocket molecule generation model training method and device. The method comprises: obtaining training data, wherein the training data comprises a protein pocket topology file of a protein molecule; inputting the protein pocket topology file into a to-be-trained pocket molecule generation model, wherein the to-be-trained pocket molecule generation model comprises an interaction learning network layer and a generation network layer; calling the interaction learning network layer to process pocket surface features in the protein pocket topology file to obtain interaction features corresponding to pocket surface features of the protein molecule; calling the generation network layer to process the interaction features and the pocket surface features to obtain predicted structure features corresponding to the pocket surface; calculating a loss value of the to-be-trained pocket molecule generation model based on the predicted structure features; and in a case where the loss value is within a preset range, training the to-be-trained pocket molecule generation model to obtain a pocket molecule generation model. The application can achieve the purpose of generating active molecules and conformations corresponding to a protein target structure designated for drug design.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a training method and apparatus for a pocket-based molecular generation model. Background Technology

[0002] De novo molecular design is a crucial step in rational drug design. Based on the principles of molecular generation, it can be divided into ligand-based methods and structure-based methods. However, most current molecular generation models learn a chemical latent space and then sample new two-dimensional molecules, falling under the ligand-based category. These methods cannot explicitly capture the interactions between small molecules and protein pocket structures, thus preventing the models from designing drugs specifically targeting a particular protein. Furthermore, these models suffer from overfitting and poor generalization.

[0003] Currently, novel drug design methods based on deep learning frameworks can be divided into two categories according to the underlying logic of the model: the first category is novel drug design methods based on ligands; the second category is novel drug design methods based on pockets.

[0004] However, the main problems with ligand-based molecular generation are: 1) neglecting the molecule's stereoconformation and binding mode within the protein pocket; 2) lacking underlying physics support, the model faces overfitting, low generalization ability, and the generated molecules may be reasonable but not specific to any particular target. The main problems with existing structure-based molecular generation are: 1) the model cannot effectively capture the interaction patterns between the molecule and the protein pocket, resulting in unreasonable molecular structures and affinities that cannot exceed those of in-situ ligands; 2) mainstream methods can only predict the two-dimensional structure of the molecule and cannot generate the corresponding binding mode between the molecule and the pocket. Summary of the Invention

[0005] The technical problem to be solved by the embodiments of this application is to provide a training method and apparatus for an in-pocket molecular generation model, so as to achieve the purpose of generating corresponding active molecules and their conformations from a specified protein target structure for drug design.

[0006] In a first aspect, embodiments of this application provide a training method for an in-pocket molecule generation model, the method comprising:

[0007] Acquire training data, which includes: protein pocket topology files of protein molecules;

[0008] The protein pocket topology file is input into the in-pocket molecule generation model to be trained, which includes an interaction learning network layer and a generation network layer.

[0009] The interaction learning network layer is invoked to process the pocket surface features in the protein pocket topology file to obtain the interaction features corresponding to the pocket surface of the protein molecule;

[0010] The generative network layer is invoked to process the interaction features and the pocket surface features to obtain the predicted structural features corresponding to the pocket surface;

[0011] Based on the predicted structural features, the loss value of the pocket-in-molecule generation model to be trained is calculated;

[0012] If the loss value is within a preset range, the trained in-pocket molecular generation model is used as the in-pocket molecular generation model.

[0013] Optionally, the pocket surface features include at least one of the following: surface charge features, hydrogen bonding potential features, hydrophobic interaction features, Gaussian curvature features, mean curvature features, and shape index features.

[0014] Optionally, the interaction learning network layer includes: a first GNN network layer and a second GNN network layer.

[0015] The step of calling the interaction learning network layer to process the pocket surface features in the protein pocket topology file to obtain the interaction features corresponding to the pocket surface of the protein molecule includes:

[0016] The first GNN network layer is invoked to process the pocket surface features to obtain the topological interaction features corresponding to the pocket surface;

[0017] The second GNN network layer is invoked to process the pocket surface features and the topological interaction features to obtain the geometric interaction features corresponding to the pocket surface;

[0018] The topological interaction features and the geometric interaction features are used as the interaction features.

[0019] Optionally, the generated network layer includes: a topology generation module and a geometry generation module.

[0020] The step of calling the generative network layer to process the interaction features and the pocket surface features to obtain the predicted structural features corresponding to the pocket surface includes:

[0021] The topology generation module is invoked to process the interaction features and the pocket surface features to obtain the predicted topological structure features corresponding to the pocket surface;

[0022] The geometry generation module is invoked to process the interaction features, the pocket surface features, and the predicted topological structure features to obtain the predicted geometric structure features corresponding to the pocket surface;

[0023] The predicted topological features and the predicted geometric features are used as the predicted structural features.

[0024] Optionally, the predicted topological features include: atom type features and bond relationship features, and the predicted geometric features include: atom position features.

[0025] Optionally, after using the trained in-pocket molecular generation model as the in-pocket molecular generation model, the method further includes:

[0026] Obtain the target protein pocket topology file corresponding to the protein target to be predicted;

[0027] The target protein pocket topology file is input into the pocket-in-molecule generation model;

[0028] The interaction learning network layer is invoked to process the target pocket surface features in the target protein pocket topology file to obtain the target interaction features corresponding to the pocket surface of the target protein to be predicted.

[0029] The generator network layer is invoked to process the target interaction features and the target pocket surface features to obtain the target predicted structure features corresponding to the target pocket surface;

[0030] Based on the predicted structural features of the target, the predicted protein molecule corresponding to the protein target to be predicted is determined.

[0031] Secondly, embodiments of this application provide a training device for an in-pocket molecular generation model, the device comprising:

[0032] The training data acquisition module is used to acquire training data, which includes: protein pocket topology files of protein molecules;

[0033] The topology file input module is used to input the protein pocket topology file into the in-pocket molecule generation model to be trained. The in-pocket molecule generation model to be trained includes: an interaction learning network layer and a generation network layer.

[0034] The interaction feature acquisition module is used to call the interaction learning network layer to process the pocket surface features in the protein pocket topology file to obtain the interaction features corresponding to the pocket surface of the protein molecule.

[0035] The predicted structural feature acquisition module is used to call the generator network layer to process the interaction features and the pocket surface features to obtain the predicted structural features corresponding to the pocket surface;

[0036] The loss value calculation module is used to calculate the loss value of the in-pocket molecule generation model to be trained based on the predicted structural features.

[0037] The molecular generation model acquisition module is used to use the trained in-pocket molecular generation model as the in-pocket molecular generation model when the loss value is within a preset range.

[0038] Optionally, the pocket surface features include at least one of the following: surface charge features, hydrogen bonding potential features, hydrophobic interaction features, Gaussian curvature features, mean curvature features, and shape index features.

[0039] Optionally, the interaction learning network layer includes: a first GNN network layer and a second GNN network layer.

[0040] The interaction feature acquisition module includes:

[0041] The topology feature acquisition unit is used to call the first GNN network layer to process the pocket surface features and obtain the topology interaction features corresponding to the pocket surface;

[0042] The geometric feature acquisition unit is used to call the second GNN network layer to process the pocket surface features and the topological interaction features to obtain the geometric interaction features corresponding to the pocket surface;

[0043] An interaction feature acquisition unit is used to acquire the topological interaction features and the geometric interaction features as the interaction features.

[0044] Optionally, the generated network layer includes: a topology generation module and a geometry generation module.

[0045] The predicted structural feature acquisition module includes:

[0046] The predicted topology feature acquisition unit is used to call the topology generation module to process the interaction features and the pocket surface features to obtain the predicted topology structure features corresponding to the pocket surface;

[0047] The predicted geometric feature acquisition unit is used to call the geometry generation module to process the interaction features, the pocket surface features and the predicted topological structure features to obtain the predicted geometric structure features corresponding to the pocket surface;

[0048] A predictive structural feature acquisition unit is used to obtain the predicted topological structural features and the predicted geometric structural features as the predicted structural features.

[0049] Optionally, the predicted topological features include: atom type features and bond relationship features, and the predicted geometric features include: atom position features.

[0050] Optionally, the device further includes:

[0051] The target file acquisition module is used to acquire the target protein pocket topology file corresponding to the protein target to be predicted.

[0052] The target file input module is used to input the target protein pocket topology file into the pocket-in-molecule generation model;

[0053] The target interaction feature acquisition module is used to call the interaction learning network layer to process the target pocket surface features in the target protein pocket topology file to obtain the target interaction features corresponding to the pocket surface of the target protein to be predicted.

[0054] The target prediction structure feature acquisition module is used to call the generator network layer to process the target interaction features and the target pocket surface features to obtain the target prediction structure features corresponding to the target pocket surface;

[0055] The predicted protein molecule determination module is used to determine the predicted protein molecule corresponding to the target protein based on the predicted structural features of the target.

[0056] Thirdly, embodiments of this application provide an electronic device, including:

[0057] A processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the training method for the pocket-based molecular generation model described in any of the preceding claims.

[0058] Fourthly, embodiments of this application provide a computer-readable storage medium that, when the instructions in the storage medium are executed by a processor of an electronic device, enables the electronic device to perform the training method for the pocket-based molecular generation model described in any of the preceding claims.

[0059] Compared with the prior art, the embodiments of this application have the following advantages:

[0060] In this embodiment, training data is acquired, including a protein pocket topology file of a protein molecule. The protein pocket topology file is input into a cell-in-pocket molecule generation model to be trained. This model includes an interaction learning network layer and a generation network layer. The interaction learning network layer processes the pocket surface features in the protein pocket topology file to obtain the interaction features corresponding to the pocket surface of the protein molecule. The generation network layer processes the interaction features and pocket surface features to obtain the predicted structural features corresponding to the pocket surface. Based on the predicted structural features, the loss value of the cell-in-pocket molecule generation model to be trained is calculated. If the loss value is within a preset range, the trained cell-in-pocket molecule generation model is used as the cell-in-pocket molecule generation model. This embodiment learns the topological interactions on the surface of the protein pocket and the spatial interactions between ligand atoms and the pocket, and uses these interaction features to generate 3D molecules within a specified pocket through a geometry and topology generation module. This allows for the generation of corresponding active molecules and their conformations from a specified protein target structure for drug design during use.

[0061] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0062] Figure 1 A flowchart illustrating the steps of a training method for an in-pocket molecular generation model provided in an embodiment of this application;

[0063] Figure 2 A schematic diagram of a model processing flow provided in an embodiment of this application;

[0064] Figure 3 A schematic diagram of a molecular binding energy distribution provided for an embodiment of this application;

[0065] Figure 4 A schematic diagram of molecular distribution within a pocket provided for an embodiment of this application;

[0066] Figure 5 A schematic diagram of the structure of a training device for an in-pocket molecular generation model provided in an embodiment of this application;

[0067] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0068] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0069] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0070] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes said element.

[0071] Reference Figure 1 The diagram illustrates a flowchart of the training method for an in-pocket molecular generation model provided in an embodiment of this application. Figure 1 As shown, the training method for this pocket-based molecular generation model may include the following steps:

[0072] Step 101: Obtain training data, which includes: protein pocket topology files of protein molecules.

[0073] The embodiments of this application can be applied to scenarios where in-pocket molecular generation models are being trained.

[0074] Training data refers to the data used to train the in-pocket molecule generation model. In this example, training data may include: protein pocket topology files of protein molecules.

[0075] In practical implementation, protein pockets refer to cavities on or inside a protein that are suitable for binding with ligands. The amino acid residues surrounding the pocket determine its shape, location, physicochemical properties, and function.

[0076] In this example, the training data can be protein pocket topology files of protein molecules obtained using open-source software (such as MSMS, pymesh, etc.) to serve as training data for the in-pocket molecule generation model.

[0077] After obtaining the training data, proceed to step 102.

[0078] Step 102: Input the protein pocket topology file into the in-pocket molecule generation model to be trained. The in-pocket molecule generation model to be trained includes an interaction learning network layer and a generation network layer.

[0079] After obtaining the training data, the protein pocket topology file contained in the training data can be input into the in-pocket molecular generation model to be trained. This in-pocket molecular generation model can include an interaction learning network layer and a generation network layer.

[0080] After inputting the protein pocket topology file into the in-pocket molecular generation model to be trained, proceed to step 103.

[0081] Step 103: Call the interaction learning network layer to process the pocket surface features in the protein pocket topology file to obtain the interaction features corresponding to the pocket surface of the protein molecule.

[0082] After inputting the protein pocket topology file into the intracavitary molecule generation model to be trained, the interaction learning network layer can be called to process the pocket surface features in the protein pocket topology file to obtain the interaction features corresponding to the pocket surface of the protein molecule.

[0083] In practical applications, topological surfaces (i.e., pocket surfaces) (Multiscale Modeling) are the molecular solvent-accessible surface areas covered by triangular faces. Each node carries the physical properties corresponding to the surface, such as charge, hydrogen bond potential energy, and curvature at that point.

[0084] Solvent-accessible surface area: The surface trajectory of a target molecule obtained by rolling a spherical probe. It includes both the contact surface and the recessed surface. (The probe diameter is typically...) The area in direct contact is the contact surface.

[0085] In this embodiment, the pocket surface features may include at least one of the following: surface charge features, hydrogen bonding potential features, hydrophobic interaction features, Gaussian curvature features, mean curvature features, and shape index features. Figure 2 As shown, the original features of the pocket surface are: I represents Gaussian and mean curvature, II represents shape index, III represents surface charge of the biological system in solution, IV represents hydrogen bonding potential, and V represents nonpolar interaction.

[0086] In practical implementation, the nodes of a protein channel include three geometric features and three energy features. The geometric features may include Gaussian curvature, mean curvature, and shape index; the energy features may include Poisson-Boltzmann charge, hydrogen bond donor / acceptor, and hydrophilicity / hydrophobicity, specifically defined as follows:

[0087] Gaussian curvature: It can measure the inherent curvature of a surface. No non-stretching transformation of a surface will change its Gaussian curvature; it is an intrinsic invariant. The formula for calculating the Gaussian curvature of a surface is shown in the following formula (1):

[0088]

[0089] In the above formula (1), Let g be the covariant derivative and g be the metric. Specifically, for the representation of the triangular facet, the calculation formula can be shown in the following formula (2):

[0090]

[0091] Where j is a first-order neighbor of i, θ j The definition is as follows Figure 2 As shown in FI, A i It is the area of ​​the mixed voronoi region.

[0092] Mean curvature: Defined as the average of the principal curvatures. Unlike Gaussian curvature, mean curvature is an external measure of surface curvature, measuring the degree of curvature of the surface in space. As shown in the following formula (3):

[0093]

[0094] Where X is a family of smooth embedded hypersurfaces. g is the unit normal vector. ij For measurement standards.

[0095] For protein triangles, the calculation can be simplified to the following formula (4):

[0096]

[0097] In the above formula (4), j is a first-order neighbor of i, α ij and β ij like Figure 2 As shown in FI, p j p i These are the Cartesian coordinates of the compute node and its neighboring nodes, respectively.

[0098] Shape index: a curvature describing the topology of a local surface, proposed by Koenderink and van Doom, is a single-valued and angled measure. Its definition can be given by the following formula (5):

[0099]

[0100] In the above formula (5), K1 and K2 are the principal curvatures of the surface at a certain point, and the calculation formula is shown in the following formula (6):

[0101]

[0102] The Poisson-Boltzmann equation describes the continuous charge distribution of biological macromolecules in solution, and the specific formula is shown in formula (7) below:

[0103]

[0104] In the above formula (7), φ is the system potential, and ε is the dielectric constant of the solution. and z i The bulk concentration and charge of the i-th ion are given by β = 1 / K. B T, K B It is the Boltzmann constant. The PB equation is obtained by solving PDB2PQR and APBS.

[0105] Hydrogen bond donors and acceptors: Hydrogen bonds are essentially dipole-dipole interactions, usually represented by XH…Y. X is covalently (or ionicly) bonded to hydrogen and has high electronegativity, while Y is generally an atom containing a lone pair of electrons. Hydrogen bond interactions are widely used in the construction of force fields. Methods similar to masif and hydrogen potential can be used to calculate the hydrogen bond potential energy at surface nodes.

[0106] Hydrophobic interactions are weak, non-covalent interactions between nonpolar groups and are widely present in biological systems. Similar to hydropathic protocols, hydrophilic-hydrophobic interactions at surface nodes can be calculated.

[0107] In this embodiment, the interaction learning network layer may include two GNN network layers (i.e., a first GNN network layer and a second GNN network layer). The first GNN network layer learns the topological interactions of the pocket surface, and the second GNN network layer learns the geometric interactions of the pocket surface. In a specific implementation, the first GNN network layer can be a Geodesic-GNN, and the second GNN network layer can be a Geoattn-GNN. The processing of pocket surface features by the interaction learning network layer can be described in detail below with reference to the specific implementation method.

[0108] In one specific implementation of this application, step 103 may include:

[0109] Sub-step S1: Call the first GNN network layer to process the pocket surface features to obtain the topological interaction features corresponding to the pocket surface.

[0110] In this embodiment, after the training data is input into the pocket molecule generation model to be trained, the first GNN network layer can be called to process the pocket surface features to obtain the topological interaction features corresponding to the pocket surface.

[0111] After calling the first GNN network layer to process the pocket surface features and obtain the topological interaction features corresponding to the pocket surface, step S2 is executed.

[0112] Sub-step S2: Call the second GNN network layer to process the pocket surface features and the topological interaction features to obtain the geometric interaction features corresponding to the pocket surface.

[0113] After processing the pocket surface features using the first GNN network layer to obtain the topological interaction features corresponding to the pocket surface, the second GNN network layer can be called to process the pocket surface features and topological interaction features to obtain the geometric interaction features corresponding to the pocket surface.

[0114] After obtaining the geometric interaction features corresponding to the pocket surface, sub-step S3 is executed.

[0115] Sub-step S3: Use the topological interaction feature and the geometric interaction feature as the interaction feature.

[0116] After obtaining the geometric interaction features corresponding to the pocket surface, the topological interaction features and geometric interaction features can be used as the interaction features corresponding to the pocket surface of the protein molecule.

[0117] The above implementation process can be combined with Figure 2As shown: The original input to this pocket-based molecular generation model is a protein pocket topology file, which consists of nodes and triangular faces. The physical and geometric properties of the pocket surface are calculated and then attached to each node. These features include: (physical) surface charge, hydrogen bonding potential, and hydrophobic interactions; (geometric) Gaussian curvature, mean curvature, and shape index.

[0118] Interaction learning network layers can include Geodesic-GNN and Geoattn-GNN, and the molecular generation process is as follows: Figure 2 As shown in C, the topological interactions of the pocket surface can first be learned in the Geodesic-GNN module, and then the geometric interactions between the generated atoms and the topological surface can be learned in the Geoattn-GNN module.

[0119] In practical implementation, since protein pockets are abstracted as topological surface structures surrounded by triangular faces, directly using existing 3D graph neural representation frameworks to learn the underlying interactions presents some problems. For example... Figure 2 As shown in G, when atom i aggregates information from its surroundings, the receptive field of atom i will include, for example, information from its surroundings. Figure 2 The protein pockets shown are the "back sides", and small molecules do not directly match these back sides. Therefore, although existing geometric graph neural frameworks can better aggregate spatial information, they cannot fully consider the topological features of the pocket surface. Based on this, this embodiment designs Geodesic-GNN to aggregate information from the pocket surface. In Geodesic-GNN, the receptive field of atom i does not contain back side information. Compared with the traditional Euclidean geometric graph neural framework, Geodesic-GNN can perform more reasonable information aggregation on the topological surface. For the interaction between ligand atoms and protein pockets, spatial geometric relationships should be of greater concern. Based on this, this embodiment designs an interaction module based on geometric attention mechanism to learn the matching rules between atoms and pockets. Thus, the SurfGen representation learning framework can be constructed. First, edge features and atom features can be mapped to a higher-dimensional representation space. As shown in the following formula (8):

[0120]

[0121] Let GVP be the scalar and vector features of the i-th atom in the (l-1)ln2 iteration. Then, for each point on the topological surface, topological features can be learned between it and its neighbors, so that each point represents a surface patch.

[0122]

[0123] After aggregating the topological surface information, geometric attention convolution can be performed between the ligand atoms and the topological pocket, enabling the model to learn spatial interactions.

[0124]

[0125] This embodiment can use the learned interaction features to guide the model in molecular generation:

[0126] Geodesic-GNN:

[0127]

[0128]

[0129]

[0130]

[0131] Interaction Block for Surface and Ligand(Geometry):

[0132]

[0133]

[0134]

[0135]

[0136] A′ ij =sigmoid(A ij )

[0137]

[0138]

[0139]

[0140] After obtaining the topological and geometric interaction features corresponding to the pocket surface through the interaction learning network layer, step 104 is executed.

[0141] Step 104: Call the generator network layer to process the interaction features and the pocket surface features to obtain the predicted structural features corresponding to the pocket surface.

[0142] After obtaining the topological and geometric interaction features corresponding to the pocket surface through the interaction learning network layer, the generative network layer can be called to process the interaction features and pocket surface features to obtain the predicted structural features corresponding to the pocket surface.

[0143] In this example, the predicted structural features may include topological structural features and geometric structural features. The generative network layer may include a topology generation module and a geometry generation module. The topology generation module generates the predicted topological structural features corresponding to the pocket surface, and the geometry generation module generates the predicted geometric structural features corresponding to the pocket surface. This implementation process can be described in detail below with reference to the specific implementation method.

[0144] In one specific implementation of this application, step 104 may include:

[0145] Sub-step M1: Call the topology generation module to process the interaction features and the pocket surface features to obtain the predicted topological structure features corresponding to the pocket surface.

[0146] In this embodiment, after obtaining the interaction features (i.e., geometric interaction features and topological interaction features) through the interaction learning network layer, the topology generation module can be called to process the interaction features and pocket surface features to obtain the predicted topological structure features corresponding to the pocket surface.

[0147] After calling the topology generation module to process the interaction features and pocket surface features to obtain the predicted topological structure features corresponding to the pocket surface, sub-step M2 is executed.

[0148] Sub-step M2: Call the geometry generation module to process the interaction features, the pocket surface features, and the predicted topological structure features to obtain the predicted geometric structure features corresponding to the pocket surface.

[0149] After processing the interaction features and pocket surface features using the topology generation module to obtain the predicted topological structure features corresponding to the pocket surface, the geometry generation module can then be called to process the interaction features, pocket surface features, and predicted topological structure features to obtain the predicted geometric structure features corresponding to the pocket surface.

[0150] Sub-step M3: Use the predicted topological structure features and the predicted geometric structure features as the predicted structure features.

[0151] After obtaining the predicted topological structure features and predicted geometric structure features, these can be used as the predicted structural features. In this example, the predicted topological structure features may include: atom type features and bond relationship features, and the predicted geometric structure features may include: atom position features (such as atom coordinates).

[0152] In the specific implementation, after obtaining the interaction characteristics, these characteristics are sent to the generation module to generate the geometric structure (atomic coordinates) and topological structure (atomic type and bond relationships) of the next atom. Assuming the generated molecule has N (N is a positive integer) atoms, then the generation module needs to iterate N times (e.g., ...). Figure 2 (See module C in the middle).

[0153] In this embodiment, the molecular generation methods can be divided into autoregressive generation and one-time generation. Autoregressive generation methods include ResGen, Pkt2mol, GraphBP, and GraphAF; one-time generation methods include DiffSBDD and DiffBP. One-time generation methods generally require pre-setting a molecular matrix, that is, first predicting the number of atoms, and then predicting the atom categories, position coordinates, and bond relationships of the molecular matrix. In the model provided in this embodiment, an autoregressive approach can be adopted, which can fully interact with the topological pocket and thus learn this geometric matching effect more fully. The specific process is as follows: 1. Anchoring atom generation: After the interaction on the surface of the aggregation pocket, SurfGen selects the starting node from the nodes of the topological pocket. That is, a network is trained to predict the starting probability for each node on the topological pocket, as shown in the formula below:

[0154] (s f V f ) = GVP(s h V h )

[0155]

[0156] Among them, (s h V h Let be a vector in the latent space, and σ be the sigmoid function, which outputs a probability value between [0,1]. The pocket contains generated fragments (t>0).

[0157] When there are already generated atoms inside the protein pocket, the starting atom for the next step can be selected from the generated atoms, and the formula is the same as when t=0.

[0158] To satisfy the SE(3) isovariability of the predicted geometric coordinates, VP can be chosen as the main framework for geometric prediction. The value output by VP satisfies rotation invariance, that is:

[0159] In this embodiment, an NLL loss function can be used as a basis to fit a three-dimensional multivariate Gaussian distribution, namely:

[0160]

[0161]

[0162] (s π ,π π ) = GVP p (s h V h )

[0163] (s μ ,π μ ) = GVP μ (s h V h )

[0164] (s ∑ ,π ∑ ) = GVP ∑ (s h V h )

[0165] μ = μ v

[0166] ∑=MLP ∑ (S ∑ )

[0167] π = MLP π2 (MLP π1 S π +||V π ||2)

[0168] in, π, μ, and ∑ represent the combination coefficients, mean, and variance of the Gaussian mixture density network, respectively.

[0169] Once the positional relationships between an atom and other atoms are known, the type of the atom and its bond relationships with already generated atoms can be predicted. However, considering that the distribution of atom types within the protein pocket also has a certain bias, this embodiment uses the structural aggregation of a geometric kNN graph to predict the characteristics between the atom and its k nearest neighbors in space, and then predicts the atom type and bond length.

[0170]

[0171]

[0172]

[0173] e′ ij ,U′ ij =Attention(e′) ij ,U′ ij )+(e′ ij ,U′ ij )

[0174]

[0175] The Attention module uses a triangular attention mechanism network.

[0176] After obtaining the predicted structural features, proceed to step 105.

[0177] Step 105: Based on the predicted structural features, calculate the loss value of the in-pocket molecule generation model to be trained.

[0178] After obtaining the predicted structural features through the above process, the loss value of the in-pocket molecule generation model to be trained can be calculated based on the predicted structural features. In this example, the loss value of the in-pocket molecule generation model to be trained can be calculated using methods such as cross-entropy loss. Specifically, the method for calculating the loss value can be determined according to business needs, and this embodiment does not impose any restrictions on it.

[0179] After calculating the loss value of the in-pocket molecular generation model to be trained based on the predicted structural features, step 106 is executed.

[0180] Step 106: If the loss value is within a preset range, the trained in-pocket molecule generation model is used as the in-pocket molecule generation model.

[0181] After calculating the loss value of the in-pocket molecular generation model to be trained based on the predicted structural features, it can be determined whether the loss value is within the preset range.

[0182] If the loss value is within a preset range, the trained pocket molecule generation model can be used as the final pocket molecule generation model for subsequent pocket molecule inference scenarios.

[0183] This application's embodiments learn the topological interactions on the surface of protein pockets and the spatial interactions between ligand atoms and pockets, and utilize these interaction features to generate 3D molecules within specified pockets through a geometry and topology generation module. This allows for the generation of corresponding active molecules and their conformations from specified protein target structures that require drug design during use.

[0184] After training the in-pocket molecule generation model, it can be applied to subsequent inference scenarios. The inference process can be described in detail below with reference to the specific implementation method.

[0185] In another specific implementation of this application, after step 106 above, the following may also be included:

[0186] Step N1: Obtain the target protein pocket topology file corresponding to the protein target to be predicted.

[0187] In this embodiment, after training the in-pocket molecule generation model, the target protein pocket topology file corresponding to the protein target to be predicted can be obtained during model inference. Specifically, the protein target PDB structure file can be prepared in advance, specifying the pocket location. Then, given the pocket truncation distance, the pocket topology file is prepared using MSMS, pymesh (both open-source software), etc.

[0188] After obtaining the target pocket topology file corresponding to the protein target to be predicted, proceed to step N2.

[0189] Step N2: Input the target protein pocket topology file into the pocket-based molecule generation model.

[0190] After obtaining the target pocket topology file corresponding to the protein target to be predicted, the target protein pocket topology file can be input into the pocket molecular generation model, and then step N3 can be executed.

[0191] Step N3: Call the interaction learning network layer to process the target pocket surface features in the target protein pocket topology file to obtain the target interaction features corresponding to the pocket surface of the target protein to be predicted.

[0192] After inputting the target protein pocket topology file into the pocket molecule generation model, the interaction learning network layer can be called to process the target pocket surface features in the target protein pocket topology file to obtain the target interaction features corresponding to the pocket surface of the protein target to be predicted.

[0193] After obtaining the target interaction features corresponding to the pocket surface of the protein target to be predicted, step N4 is executed.

[0194] Step N4: Call the generator network layer to process the target interaction features and the target pocket surface features to obtain the target predicted structure features corresponding to the target pocket surface.

[0195] After obtaining the target interaction features corresponding to the pocket surface of the protein target to be predicted, the generative network layer can be called to process the target interaction features and the target pocket surface features to obtain the target prediction structural features corresponding to the target pocket surface.

[0196] The processing procedures for the interaction learning network layer and the generative network layer can be combined with the description of the model training process above, and will not be repeated here in this embodiment.

[0197] Step N5: Based on the predicted structural features of the target, determine the predicted protein molecule corresponding to the protein target to be predicted.

[0198] After obtaining the target predicted structural features, the predicted protein molecule corresponding to the protein target to be predicted can be determined based on the target predicted structural features.

[0199] In this example, the targets can be selected as the COVID-19 target 3CL, and the kinase targets AKT1 and CDK2. The specific process is as follows:

[0200] 1. Prepare the protein target pdb (protein data bank, protein three-dimensional structure data file) structure file and specify the pocket location;

[0201] 2. Given the pocket cutoff distance, prepare the pocket topology file using MSMS, pymesh (both open source software), etc. The default pocket size is 8 angstroms and the resolution is 1.5.

[0202] 3. Using SurfGen to sample the protein topological surface, 200 new active molecules were generated.

[0203] Simultaneously, active molecule data on targets 3CL, AKT1, and CDK2 can be collected. Two hundred molecules are randomly selected from the drug library as random molecules, and two sets of molecules are generated using GraphBP and Pocket2Mol to analyze their binding energy distribution. Figure 3 As shown, the overall activity distribution of molecules generated by SurfGen (i.e., the pocket molecule generation model in this example) is better than that of Pocket2Mol and GraohBP methods, and some molecules even outperform those generated by existing experimental data. Molecules generated by Pocket2Mol are closer to random molecules; molecules generated by GraphBP even have lower activity than randomly selected molecules.

[0204] In practical applications, various drug resistance problems often arise, meaning that the therapeutic effect of a certain drug decreases significantly or even becomes ineffective. This situation is particularly common in cancer medication. The main mechanism of drug resistance is the mutation of the protein pocket that exerts the drug's effect, altering the interaction between the pocket and the molecule, thus causing the drug to go off-target. Drug resistance is also an important aspect of personalized medicine.

[0205] SurfGen offers a novel molecular generation approach to address drug resistance: when a drug target mutates, the physical and geometric properties of its pocket change accordingly. This change can be captured by SurfGen, generating molecules that interact with the corresponding mutated target. Using an antibacterial target—shikimate kinase—as an example, SurfGen demonstrates its ability to capture the molecular distribution within the pocket under different mutations, such as... Figure 4 As shown, first from Figure 4 As seen in C, the molecular backbone generated based on the wild-type structure is similar to the native ligand, both using a six-membered ring as the backbone structure, and the ligands occupying different spatial positions within the protein pockets differ significantly. This is because all amino acid side chain residues within the pocket are mutated to alanine, and all R groups are mutated to the smallest methyl groups, significantly increasing the space of the protein pocket. Therefore, SurfGen generates larger molecules to geometrically fit the larger mutated pocket. In the all-interact mutant system, six key residues are mutated to methyl groups, resulting in molecules that differ significantly from the native ligands. However, compared to the All Pocket system, which mutates all pockets, the resulting molecules are smaller, as expected. The all_no_interact system mutates all residues in the pocket except for the six key residues. The backbone of the resulting molecules is similar to the ligands in the 1zyu pocket, retaining the carboxyl groups involved in the interaction on the six-membered ring and preserving more energy-matching characteristics. However, geometrically, because all_no_interact also mutates many residues to the smaller alanine, the pocket structure is still much larger than WT, resulting in a larger molecule size.

[0206] Next, we will discuss the experiments involving single-point mutations. In both the Arg58Ala and Arg136Ala systems, it can be observed that the molecules generated by Surfgen still retain the ring backbone and the oxygen-containing group at position 1. This is because the guanidino groups in Arg58 and Arg136 can interact with the oxygen-containing group at position 1 through hydrogen bonding or salt bridging. If these two key residues are mutated (as shown in Arg(58,136)Ala), the generated molecule no longer carries a carboxyl group at position 1 of the six-membered ring, while the rest of the backbone retains its similarity to the native ligand. A similar phenomenon was observed in the Asp34Ala mutation experiment. Since Asp34Ala only mutates the residue at position 34, retaining the two Arg groups at positions 58 and 136, the generated molecule retains the carboxyl groups interacting with positions 58 and 136, but loses the hydroxyl group interacting with Asp34. The molecules generated by the other three mutant systems showed less change in the groups on the ring, possibly because the residues at these three positions did not contribute as significantly to the formation of hydrogen bonds as residues such as Arg.

[0207] The training method for the in-pocket molecule generation model provided in this application embodiment acquires training data, including a protein pocket topology file of a protein molecule. The protein pocket topology file is input into the in-pocket molecule generation model to be trained. The in-pocket molecule generation model to be trained includes an interaction learning network layer and a generation network layer. The interaction learning network layer is invoked to process the pocket surface features in the protein pocket topology file to obtain the interaction features corresponding to the pocket surface of the protein molecule. The generation network layer is invoked to process the interaction features and pocket surface features to obtain the predicted structural features corresponding to the pocket surface. Based on the predicted structural features, the loss value of the in-pocket molecule generation model to be trained is calculated. If the loss value is within a preset range, the trained in-pocket molecule generation model to be trained is used as the in-pocket molecule generation model. This application embodiment learns the topological interactions on the surface of the protein pocket and the spatial interactions between ligand atoms and the pocket, and uses these interaction features to generate 3D molecules within a specified pocket through a geometry and topology generation module. This allows for the generation of corresponding active molecules and their conformations from a specified protein target structure for drug design during use.

[0208] Reference Figure 5 The diagram shows a schematic representation of a training device for a pocket-based molecular generation model provided in an embodiment of this application. Figure 5 As shown, the training device 500 for the in-pocket molecular generation model may include the following modules:

[0209] The training data acquisition module 510 is used to acquire training data, which includes: protein pocket topology files of protein molecules;

[0210] The topology file input module 520 is used to input the protein pocket topology file into the in-pocket molecule generation model to be trained. The in-pocket molecule generation model to be trained includes: an interaction learning network layer and a generation network layer.

[0211] The interaction feature acquisition module 530 is used to call the interaction learning network layer to process the pocket surface features in the protein pocket topology file to obtain the interaction features corresponding to the pocket surface of the protein molecule.

[0212] The predicted structural feature acquisition module 540 is used to call the generator network layer to process the interaction features and the pocket surface features to obtain the predicted structural features corresponding to the pocket surface;

[0213] The loss value calculation module 550 is used to calculate the loss value of the in-pocket molecule generation model to be trained based on the predicted structural features.

[0214] The molecular generation model acquisition module 560 is used to use the trained in-pocket molecular generation model as the in-pocket molecular generation model when the loss value is within a preset range.

[0215] Optionally, the pocket surface features include at least one of the following: surface charge features, hydrogen bonding potential features, hydrophobic interaction features, Gaussian curvature features, mean curvature features, and shape index features.

[0216] Optionally, the interaction learning network layer includes: a first GNN network layer and a second GNN network layer.

[0217] The interaction feature acquisition module includes:

[0218] The topology feature acquisition unit is used to call the first GNN network layer to process the pocket surface features and obtain the topology interaction features corresponding to the pocket surface;

[0219] The geometric feature acquisition unit is used to call the second GNN network layer to process the pocket surface features and the topological interaction features to obtain the geometric interaction features corresponding to the pocket surface;

[0220] An interaction feature acquisition unit is used to acquire the topological interaction features and the geometric interaction features as the interaction features.

[0221] Optionally, the generated network layer includes: a topology generation module and a geometry generation module.

[0222] The predicted structural feature acquisition module includes:

[0223] The predicted topology feature acquisition unit is used to call the topology generation module to process the interaction features and the pocket surface features to obtain the predicted topology structure features corresponding to the pocket surface;

[0224] The predicted geometric feature acquisition unit is used to call the geometry generation module to process the interaction features, the pocket surface features and the predicted topological structure features to obtain the predicted geometric structure features corresponding to the pocket surface;

[0225] A predictive structural feature acquisition unit is used to obtain the predicted topological structural features and the predicted geometric structural features as the predicted structural features.

[0226] Optionally, the predicted topological features include: atom type features and bond relationship features, and the predicted geometric features include: atom position features.

[0227] Optionally, the device further includes:

[0228] The target file acquisition module is used to acquire the target protein pocket topology file corresponding to the protein target to be predicted.

[0229] The target file input module is used to input the target protein pocket topology file into the pocket-in-molecule generation model;

[0230] The target interaction feature acquisition module is used to call the interaction learning network layer to process the target pocket surface features in the target protein pocket topology file to obtain the target interaction features corresponding to the pocket surface of the target protein to be predicted.

[0231] The target prediction structure feature acquisition module is used to call the generator network layer to process the target interaction features and the target pocket surface features to obtain the target prediction structure features corresponding to the target pocket surface;

[0232] The predicted protein molecule determination module is used to determine the predicted protein molecule corresponding to the target protein based on the predicted structural features of the target.

[0233] The training device for the in-pocket molecule generation model provided in this application embodiment acquires training data, including a protein pocket topology file of a protein molecule. The protein pocket topology file is input into the in-pocket molecule generation model to be trained. The in-pocket molecule generation model to be trained includes an interaction learning network layer and a generation network layer. The interaction learning network layer is invoked to process the pocket surface features in the protein pocket topology file to obtain the interaction features corresponding to the pocket surface of the protein molecule. The generation network layer is invoked to process the interaction features and pocket surface features to obtain the predicted structural features corresponding to the pocket surface. Based on the predicted structural features, the loss value of the in-pocket molecule generation model to be trained is calculated. If the loss value is within a preset range, the trained in-pocket molecule generation model to be trained is used as the in-pocket molecule generation model. This application embodiment learns the topological interactions on the surface of the protein pocket and the spatial interactions between ligand atoms and the pocket, and uses these interaction features to generate 3D molecules within a specified pocket through a geometry and topology generation module. This allows for the generation of corresponding active molecules and their conformations from a specified protein target structure for drug design during use.

[0234] This application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the training method of the above-described pocket-based molecular generation model.

[0235] Figure 6 A schematic diagram of the structure of an electronic device 600 according to an embodiment of the present invention is shown. Figure 6 As shown, the electronic device 600 includes a central processing unit (CPU) 601, which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) 602 or loaded from storage unit 608 into random access memory (RAM) 603. The RAM 603 may also store various programs and data required for the operation of the electronic device 600. The CPU 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0236] Multiple components in electronic device 600 are connected to I / O interface 605, including: input unit 606, such as keyboard, mouse, microphone, etc.; output unit 607, such as various types of displays, speakers, etc.; storage unit 608, such as disk, optical disk, etc.; and communication unit 609, such as network card, modem, wireless transceiver, etc. Communication unit 609 allows electronic device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0237] The various processes and handling described above can be executed by processing unit 601. For example, the methods of any of the above embodiments can be implemented as computer software programs tangibly contained in a computer-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 600 via ROM 602 and / or communication unit 609. When the computer program is loaded into RAM 603 and executed by CPU 601, one or more actions of the methods described above can be performed.

[0238] This application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the training method for the above-described pocket-based molecular generation model.

[0239] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0240] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of this application can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0241] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, terminals (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0242] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0243] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal, causing a series of operational steps to be executed on the computer or other programmable terminal to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0244] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0245] The foregoing has provided a detailed description of a training method for an in-pocket molecular generation model, a training device for an in-pocket molecular generation model, an electronic device, and a computer-readable storage medium provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and its core ideas. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A training method for a pocket-based molecular generation model, characterized in that, The method includes: Acquire training data, which includes: protein pocket topology files of protein molecules; The protein pocket topology file is input into the in-pocket molecule generation model to be trained, which includes an interaction learning network layer and a generation network layer. The interaction learning network layer is invoked to process the pocket surface features in the protein pocket topology file to obtain the interaction features corresponding to the pocket surface of the protein molecule; the pocket surface features include at least one of the following: surface charge features, hydrogen bond potential features, hydrophobic interaction features, Gaussian curvature features, mean curvature features, and shape index features; The generative network layer is invoked to process the interaction features and the pocket surface features to obtain the predicted structural features corresponding to the pocket surface; Based on the predicted structural features, the loss value of the pocket-in-molecule generation model to be trained is calculated; If the loss value is within a preset range, the trained in-pocket molecule generation model is used as the in-pocket molecule generation model. The interaction learning network layer includes: a first GNN network layer and a second GNN network layer; The step of calling the interaction learning network layer to process the pocket surface features in the protein pocket topology file to obtain the interaction features corresponding to the pocket surface of the protein molecule includes: The first GNN network layer is invoked to process the pocket surface features to obtain the topological interaction features corresponding to the pocket surface; The second GNN network layer is invoked to process the pocket surface features and the topological interaction features to obtain the geometric interaction features corresponding to the pocket surface; The topological interaction features and the geometric interaction features are used as the interaction features; The generated network layer includes a topology generation module and a geometry generation module. The step of calling the generative network layer to process the interaction features and the pocket surface features to obtain the predicted structural features corresponding to the pocket surface includes: The topology generation module is invoked to process the interaction features and the pocket surface features to obtain the predicted topological structure features corresponding to the pocket surface; The geometry generation module is invoked to process the interaction features, the pocket surface features, and the predicted topological structure features to obtain the predicted geometric structure features corresponding to the pocket surface; The predicted topological features and the predicted geometric features are used as the predicted structural features.

2. The method according to claim 1, characterized in that, The predicted topological features include: atom type features and bond relationship features, and the predicted geometric features include: atom position features.

3. The method according to claim 1, characterized in that, After using the trained in-pocket molecular generation model as the in-pocket molecular generation model, the method further includes: Obtain the target protein pocket topology file corresponding to the protein target to be predicted; The target protein pocket topology file is input into the pocket-in-molecule generation model; The interaction learning network layer is invoked to process the target pocket surface features in the target protein pocket topology file to obtain the target interaction features corresponding to the pocket surface of the target protein to be predicted. The generator network layer is invoked to process the target interaction features and the target pocket surface features to obtain the target predicted structure features corresponding to the target pocket surface; Based on the predicted structural features of the target, the predicted protein molecule corresponding to the protein target to be predicted is determined.

4. A training device for a pocket-based molecular generation model, characterized in that, The device includes: The training data acquisition module is used to acquire training data, which includes: protein pocket topology files of protein molecules; The topology file input module is used to input the protein pocket topology file into the in-pocket molecule generation model to be trained. The in-pocket molecule generation model to be trained includes: an interaction learning network layer and a generation network layer. The interaction feature acquisition module is used to call the interaction learning network layer to process the pocket surface features in the protein pocket topology file to obtain the interaction features corresponding to the pocket surface of the protein molecule; the pocket surface features include at least one of the following: surface charge features, hydrogen bond potential features, hydrophobic interaction features, Gaussian curvature features, mean curvature features, and shape index features; The predicted structural feature acquisition module is used to call the generator network layer to process the interaction features and the pocket surface features to obtain the predicted structural features corresponding to the pocket surface; The loss value calculation module is used to calculate the loss value of the in-pocket molecule generation model to be trained based on the predicted structural features. The molecular generation model acquisition module is used to use the trained in-pocket molecular generation model as the in-pocket molecular generation model when the loss value is within a preset range. The interaction learning network layer includes: a first GNN network layer and a second GNN network layer; The interaction feature acquisition module includes: The topology feature acquisition unit is used to call the first GNN network layer to process the pocket surface features and obtain the topology interaction features corresponding to the pocket surface; The geometric feature acquisition unit is used to call the second GNN network layer to process the pocket surface features and the topological interaction features to obtain the geometric interaction features corresponding to the pocket surface; An interaction feature acquisition unit is used to acquire the topological interaction features and the geometric interaction features as the interaction features; The generated network layer includes a topology generation module and a geometry generation module. The predicted structural feature acquisition module includes: The predicted topology feature acquisition unit is used to call the topology generation module to process the interaction features and the pocket surface features to obtain the predicted topology structure features corresponding to the pocket surface; The predicted geometric feature acquisition unit is used to call the geometry generation module to process the interaction features, the pocket surface features and the predicted topological structure features to obtain the predicted geometric structure features corresponding to the pocket surface; A predictive structural feature acquisition unit is used to obtain the predicted topological structural features and the predicted geometric structural features as the predicted structural features.

5. The apparatus according to claim 4, characterized in that, The predicted topological features include: atom type features and bond relationship features, and the predicted geometric features include: atom position features.

6. The apparatus according to claim 4, characterized in that, The device further includes: The target file acquisition module is used to acquire the target protein pocket topology file corresponding to the protein target to be predicted. The target file input module is used to input the target protein pocket topology file into the pocket-in-molecule generation model; The target interaction feature acquisition module is used to call the interaction learning network layer to process the target pocket surface features in the target protein pocket topology file to obtain the target interaction features corresponding to the pocket surface of the target protein to be predicted. The target prediction structure feature acquisition module is used to call the generator network layer to process the target interaction features and the target pocket surface features to obtain the target prediction structure features corresponding to the target pocket surface; The predicted protein molecule determination module is used to determine the predicted protein molecule corresponding to the target protein based on the predicted structural features of the target.

Images