A method and related apparatus for three-dimensional molecule generation

Abstract

The application discloses a three-dimensional molecule generation method and related equipment, and the method comprises the following steps: acquiring a data set of a drug molecule; constructing a three-dimensional molecule graph and a pharmacophore graph of each drug molecule based on the data set; and performing optimization training on a preset model based on the three-dimensional molecule graph and the pharmacophore graph to obtain a target model; wherein the preset model comprises a pre-training control-free molecule generation model and a pharmacophore control module; and performing three-dimensional molecule generation of a target pharmacodynamic effect by using the target model. The application directly generates a brand-new three-dimensional molecule meeting the target pharmacodynamic effect by using the pre-training model and the pharmacophore control module, breaks through the chemical space limitation of a compound library, avoids one-by-one screening of a large number of compounds, and significantly reduces the consumption of computing resources and the time cost; in addition, by fusing the pharmacophore graph and the three-dimensional molecule graph, it is ensured that the generated molecule meets the spatial conformation and functional group matching at the same time; and the related prediction model technology can be widely applied to the technical field of molecule design.

Description

Technical Field

[0001] This application relates to the field of molecular design technology, and in particular to a method and related equipment for generating three-dimensional molecules. Background Technology

[0002] Ligand-Based Drug Design (LBDD) is a strategy for discovering or optimizing new drug candidates by analyzing the structural features of known active ligands when the structural information of the target protein is lacking. This method typically relies on extracting key structural features of active ligands, such as pharmacophore features, shape matching, or physicochemical properties, and constructing virtual screening models based on these features. Potential candidates are then screened in large compound libraries. However, this traditional approach heavily depends on the chemical spatial coverage and molecular weight of existing compound libraries, resulting in limited screening efficiency and often failing to cover a sufficient number of skeletal structures, thus limiting the discovery of novel compounds. Furthermore, this process often involves significant computational resource consumption, making the screening process expensive and time-consuming. Existing techniques such as the PGMG model (based on SMILES generation) cannot guarantee three-dimensional conformation matching, while the ShEPhERD model exhibits limited performance in the generation of complex drug molecules. Summary of the Invention

[0003] The main objective of this application is to provide a three-dimensional molecular generation method, apparatus, electronic device, storage medium, and program product, aiming to solve at least one problem of the prior art.

[0004] To achieve the above objectives, one aspect of this application proposes a method for generating three-dimensional molecules, the method comprising:

[0005] Obtain a dataset of drug molecules;

[0006] Three-dimensional molecular maps and pharmacophore maps of each drug molecule were constructed based on the dataset;

[0007] The preset model is optimized and trained based on the three-dimensional molecular diagram and pharmacophore diagram to obtain the target model;

[0008] The preset model includes a pre-trained uncontrolled molecule generation model and a pharmacophore control module;

[0009] Three-dimensional molecular generation of target drug efficacy is performed using the target model.

[0010] In some embodiments, constructing a three-dimensional molecular map and pharmacophore map for each drug molecule based on the dataset includes the following steps:

[0011] A three-dimensional molecular map is constructed based on the molecular information of drug molecules;

[0012] The three-dimensional molecular diagram includes an atomic feature matrix, a formal charge matrix, an edge feature matrix, and an atomic coordinate matrix. The atomic feature matrix is ​​obtained by uniquely encoding the chemical type of each atom in the drug molecule based on the molecular information.

[0013] The pharmacophore information of drug molecules is identified using a pre-set chemical information learning tool, and a pharmacophore map is constructed based on the pharmacophore information;

[0014] The pharmacophore diagram includes pharmacophore attributes, formal charge control matrix, pharmacophore connection control information, and atom position control information. The pharmacophore attributes are obtained by one-heat encoding based on the pharmacophore category of each atom in the drug molecule in the pharmacophore information.

[0015] In some embodiments, optimizing and training a preset model based on a three-dimensional molecular map and a pharmacophore map to obtain a target model includes the following steps:

[0016] Based on the three-dimensional molecular map, the uncontrolled molecular generation model is pre-trained through diffusion and denoising processes;

[0017] The structure of the pre-trained uncontrolled molecular generation model during the denoising process is parameter-frozen and used as a molecular generation module;

[0018] The pharmacophore control module is trained based on the three-dimensional molecular map and pharmacophore map, combined with the molecular generation module.

[0019] The target model is obtained based on the pre-trained uncontrolled molecule generation model and the trained pharmacophore control module.

[0020] In some embodiments, based on a three-dimensional molecular map, a pre-training process for an uncontrolled molecular generation model is performed through diffusion and denoising processes, including the following steps:

[0021] Input the 3D molecular graph into the uncontrolled molecular generation model;

[0022] During the diffusion process, noise is added by independently perturbing different properties of the three-dimensional molecular map;

[0023] In the denoising process, based on the result of the noise addition processing, the predicted molecular map is output using an encoding and decoding architecture;

[0024] A loss function is constructed based on the 3D molecular map and the predicted molecular map, and the parameters of the uncontrolled molecular generation model are adjusted using the loss function.

[0025] In some embodiments, noise addition processing that independently perturbs different properties of a three-dimensional molecular map includes the following steps:

[0026] Apply Gaussian noise with zero centroid subspace to continuous properties;

[0027] Discrete diffusion is performed on discrete properties.

[0028] In some embodiments, the encoding / decoding architecture includes a first fully connected layer, an encoder, a decoder, and a second fully connected layer connected in sequence. The encoder and decoder include equivalent multi-level equivariant graph Transformer modules. The equivariant graph Transformer module includes a graph attention layer, a feedforward neural network, and a Dropout regularization layer. Both the 3D molecular graph and the result of the noise-adding processing include atomic feature matrices corresponding to node features, edge feature matrices corresponding to edge features, and atomic coordinate matrices corresponding to node coordinates. In the step of outputting the predicted molecular graph using the encoding / decoding architecture, the graph attention layer specifically performs the following operations:

[0029] Based on node coordinates, geometric information between any two nodes is obtained by using a spatial extraction layer.

[0030] Based on geometric information, node features, and global features, the edge features are updated using an edge information update layer;

[0031] Based on geometric information, edge features, and global features, a node information update layer using a self-attention mechanism updates node features.

[0032] The node information update layer includes a PNA layer, which is used to aggregate paired edge features and geometric information into node representations.

[0033] Based on geometric information and edge features, the EGNN update layer uses the rEGNN update mechanism to update the node coordinates.

[0034] Based on geometric information, edge features, and node features, a global feature update layer is used to perform graph-level aggregation to update global features.

[0035] In some embodiments, both the 3D molecular graph and the predicted molecular graph include atomic feature matrices corresponding to atom types, formal charge matrices corresponding to formal charges, edge feature matrices corresponding to bond types, and atomic coordinate matrices corresponding to atomic coordinates. A loss function is constructed based on the 3D molecular graph and the predicted molecular graph, including the following steps:

[0036] Based on the actual values ​​of atomic coordinates in the 3D molecular diagram and the predicted values ​​of atomic coordinates in the predicted molecular diagram, the first loss is constructed using the mean square error.

[0037] Based on the actual values ​​of atom types in the 3D molecular diagram and the predicted values ​​of atom types in the predicted molecular diagram, a second loss is constructed using cross-entropy loss.

[0038] Based on the actual value of formal charge in the 3D molecular diagram and the predicted value of formal charge in the predicted molecular diagram, a third loss is constructed using cross-entropy loss;

[0039] Based on the actual values ​​of bond types in the 3D molecular diagram and the predicted values ​​of bond types in the predicted molecular diagram, a fourth loss is constructed using cross-entropy loss.

[0040] The first loss, second loss, third loss, and fourth loss are weighted and summed to obtain the loss function;

[0041] The expression for the loss function is:

[0042]

[0043] In the formula, Represents the loss function; , , and Indicates the weight of each loss; Represents cross-entropy loss; Indicates mean square error; , , and These represent the actual values ​​of atom type, formal charge, bond type, and atom coordinates, respectively. , , and The predicted values ​​are represented in order by atom type, formal charge, bond type, and atom coordinates.

[0044] In some embodiments, prior to the step of training the pharmacophore control module, the method further includes the following steps:

[0045] Pharmacophore control strategy based on progressive control strategy configuration of pharmacophore control module;

[0046] The pharmacophore control strategy includes unfuzzy level, fuzzy connection information, centroid atoms, fuzzy partial element level, and fuzzy bond level. The denoising process of the uncontrolled molecular generation model adopts an encoding and decoding architecture. The pharmacophore control module is used to input pharmacophore control information in the decoding stage of the molecular generation module based on the three-dimensional molecular diagram and the pharmacophore diagram through the pharmacophore control strategy.

[0047] In some embodiments, the generation of three-dimensional molecules of the target drug effect using the target model includes the following steps:

[0048] Obtain the pharmacophore model corresponding to the reference molecule;

[0049] The pharmacophore model and the sampled noise molecule data are transformed into a graph structure input to the target model;

[0050] A pre-trained uncontrolled molecule generation model is used to denoise a reference molecule to generate a new molecule. The denoising process includes an encoding stage and a decoding stage. The pharmacophore control module is used to guide the generation of new molecules by inputting pharmacophore control information in the decoding stage based on the three-dimensional molecular map and the pharmacophore map through a preset pharmacophore control strategy.

[0051] Candidate molecules are obtained from new molecules based on matching scores and screening scores.

[0052] In some embodiments, the method further includes the following steps:

[0053] In response to the preservation instruction of the target object, some atoms in the reference molecule are marked as substructures by a mask;

[0054] In this process, the substructure representation must retain the structure, and the remaining parts other than the substructure are regions to be generated.

[0055] Obtain intermediate states from a forward-noised distribution;

[0056] The first region corresponding to the substructure is extracted from the intermediate state based on the mask;

[0057] The region is generated by sampling from the denoising process;

[0058] The second region corresponding to the region to be generated is extracted from the generated region based on the inversion code of the mask;

[0059] The first and second regions are fused to generate new molecules with fixed substructures.

[0060] To achieve the above objectives, another aspect of this application provides a three-dimensional molecular generation apparatus, the apparatus comprising:

[0061] The data acquisition module is used to acquire datasets of drug molecules;

[0062] The data construction module is used to build three-dimensional molecular maps and pharmacophore maps for each drug molecule based on the dataset;

[0063] The model training module is used to optimize and train a preset model based on a 3D molecular diagram and a pharmacophore diagram to obtain the target model;

[0064] The preset model includes a pre-trained uncontrolled molecule generation model and a pharmacophore control module;

[0065] The model application module is used to generate three-dimensional molecules of the target drug effect using the target model.

[0066] In some embodiments, the apparatus further includes a policy configuration module for performing the following operations:

[0067] Pharmacophore control strategy based on progressive control strategy configuration of pharmacophore control module;

[0068] The pharmacophore control strategy includes unfuzzy level, fuzzy connection information, centroid atoms, fuzzy partial element level, and fuzzy bond level. The denoising process of the uncontrolled molecular generation model adopts an encoding and decoding architecture. The pharmacophore control module is used to input pharmacophore control information in the decoding stage of the molecular generation module based on the three-dimensional molecular diagram and the pharmacophore diagram through the pharmacophore control strategy.

[0069] To achieve the above objectives, another aspect of the embodiments of this application proposes an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned method.

[0070] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0071] To achieve the above objectives, another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method.

[0072] The embodiments of this application include at least the following beneficial effects: This application provides a method, apparatus, electronic device, storage medium, and program product for generating three-dimensional molecules. This solution involves acquiring a dataset of drug molecules; constructing a three-dimensional molecular map and pharmacophore map for each drug molecule based on the dataset; optimizing and training a preset model based on the three-dimensional molecular map and pharmacophore map to obtain a target model; wherein the preset model includes a pre-trained uncontrolled molecular generation model and a pharmacophore control module; and using the target model to generate three-dimensional molecules with the target drug effect. The embodiments of this application directly generate novel three-dimensional molecules that meet the target drug effect through the pre-trained model and the pharmacophore control module, breaking through the chemical space limitations of compound libraries. Specifically, it breaks through the dependence of traditional LBDD on existing compound libraries and avoids the limitation of skeletal novelty caused by insufficient chemical space coverage. Furthermore, the embodiments of this application ensure that the generated molecules simultaneously satisfy spatial conformation and functional group matching by fusing pharmacophore maps (functional constraints) and three-dimensional molecular maps (structural constraints). The model-based generation technology of the embodiments of this application avoids the need for one-by-one screening of massive amounts of compounds, significantly reducing computational resource consumption and time costs. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of an implementation environment for the three-dimensional molecular generation method provided in this application embodiment;

[0074] Figure 2This is a schematic flowchart of a three-dimensional molecular generation method provided in an embodiment of this application;

[0075] Figure 3 This is a schematic diagram illustrating an example of the architecture of a preset model provided in the embodiments of this application;

[0076] Figure 4 This is a schematic diagram illustrating an example of the isomorphic graph Transformer module and its hierarchical expansion structure provided in the embodiments of this application;

[0077] Figure 5 This is a schematic diagram illustrating an application example of the control strategy provided in the embodiments of this application;

[0078] Figure 6 This is a schematic diagram illustrating an example of the architecture of the pharmacophore control module fused with the molecule generation module provided in the embodiments of this application;

[0079] Figure 7 This is a schematic diagram illustrating the principle of denoising and reconstruction of molecules based on fixed substructures provided in the embodiments of this application;

[0080] Figure 8 This is a comparative schematic diagram of the molecular generation examples provided in the embodiments of this application;

[0081] Figure 9 This is a comparative schematic diagram of an example of molecular generation based on a fixed substructure provided in the embodiments of this application;

[0082] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0083] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0084] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0085] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.

[0086] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0087] In related technologies, strategies for discovering or optimizing new drug candidates involve analyzing the structural features of known active ligands. This method typically relies on extracting key structural features of active ligands, such as pharmacophore characteristics, shape matching, or physicochemical properties, and then constructing a virtual screening model to screen potential candidates in a large compound library. However, this traditional method heavily depends on the chemical spatial coverage and molecular weight of existing compound libraries, resulting in limited screening efficiency and often failing to cover a sufficient number of skeletal structures, thus limiting the discovery of novel compounds. Furthermore, this process often involves significant computational resource consumption, making the screening process expensive and time-consuming.

[0088] In view of this, this application provides a method and related equipment for generating three-dimensional molecules. This method involves acquiring a dataset of drug molecules; constructing a three-dimensional molecular map and pharmacophore map for each drug molecule based on the dataset; optimizing and training a preset model based on the three-dimensional molecular map and pharmacophore map to obtain a target model; wherein the preset model includes a pre-trained uncontrolled molecular generation model and a pharmacophore control module; and using the target model to generate three-dimensional molecules with the target drug effect. This application directly generates novel three-dimensional molecules that meet the target drug effect through the pre-trained model and pharmacophore control module, breaking through the chemical space limitations of compound libraries. Specifically, it overcomes the dependence of traditional LBDD on existing compound libraries and avoids the limitation of skeletal novelty caused by insufficient chemical space coverage. Furthermore, this application integrates the pharmacophore map (functional constraints) and the three-dimensional molecular map (structural constraints) to ensure that the generated molecules simultaneously satisfy spatial conformation and functional group matching. The model-based generation technology of this application avoids the need for individual screening of massive amounts of compounds, significantly reducing computational resource consumption and time costs.

[0089] It is understood that the three-dimensional molecular generation method provided in this application can be applied to any computer device with data processing and computing capabilities, and this computer device can be various terminals or servers. When the computer device in the embodiment is a server, the server is an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Optionally, the terminal can be a smartphone, tablet, laptop, or desktop computer, but it is not limited to these.

[0090] like Figure 1 The diagram shown is a schematic representation of an implementation environment provided in an embodiment of this application. (Refer to...) Figure 1 The implementation environment includes at least one terminal 102 and a server 101. The terminal 102 and the server 101 can be connected via a network, either wirelessly or via a wired connection, to complete data transmission and exchange.

[0091] Server 101 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0092] Additionally, server 101 can also be a node server in a blockchain network. Blockchain is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms.

[0093] Terminal 102 can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. Terminal 102 and server 101 can be directly or indirectly connected via wired or wireless communication, and this embodiment of the application does not impose any limitations.

[0094] For example, based on Figure 1 The implementation environment shown in this application embodiment provides a three-dimensional molecule generation method. The following description uses the application of this three-dimensional molecule generation method in server 101 as an example. It can be understood that this three-dimensional molecule generation method can also be applied in terminal 102.

[0095] Reference Figure 2 , Figure 2 This is an optional flowchart of the three-dimensional molecule generation method provided in the embodiments of this application. The execution subject of the three-dimensional molecule generation method can be any of the aforementioned computer devices (including servers or terminals). Figure 2 The method may include, but is not limited to, steps S100 to S400.

[0096] Step S100: Obtain the dataset of drug molecules;

[0097] For example, in some specific implementations, dataset acquisition and screening can utilize two publicly available datasets, ChEMBL 35 and GEOM-Drugs, which contain approximately 1.25 million and 430,000 drug molecules, respectively. Furthermore, the Lipinski Five Rules (RO5) and the Synthesis Accessibility Score (SA) can be combined to screen molecules in both datasets, eliminating those with non-drug properties or those difficult to synthesize. This application further limits the constituent atomic types of the molecules to the 10 most common elements (H, C, N, O, F, P, S, Cl, Br, I). Finally, the screened molecules can be divided into a training set (for model training), a validation set (for model validation), and a test set (for model performance evaluation) in an 8:1:1 ratio.

[0098] Step S200: Construct a three-dimensional molecular map and pharmacophore map for each drug molecule based on the dataset;

[0099] It should be noted that in some embodiments, step S200 may include the following steps: constructing a three-dimensional molecular graph based on the molecular information of the drug molecule; wherein, the three-dimensional molecular graph includes an atomic feature matrix, a formal charge matrix, an edge feature matrix, and an atomic coordinate matrix, and the atomic feature matrix is ​​obtained by one-hot encoding based on the chemical type of each atom of the drug molecule in the molecular information; identifying the pharmacophore information of the drug molecule using a preset chemical information learning tool, and constructing a pharmacophore graph based on the pharmacophore information; wherein, the pharmacophore graph includes pharmacophore attributes, a formal charge control matrix, pharmacophore connection relationship control information, and atomic position control information, and the pharmacophore attributes are obtained by one-hot encoding based on the pharmacophore category of each atom of the drug molecule in the pharmacophore information. The three-dimensional molecular graph is thus a three-dimensional pharmacophore molecular graph.

[0100] Specifically, traditional pharmacophore models lack atomic-level geometric information integration, resulting in insufficient accuracy in virtual screening. In contrast, this application constructs a geometrically accurate molecular graph using atomic feature matrices (chemical type), coordinate matrices (spatial location), and edge feature matrices (bonding relationships), providing a structural foundation for the generated model. Furthermore, this application encodes pharmacophore attributes into a learnable graph structure (including connectivity and positional control), achieving strong constraints on functional groups and spatial positions, thereby improving the activity reliability of the generated molecules.

[0101] For example, in some specific embodiments, a molecule containing 𝑛 atoms can be represented as a three-dimensional molecular diagram. ,in It is an atomic characteristic matrix. , Indicates the first The chemical type of each atom, via a one-hot mapping function. Encoded as an m-dimensional vector . It is a formal charge matrix. . It is an edge feature matrix, representing the type of bond between atoms. It is an atomic coordinate matrix.

[0102] Similarly, pharmacophore information of molecules can be constructed into pharmacophore diagrams. , used to guide the molecular generation process. Among them: The pharmacophore properties of atoms are represented by categories including hydrogen bond donor, hydrogen bond acceptor, aromatic ring, hydrophobic group, positive charge, negative charge, and linker, which are mapped using a one-hot mapping function. Encoded as Dimensional vector. Pharmacophore information is identified by RDKit (Rational Discovery Kit, an open-source cheminformatics and machine learning toolkit designed for drug discovery, molecular simulation, chemical data analysis, etc.). Atoms not belonging to any pharmacophore are marked as connectors by default. Atoms with the same pharmacophore label and connected are grouped into a pharmacophore node, and randomly selected from it... One node is used to control the input, and the remaining atoms are uniformly labeled as connectors; Formal charge control matrix (all values ​​are 0); and These are the control information for the pharmacophore connection relationship and the atomic position, which can be adjusted according to the pharmacophore control strategy required in subsequent steps.

[0103] Step S300: Optimize and train the preset model based on the three-dimensional molecular diagram and pharmacophore diagram to obtain the target model;

[0104] The preset model includes a pre-trained uncontrolled molecule generation model and a pharmacophore control module;

[0105] For example, such as Figure 3 As shown, the model in this embodiment consists of a pharmacophore control module and a pre-trained uncontrolled molecule generation model, where a represents the first training stage and b represents the second training stage.

[0106] It should be noted that in some embodiments, step S300 may include the following steps: pre-training an uncontrolled molecular generation model based on a three-dimensional molecular map through a diffusion process and a denoising process; freezing the parameters of the pre-trained uncontrolled molecular generation model in the denoising process as a molecular generation module; training a pharmacophore control module based on a three-dimensional molecular map and a pharmacophore map, combined with the molecular generation module; and obtaining a target model based on the pre-trained uncontrolled molecular generation model and the trained pharmacophore control module.

[0107] Specifically, this application employs a decoupled training strategy, which achieves the following: the diffusion model learns general molecular distributions, mastering chemical rules and conformational diversity; the parameters of the molecule generation module are frozen, and only the pharmacophore control module is trained, avoiding catastrophic forgetting and improving training stability and generalization. Furthermore, this application supports flexible replacement of the pharmacophore control module through modular design, adapting to different drug design needs. Additionally, adjustments can be made to the pharmacophore control strategy requirements corresponding to the pharmacophore control module to achieve adaptation to different drug designs.

[0108] For example, in some specific implementations, model training is divided into two stages: First, a basic unconditional diffusion model is trained with the goal of learning how to generate reasonable complete molecular structures from noise; then, the generator parameters are frozen and only the pharmacophore control module is trained to achieve fine control over conditional generation, accelerate model convergence, and improve pharmacophore matching ability.

[0109] In some embodiments, pre-training an uncontrolled molecular generation model based on a three-dimensional molecular graph through a diffusion process and a denoising process may include the following steps: inputting the three-dimensional molecular graph into the uncontrolled molecular generation model; in the diffusion process, adding noise by independently perturbing different properties of the three-dimensional molecular graph; in the denoising process, outputting a predicted molecular graph based on the result of the noise addition using an encoding / decoding architecture; constructing a loss function based on the three-dimensional molecular graph and the predicted molecular graph, and using the loss function to adjust the parameters of the uncontrolled molecular generation model.

[0110] Specifically, the embodiments of this application employ attribute-independent perturbation, which enables the following: continuous attributes (coordinates) to use zero-centroid Gaussian noise, maintaining the overall translational invariance of the molecule; discrete attributes (atoms / bonds) to use discrete diffusion, avoiding continuous noise interference with the classification task. Furthermore, the embodiments of this application employ an encoding-decoding prediction architecture, reconstructing the original molecule through a denoising process, forcing the model to learn the structural distribution patterns in chemical space, thus laying the foundation for generative capabilities.

[0111] In some embodiments, the noise addition process of independently perturbing different properties of the three-dimensional molecular graph may include the following steps: applying Gaussian noise of the zero centroid subspace to continuous properties; and performing discrete diffusion on discrete properties.

[0112] Specifically, the embodiments of this application utilize zero centroid noise to ensure that noise injection does not disrupt the overall geometric symmetry of the molecule (such as rotation / translation invariance), thereby improving the rationality of coordinate generation. In addition, discrete diffusion is applied to categorical variables (such as atom type) to avoid prediction bias caused by continuous noise, thereby improving the accuracy of discrete attribute generation.

[0113] For example, in some specific implementations, the pre-trained uncontrolled molecular generation model uses a denoised diffusion model as its framework. During the forward noise addition process, independent perturbations are applied to different attributes of the molecular graph. For continuous attributes, such as position, Gaussian noise is applied to the zero centroid (CoM) subspace. Ensure that rotation and translation are equally variable:

[0114]

[0115] in, Control how much signal is retained in each step. The amount of noise added is controlled by the following constraints: The noise is distributed in dimension The linear subspace. For discrete properties such as atom type, formal charge, and bond type, discrete diffusion is used. Here, the transition matrix for the atom type is defined:

[0116]

[0117] in The marginal distribution of atom types in the training set. The transition matrices for formal charge and bond types are defined in the same way. and The overall noise model can be jointly expressed as:

[0118]

[0119] In some embodiments, the encoding / decoding architecture includes a first fully connected layer, an encoder, a decoder, and a second fully connected layer connected sequentially. The encoder and decoder include equivalent multi-level equivariant graph Transformer modules. The equivariant graph Transformer module includes a graph attention layer, a feedforward neural network, and a Dropout regularization layer. Both the 3D molecular graph and the result of noise addition processing include atomic feature matrices corresponding to node features, edge feature matrices corresponding to edge features, and atomic coordinate matrices corresponding to node coordinates. In the step of outputting the predicted molecular graph using the encoding / decoding architecture, the graph attention layer specifically performs the following operation: based on the node coordinates, it uses a spatial extraction layer to process and obtain... The graph considers geometric information between any two nodes. Based on geometric information, node features, and global features, an edge information update layer updates edge features. A node information update layer using a self-attention mechanism updates node features based on geometric information, edge features, and global features. The node information update layer includes a PNA layer, which aggregates paired edge features and geometric information into node representations. An EGNN update layer using positional information updates node coordinates based on geometric information and edge features, employing an rEGNN update mechanism. Finally, a global feature update layer performs graph-level aggregation based on geometric information, edge features, and node features to update global features.

[0120] Specifically, the embodiments of this application utilize a spatial extraction layer to explicitly model geometric relationships such as distances / angles between atoms; furthermore, the rEGNN coordinate update strictly maintains the equivariance of coordinates; simultaneously, the PNA layer aggregates edge information to enhance the node representation's ability to perceive local structures; finally, global feature aggregation supports molecular-level attribute (such as polarity) control, realizing multi-scale molecular generation.

[0121] Exemplary examples, such as in some specific implementations, Figure 3 As shown, the denoising process can predict atom type, charge, bond type, and coordinates based on a 12-layer E(3) isomorphic graph Transformer module (EGT) model architecture, which includes a 6-layer encoder and a 6-layer decoder. The encoder and decoder modules have the same architecture, both adapted from the traditional transformer architecture. Figure 4 As shown, the isomorphic graph Transformer module mainly consists of a graph attention layer, a feedforward neural network, and a Dropout regularization layer. The Add& normalization layer represents the residual connections and the normalized network layer. Specifically, this application incorporates a novel rEGNN layer into the graph attention layer, which fully utilizes the information interaction between various features during the update process. Edge features are first processed through the geometric information between nodes. Node features and global features yUpdates are then performed. Node features are updated through a self-attention mechanism, where attention weights depend on both edge features and... After the attention head is expanded, the obtained value will be further combined with the converged edge features, And global features are fused. For paired features (e.g., edge features) and The nodes are aggregated and represented in this application using a PNA (Principal Neighbourhood Aggregation) layer, which takes the following form:

[0122]

[0123] Global features are updated by aggregating all other features at the whole-image level. Coordinate updates use the rEGNN update mechanism, whose message function is... The updated edge features are used as input.

[0124] In some embodiments, both the 3D molecular graph and the predicted molecular graph include atomic feature matrices corresponding to atom types, formal charge matrices corresponding to formal charges, edge feature matrices corresponding to bond types, and atomic coordinate matrices corresponding to atomic coordinates. Constructing a loss function based on the 3D molecular graph and the predicted molecular graph may include the following steps: constructing a first loss using mean squared error based on the actual values ​​of atomic coordinates in the 3D molecular graph and the predicted values ​​of atomic coordinates in the predicted molecular graph; constructing a second loss using cross-entropy loss based on the actual values ​​of atom types in the 3D molecular graph and the predicted values ​​of atom types in the predicted molecular graph; constructing a third loss using cross-entropy loss based on the actual values ​​of formal charges in the 3D molecular graph and the predicted values ​​of formal charges in the predicted molecular graph; constructing a fourth loss using cross-entropy loss based on the actual values ​​of bond types in the 3D molecular graph and the predicted values ​​of bond types in the predicted molecular graph; and weighted summing the first, second, third, and fourth losses to obtain the loss function; wherein the expression for the loss function is:

[0125]

[0126] In the formula, Represents the loss function; , , and Indicates the weight of each loss; Represents cross-entropy loss; Indicates mean square error; , , and These represent the actual values ​​of atom type, formal charge, bond type, and atom coordinates, respectively. , , and The predicted values ​​are represented in order by atom type, formal charge, bond type, and atom coordinates.

[0127] Specifically, the embodiments of this application apply attribute-based loss design: coordinates (continuous) → mean square error (MSE), atoms / bonds / charges (discrete) → cross-entropy (CE); finally, weighted fusion is performed through weight coefficients to balance the contributions of various losses and simultaneously optimize structural rationality and chemical rationality.

[0128] For example, in some specific implementations, an uncontrolled molecule generation model is first trained. The mean squared error of coordinates (MSE) is jointly optimized with the cross-entropy loss of formal charge / atom / bond type (using...). express):

[0129]

[0130] in Represents coordinates, , , These represent predictions for atom type, formal charge, and bond type, respectively. This indicates the weight of each loss.

[0131] In addition, in some alternative implementations, after training the uncontrolled molecule generation model, the parameters of the pre-trained generation module are fixed, and only the network layers in the pharmacophore control module are trained.

[0132] In some embodiments, prior to training the pharmacophore control module, the method may further include the following steps: configuring a pharmacophore control strategy for the pharmacophore control module based on a progressive control strategy; wherein the pharmacophore control strategy includes unfuzzy levels, fuzzy connection information, centroidal atoms, fuzzy partial element levels, and fuzzy bond levels; the denoising process of the uncontrolled molecular generation model adopts an encoding / decoding architecture, and the pharmacophore control module is used to input pharmacophore control information in the decoding stage of the molecular generation module based on the three-dimensional molecular graph and the pharmacophore graph, through the pharmacophore control strategy.

[0133] Specifically, this application employs a progressive control strategy, moving from strict constraints (unambiguous) to weak constraints (fuzzy keys) in stages to balance pharmacophore conformance and structural degrees of freedom. Furthermore, this application injects control signals only during the denoising decoding stage, avoiding disruption of the pre-trained model's knowledge representation and enhancing generative diversity.

[0134] For example, in some specific embodiments, the pharmacophore control strategy can be implemented as follows:

[0135] The goal of this application's model is to generate three-dimensional pharmacophore conformations that match the template molecule or pharmacophore hypothesis. During the generation process, this application employs a progressive control strategy, gradually simplifying the input control information to improve the model's generation capability under pharmacophore guidance and its matching ability with the pharmacophore model. The control strategy includes:

[0136] 1. No fuzzy level: Preserves the true pharmacophore type and 3D coordinates of all atoms, and simplifies edge features to single bond connections;

[0137] 2. Fuzzy edge: Preserve the connection bonds between pharmacophores, and delete the bonds between atoms within the same pharmacophore;

[0138] 3. Centroid alignment: Merging atoms within the same pharmacophore to their geometric center;

[0139] 4. Fuzzy element level: Randomly convert some pharmacophores and their atoms into connectors, and merge adjacent atoms that are both connectors to the centroid position;

[0140] 5. Fuzzy Bond Level (Free): Remove all interatomic bonds and move all connecting atoms to the origin in space.

[0141] The definitions and settings of each control strategy are shown in Table 1 and Figure 5 ( Figure 5 The molecular structures obtained by representing the structures corresponding to the labels in the diagram using the aforementioned labeling strategy (the black and white line structures represent the original molecular structures) are shown in the figure. It is worth noting that this application assumes all atoms are connected, thus avoiding redundant judgments during the mapping process. Through the above stepwise control, each atom in the DiffPharm model carries two types of control information: its pharmacophore attribute and spatial position. This control information is input into the pharmacophore control module, guiding the generation of a molecular conformation matching the expected pharmacophore. Experiments show that this series of control operations plays a crucial role in generating target molecules with reasonable structures and well-defined functions.

[0142] It should also be noted that, such as Figure 3 and Figure 6 As shown, the pharmacophore control module in this embodiment of the application: introduces a zero-weight fully connected layer to reference the pharmacophore conditions. It is introduced and fused into the molecular generation module through the isomorphic graph Transformer.

[0143] Furthermore, to meet the requirement of preserving known active structures in some tasks, this invention further supports masking and fixing some atoms in the reference molecule during molecular generation, reconstructing and growing only the remaining regions. By combining pharmacophore maps with fixed substructure masks, the generative model is guided to achieve local optimization and expansion of the structure while keeping the known key structures unchanged, making it suitable for tasks such as skeleton skipping, fragment connection, and fragment growth.

[0144] Step S400: Use the target model to generate three-dimensional molecules of the target drug effect;

[0145] It should be noted that in some embodiments, step S400 may include the following steps: obtaining the pharmacophore model corresponding to the reference molecule; converting the pharmacophore model and the sampled noisy molecule data into a graph structure input target model; denoising the reference molecule using a pre-trained uncontrolled molecule generation model to generate a new molecule; wherein, the denoising process includes an encoding stage and a decoding stage, and the pharmacophore control module is used to input pharmacophore control information in the decoding stage based on the three-dimensional molecular graph and the pharmacophore graph to guide the generation of new molecules through a preset pharmacophore control strategy; and selecting candidate molecules from the new molecules based on the matching score and the screening score.

[0146] Specifically, this embodiment automates the entire process by using a reference molecule + pharmacophore model → graph structure input → diffusion model generation → screening. Furthermore, this embodiment dynamically injects pharmacophore constraints during the denoising process to ensure the generated molecules have clearly defined functional orientations. Finally, the quality of candidate molecules is guaranteed through dual filtering using matching scores (pharmacophore conformity) and screening scores (drug-likeness).

[0147] For example, in some specific implementations, during the application phase, the user inputs a reference molecule and its pharmacophore model. Based on the trained model, the molecule generation process is guided by pharmacophore control information, and a new molecule that meets the three-dimensional matching requirements is output. The output results are then calculated using a match score and a screening score (such as a phase screen score) to select candidate molecules with potential activity.

[0148] In this process, molecules are generated using a pre-trained model, and samples are collected only under pharmacophore control. Molecules are filtered using a match score, assuming a target pharmacophore model. Include There are 1 pharmacophore points, each represented as... ,in Represents three-dimensional spatial coordinates. Indicates the type of pharmacophore. The resulting molecule is represented as a set of atoms. Each atom Corresponding to spatial location and pharmacophore type tags The matching score is defined as follows:

[0149]

[0150] In some embodiments, the method may further include the following steps: in response to a retention instruction for the target object, marking a portion of atoms in a reference molecule as substructures using a mask; wherein the substructures characterize the structures to be retained, and the remaining portion excluding the substructures is the region to be generated; obtaining an intermediate state from a forward-noised distribution; extracting a first region corresponding to the substructure from the intermediate state according to the mask; sampling the generated region from the denoising process; extracting a second region corresponding to the region to be generated from the generated region according to the inversion code of the mask; and fusing the first region and the second region to generate a new molecule with a fixed substructure.

[0151] Specifically, this application embodiment generates the structure through mask partitioning: the original state of the retained substructure (mask region) is extracted from the forward noise; the region to be generated (inverse mask) is sampled from the denoising process. Furthermore, through structural fusion: the retained part is spliced ​​with the newly generated part to achieve local optimization under the fixed core framework, thereby accelerating the iteration of the lead compound.

[0152] Exemplary examples, in some specific embodiments, in pharmacophore-guided molecular repair tasks, embodiments of this application can utilize molecules... A subset of atoms is defined as the substructure that must be retained. The remaining part is defined as the region to be generated. Substructure The atoms are uniquely identified throughout the molecule using a set of mask indices, which are used for explicit fixation in the diffusion model. At each time step of the denoising sampling, the fixed substructure and the generated region are fused in the following manner:

[0153] (1) Obtain from the forward-noiseed distribution t intermediate state at time -1 : ;

[0154] (2) Extract the fixed substructure part based on the mask: ;

[0155] (3) Sample the generated region from the conditional diffusion model: ;

[0156] (4) Extract information about the region to be generated: ;

[0157] (5) Merge the fixed region and the generated region: ;

[0158] By from arrive By traversing the Markov chain in reverse order, this application progressively completes the denoising and reconstruction of the molecule. An overall schematic diagram is shown below. Figure 7 As shown. Due to the diffusion model in When the noise level approaches zero, the resulting molecule can completely retain the specified substructure. This avoids disturbance during the generation process. At the same time, this mechanism also allows the model to perceive the structural information of fixed regions at each denoising step, thereby improving the consistency between the overall generated result and the prior substructure and avoiding the problem of "ignoring the free generation of fixed regions".

[0159] To explain in detail the principles of the technical solution of this application, the overall process of this application will be described below with reference to some specific embodiments. It is easy to understand that the following is an explanation of the technical principles of this application and should not be regarded as a limitation of this application.

[0160] First, it's important to note that with the development of deep learning, molecular generation models have become an emerging direction in LBDD (Label-Based Drug Discovery). These models can directly generate drug-like molecules at the SMILES (Simplified Molecular Input Line Entry System) representation, molecular graph, or 3D structural level, thus serving as a powerful supplement to traditional compound libraries. Especially with the advancements in architectures such as molecular graph neural networks, graph generation models, variational autoencoders (VAEs), generative adversarial networks (GANs), and diffusion models, the quality and diversity of generated molecules have significantly improved. However, despite these generative models enhancing the structural novelty of molecules, the generation process is mostly unconstrained and spontaneous, lacking property control tailored to task requirements. This means that the generated molecules still require large-scale screening using virtual screening methods, thus not substantially reducing the cost of drug screening.

[0161] To address this, researchers have attempted to incorporate structure- or function-guided information into the molecular generation process, particularly utilizing the pharmacophore, a widely used structural simplification model in drug design, to guide molecular generation. Early studies often employed fragment-based editing methods, such as the 3D-MCTS method, which combines three-dimensional pharmacophores with structure optimization strategies to generate molecules through fragment combination. However, this approach still adheres to the paradigm of "modifying existing molecules" rather than "generating from scratch." Furthermore, there is the PGMG model, a 2D graph generation model based on pharmacophore constraints. This model uses graph neural networks to encode pharmacophore features and a Transformer as the decoder to generate molecules that conform to the constraints. Xie et al.'s TransPharmer model further incorporates the GPT framework, jointly inputting pharmacophore fingerprints and SMILES encodings to achieve scaffold-transition-style structure generation under pharmacophore constraints.

[0162] Meanwhile, diffusion models, as a type of neural network with strong generative capabilities, are widely used in LBDD and Structural-Based Drug Design (SBDD). The guided generation methods of diffusion models are mainly divided into two categories: conditional injection during the training phase and conditional control during the sampling phase. In terms of conditional injection during the training phase, the related technology DiffLinker uses the E(3) equal-variable diffusion mechanism to take the initial fragment as the conditional input to realize the generation of connecton molecules. The Graph DiT developed by Liu et al. combines the Transformer architecture and introduces attribute information through the encoder to complete the generation of multi-attribute conditional molecules. The related technology ShEPhERD model further integrates three-dimensional molecular structure, electrostatic potential, and pharmacophore features, and performs joint modeling using the SE(3) equal-variable diffusion method. In terms of guidance during the sampling phase, the GaUDI model of Tomer et al. introduces the gradient of the objective function in each step of the denoising process to realize reverse design guided by the target property. The ShapeMol of Chen et al. introduces a shape embedding module to ensure that the generated molecule is highly similar to the reference molecule in three-dimensional morphology.

[0163] While the aforementioned methods significantly improve the control over the properties of generated molecules, several key issues remain. Firstly, most methods require de novo model training for specific tasks, resulting in high training costs and a lack of efficient utilization of pre-trained molecule generation models. Secondly, existing models have limited pharmacophore control capabilities, often lacking precise constraints on pharmacophore matching in three-dimensional space. Furthermore, some models still rely on structure optimization or fragment editing, making it difficult to fully leverage the "de novo generation" potential of deep generative models.

[0164] Furthermore, some attempts have focused on combining pharmacophores with 3D diffusion models for molecular conformation alignment and generation. This approach aligns molecular conformations with pharmacophore models to obtain conformations with higher matching degrees. While this has shown good results in virtual screening, its focus remains on structural alignment and screening, lacking an end-to-end generation strategy and control over complex structural features. Therefore, there is an urgent need for a generation method that can flexibly incorporate functional constraints such as pharmacophores, possesses 3D spatial modeling capabilities, and fully utilizes the capabilities of pre-trained models to address the problems of low efficiency, weak constraints, and insufficient generalization ability in existing drug generation methods.

[0165] It's important to note that PGMG is a pharmacophore-guided deep learning method designed to generate biologically active molecules. This method represents pharmacophores as complete graphs with node and edge attributes, encodes them using graph neural networks, and generates molecules using a Transformer decoder. Furthermore, PGMG introduces latent variables to handle the many-to-many mapping between pharmacophores and molecules, thereby increasing the diversity of generated molecules. PGMG can handle different types of pharmacophore inputs and is suitable for ligand-based and structure-based drug design; the introduction of latent variables results in highly diverse structurally generated molecules. However, it has the following drawbacks: a lack of three-dimensional structural information; PGMG is primarily based on SMILES representations and does not fully utilize the three-dimensional conformational information of molecules, potentially limiting its application in tasks requiring precise three-dimensional ligand conformations. PGMG also has relatively limited control over preserving specific substructures or performing targeted molecule optimization.

[0166] ShEPhERD is an SE(3)-isovariant diffusion model designed to jointly process the 3D shape, charge distribution, and pharmacophore information of molecules. This method achieves fine control over the molecular generation process by simultaneously modeling the 3D structure and interaction characteristics of molecules during diffusion. ShEPhERD can simultaneously process the 3D shape, charge distribution, and pharmacophore information of molecules, improving the bioactivity relevance of generated molecules and demonstrating good performance in tasks such as natural product analogy, protein-blind active molecule diversification, and bioisosteric fragment merging. However, it has the following drawbacks: due to the need to process complex 3D information, ShEPhERD requires high computational resources during training and inference; it is not flexible enough for tasks requiring the preservation of specific molecular substructures; and the ShEPhERD conditional model is trained on the GDB17 dataset, but the molecular structures in this dataset are relatively simple, with a limited number and types of atoms, which may not be sufficient to cover the complexity of real drug molecules.

[0167] Therefore, the purpose of this application is to construct a molecular generation model (i.e., DiffPharm of this application) that can integrate three-dimensional pharmacophore constraints, support substructure preservation, and has efficient and controllable generation capabilities, so as to improve the bioactivity relevance, structural diversity and task adaptability of generated molecules.

[0168] To address the shortcomings of existing molecular generation methods, such as the lack of three-dimensional pharmacophore control, difficulty in preserving substructures, and uncontrollable generation processes, this application proposes a molecular diffusion generation method, DiffPharm, which integrates three-dimensional pharmacophore guidance and structure preservation mechanisms. This method comprises the following key modules and technical processes:

[0169] 1. Model Architecture Design:

[0170] The core system DiffPharm of this application includes the following three main modules:

[0171] (1) Pre-trained unconditional diffusion model module:

[0172] An unconditional 3D molecular generator is constructed based on E(3)-equal variation graphical neural network (EGNN) to progressively generate drug-like molecules with reasonable conformations, atomic properties, and bonding structures from noise. This module is trained based on general molecular data and has strong conformational modeling capabilities.

[0173] (2) Pharmacophore control module:

[0174] Employing a zero-initialization multilayer perceptron (MLP) and an isomorphic graph Transformer structure, pharmacophore features (including pharmacophore type, three-dimensional position, connectivity, and charge) from a reference molecule are introduced as conditions to dynamically adjust the prediction at each step of the diffusion process, thereby achieving pharmacophore guidance at a three-dimensional level.

[0175] (3) Substructure preservation function (DiffPharm-fix):

[0176] To address the need for certain structural elements (such as key fragments and the backbone) to remain unchanged in molecular optimization tasks, a substructure conditional injection mechanism was designed. During diffusion-backward denoising, the system automatically retains information from a specified subset of atoms, optimizing only the remaining parts, thereby achieving "repair-type" molecular generation.

[0177] 2. Data preprocessing workflow:

[0178] A training sample library is constructed using public datasets (such as GEOM-DRUGS and ChEMBL). Molecules are preprocessed in multiple steps, including low-energy conformation screening, atomic pharmacophore attribute standardization (such as pharmacophore type linkerization), structural simplification (such as bond type normalization), and point cloudification, to construct multi-scale pharmacophore control input data.

[0179] 3. Model training mechanism:

[0180] DiffPharm model training consists of two stages:

[0181] First, a basic unconditional diffusion model is trained with the goal of learning how to generate reasonable and complete molecular structures from noise.

[0182] Subsequently, the generator parameters were frozen, and only the pharmacophore control module was trained to achieve fine control over conditional generation, accelerate model convergence, and improve pharmacophore matching ability.

[0183] A multi-objective loss function was designed during training, where the estimation of coordinates R can be optimized using mean squared error, while the prediction of atom types... Prediction of formal charge Key type prediction This corresponds to classification problems, and can be solved using cross-entropy loss (in the formula). The optimization is performed using these four methods. The ultimate training goal of DiffPharm is the weighted sum of these four losses:

[0184]

[0185] This represents the weight of each item. This ensures that all dimensions of the generated molecule satisfy the target distribution.

[0186] 4. Molecular generation and screening process:

[0187] During the application phase, the user inputs a reference molecule and its pharmacophore model. The DiffPharm system guides the molecule generation process based on the pharmacophore control information, outputting new molecules that meet the three-dimensional matching requirements. The output results are then calculated using a match score and a screening score (such as a phase screen score) to select candidate molecules with potential activity.

[0188] 5. Molecular repair and directed structure generation:

[0189] To support the structural optimization of existing molecules, this application introduces a dual-guided mechanism of "substructure preservation + pharmacophore control". The user specifies the subset of atoms to be preserved, while the remaining parts are noise-added and recovered during the reverse diffusion process. This generates optimized molecules without destroying the critical structure, enabling advanced functions such as targeted repair and skip design.

[0190] For example, the generation of experiments through screening molecules is as follows: Figure 8 As shown, Figure 8 Example diagrams are shown of molecules generated by the model of this application (DiffPharm) and other comparative models (PGMG and ShEPhERD) under the control of pharmacophore models (References 1 to References 6) of six reference molecules. The sampling process of the target model of this application under dual control of pharmacophore and substructure is as follows: Figure 3 As shown, another example of molecule formation is as follows: Figure 9 As shown, Figure 9 The pharmacophore models of four reference molecules are shown (e.g.) Figure 9 (Example of the References column) and fixed substructures (fixed substructures such as...) Figure 9 Example diagram of molecules generated under dual control of the present application model (DiffPharm) and other comparative models (PGMG and ShEPhERD) in the Pharmacophore column.

[0191] In summary, considering the shortcomings of the prior art, the embodiments of this application implement the following technical solution:

[0192] 1. A three-dimensional pharmacophore-guided diffusion molecule generation mechanism: This application introduces pharmacophore constraints into the diffusion model in a three-dimensional form and constructs a pharmacophore control module, enabling the generated molecules to highly match the reference pharmacophore at the three-dimensional conformation level. This mechanism not only enhances the bioactivity relevance of the generated molecules but also overcomes the limitations of traditional two-dimensional structure or fragment splicing methods in terms of three-dimensional spatial constraints.

[0193] 2. Model parameter reuse and control module decoupling: By pre-training the unconditional diffusion model and freezing the backbone parameters, only the pharmacophore control module parameters are trained, which greatly reduces training costs and accelerates model convergence. At the same time, the generality of the main model is maintained, making this approach highly portable and scalable.

[0194] 3. Predefined substructure retention mechanism (DiffPharm-fix): Introducing a substructure retention mechanism, during the diffusion generation process, the retention and surrounding reconstruction of a specified subset of atoms (such as the core framework or pharmacophore) are achieved through structural masks and conditional constraints. This supports more complex optimization tasks, such as "local modification optimization based on the original drug", and improves the practicality and targeted design capabilities of the generated product.

[0195] 4. Multi-level pharmacophore representation and fusion mechanism. This application designs a multi-level pharmacophore feature encoding method (point cloudification, attribute simplification, etc.) and fuses them through an isomorphic graph Transformer, which effectively improves the expressive power and guidance effect of pharmacophore features, and is more robust and expressive than the traditional simple vector embedding method.

[0196] Compared to existing technologies, the embodiments of this application have at least the following beneficial effects:

[0197] (1) Achieve precise control of three-dimensional pharmacophores in molecular generation. Compared with traditional pharmacophore-guided methods that are based solely on two-dimensional structure or fragment optimization, DiffPharm introduces three-dimensional pharmacophore information as a "plug-in" through the pharmacophore control module. This ensures a high degree of spatial matching between the generated molecules and the pharmacophore model at the molecular conformation level, effectively enhancing the bioactivity relevance of the generated molecules.

[0198] (2) Supports partial structure retention and targeted repair, improving design flexibility. The system in this application supports users to specify key structures (such as core skeleton, active groups, etc.) in the molecule to remain unchanged, and only guides the generation of the rest, which meets common needs in molecular optimization and is suitable for actual drug design tasks such as hit-to-lead or lead optimization.

[0199] (3) Modular control mechanism based on pre-trained model to improve training efficiency and transferability. The pharmacophore control module of this application is superimposed on the existing diffusion generator. It does not require training the entire model from scratch. Only the control module needs to be tuned to adapt to different pharmacophore requirements, which significantly reduces training cost and improves the transfer and generalization ability of the model in new tasks.

[0200] (4) Enhance the diversity of the generated molecular skeleton and overcome library dependence limitations. By generating molecular structures from scratch, DiffPharm overcomes the dependence of traditional virtual screening methods on existing compound libraries. It can generate candidate molecules with novel structures and rich skeletons while maintaining pharmacophore constraints, thus expanding the chemical space coverage capability.

[0201] (5) Unify the adaptation of ligand-guided and structure-guided scenarios to improve the scope of application. The model in this application can be applied to ligand-based pharmacophore models (such as known active molecules) and structure-guided pharmacophores (such as pharmacophores constructed from protein binding pocket information), and has good adaptability in both LBDD and SBDD.

[0202] (6) Significantly reduce the workload of later screening and improve screening efficiency. Since the generated molecules already meet multiple conditions such as the spatial distribution and structural rationality of the pharmacophore, the burden of scoring and filtering in the screening process is significantly reduced, which can increase the molecular screening throughput without sacrificing quality and significantly accelerate the drug discovery process.

[0203] (7) It has modularity and scalability, which facilitates the subsequent integration of more guiding dimensions. The system structure is reasonably designed and the modules are well decoupled. In addition to pharmacophores, it can also be expanded to include guiding factors such as properties, charges, and hydrophobicity. It has good scalability and integration and can serve as the basic architecture for future multi-attribute molecular design platforms.

[0204] This application embodiment also provides a three-dimensional molecular generation apparatus that can implement the above-described method. The apparatus includes:

[0205] The data acquisition module is used to acquire datasets of drug molecules;

[0206] The data construction module is used to build three-dimensional molecular maps and pharmacophore maps for each drug molecule based on the dataset;

[0207] The model training module is used to optimize and train a preset model based on a 3D molecular diagram and a pharmacophore diagram to obtain the target model;

[0208] The preset model includes a pre-trained uncontrolled molecule generation model and a pharmacophore control module;

[0209] The model application module is used to generate three-dimensional molecules of the target drug effect using the target model.

[0210] In some embodiments, the apparatus further includes a policy configuration module for performing the following operations:

[0211] Pharmacophore control strategy based on progressive control strategy configuration of pharmacophore control module;

[0212] The pharmacophore control strategy includes unfuzzy level, fuzzy connection information, centroid atoms, fuzzy partial element level, and fuzzy bond level. The denoising process of the uncontrolled molecular generation model adopts an encoding and decoding architecture. The pharmacophore control module is used to input pharmacophore control information in the decoding stage of the molecular generation module based on the three-dimensional molecular diagram and the pharmacophore diagram through the pharmacophore control strategy.

[0213] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0214] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0215] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0216] like Figure 10 As shown, Figure 10 The hardware structure of an electronic device 1000 according to another embodiment is illustrated. The electronic device 1000 includes:

[0217] The processor 1001 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (aSIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0218] The memory 1002 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RaM). The memory 1002 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1002 and is called and executed by the processor 1001.

[0219] Input / output interface 1003 is used to implement information input and output;

[0220] The communication interface 1004 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0221] Bus 1005 transmits information between various components of the device (e.g., processor 1001, memory 1002, input / output interface 1003, and communication interface 1004);

[0222] The processor 1001, memory 1002, input / output interface 1003 and communication interface 1004 are connected to each other within the device via bus 1005.

[0223] The electronic device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0224] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0225] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0226] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0227] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0228] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0229] The three-dimensional molecule generation method, apparatus, electronic device, storage medium, and program product provided in this application embodiment acquire a dataset of drug molecules; construct a three-dimensional molecular map and pharmacophore map for each drug molecule based on the dataset; optimize and train a preset model based on the three-dimensional molecular map and pharmacophore map to obtain a target model; wherein, the preset model includes a pre-trained uncontrolled molecule generation model and a pharmacophore control module; and use the target model to generate a three-dimensional molecule with the target drug effect. This application embodiment directly generates novel three-dimensional molecules that meet the target drug effect through the pre-trained model and pharmacophore control module, breaking through the chemical space limitations of compound libraries. Specifically, it breaks through the dependence of traditional LBDD on existing compound libraries and avoids the limitation of skeletal novelty caused by insufficient chemical space coverage. In addition, this application embodiment ensures that the generated molecule simultaneously satisfies spatial conformation and functional group matching by fusing pharmacophore maps (functional constraints) and three-dimensional molecular maps (structural constraints). The model-based generation technology solution of this application embodiment avoids the need for one-by-one screening of massive amounts of compounds, significantly reducing computational resource consumption and time costs. Compared with existing pharmacophore-based molecular screening or generation methods, this invention has the following advantages: (1) High efficiency control capability: By freezing the pre-trained molecular generation module, only the pharmacophore control module needs to be trained, which can achieve effective target pharmacophore constraints without sacrificing the model's generalization ability; (2) Flexible control granularity: It supports five pharmacophore control granularities, covering a variety of task requirements from strict alignment to fuzzy constraints, and improving the structural diversity of generated molecules; (3) Support for substructure preservation design: By introducing a masking mechanism, it supports conditional generation guided by substructures such as scaffold hopping and fragment linking; (4) Strong practicality: Combining pharmacophore matching and structural rationality, the generated molecules perform well in docking scoring and pharmacophore matching scoring, and have the potential for practical drug development applications.

[0230] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0231] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0232] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0233] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0234] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0235] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0236] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0237] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0238] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0239] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0240] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for generating three-dimensional molecules, characterized in that, The method includes the following steps: Obtain a dataset of drug molecules; Based on the dataset, construct a three-dimensional molecular map and pharmacophore map for each drug molecule; The preset model is optimized and trained based on the three-dimensional molecular diagram and the pharmacophore diagram to obtain the target model; The preset model includes a pre-trained uncontrolled molecule generation model and a pharmacophore control module; the optimization training of the preset model based on the three-dimensional molecular map and the pharmacophore map to obtain the target model includes the following steps: Based on the aforementioned three-dimensional molecular map, the uncontrolled molecular generation model is pre-trained through diffusion and denoising processes; The pre-trained uncontrolled molecular generation model is parameter-frozen in the structure of the denoising process as a molecular generation module; Based on the three-dimensional molecular map and the pharmacophore map, the pharmacophore control module is trained in conjunction with the molecular generation module; The target model is obtained based on the pre-trained uncontrolled molecule generation model and the trained pharmacophore control module; The pre-training of the uncontrolled molecule generation model based on the three-dimensional molecular map through diffusion and denoising processes includes the following steps: Input the three-dimensional molecular graph into the uncontrolled molecular generation model; During the diffusion process, different properties of the three-dimensional molecular map are subjected to independent perturbation and noise addition processing. In the denoising process, based on the result of the noise addition process, a predicted molecular map is output using an encoding / decoding architecture; A loss function is constructed based on the three-dimensional molecular map and the predicted molecular map, and the parameters of the uncontrolled molecular generation model are adjusted using the loss function. The target model is used to generate three-dimensional molecules for the target drug effect.

2. The method according to claim 1, characterized in that, The construction of a three-dimensional molecular map and pharmacophore map for each drug molecule based on the dataset includes the following steps: The three-dimensional molecular map is constructed based on the molecular information of the drug molecule; The three-dimensional molecular diagram includes an atomic feature matrix, a formal charge matrix, an edge feature matrix, and an atomic coordinate matrix. The atomic feature matrix is ​​obtained by uniquely encoding the chemical type of each atom in the drug molecule in the molecular information. The pharmacophore information of the drug molecule is identified using a pre-set chemical information learning tool, and the pharmacophore map is constructed based on the pharmacophore information. The pharmacophore diagram includes pharmacophore attributes, formal charge control matrix, pharmacophore connection control information, and atom position control information. The pharmacophore attributes are obtained by one-heat encoding based on the pharmacophore category of each atom of the drug molecule in the pharmacophore information.

3. The method according to claim 1, characterized in that, The noise addition process, which involves independently perturbing different properties of the three-dimensional molecular map, includes the following steps: Apply Gaussian noise with zero centroid subspace to continuous properties; Discrete diffusion is performed on discrete properties.

4. The method according to claim 1, characterized in that, The encoding / decoding architecture includes a first fully connected layer, an encoder, a decoder, and a second fully connected layer connected sequentially. The encoder and the decoder include equivalent multi-level isomorphic graph Transformer modules. The isomorphic graph Transformer module includes a graph attention layer, a feedforward neural network, and a Dropout regularization layer. Both the 3D molecular graph and the result of the noise addition process include atomic feature matrices corresponding to node features, edge feature matrices corresponding to edge features, and atomic coordinate matrices corresponding to node coordinates. In the step of outputting the predicted molecular graph using the encoding / decoding architecture, the graph attention layer specifically performs the following operations: Based on the node coordinates, geometric information between any two nodes is obtained by using a spatial extraction layer. Based on the geometric information, the node features, and the global features, the edge features are updated using an edge information update layer; Based on the geometric information, the edge features, and the global features, the node features are updated using a node information update layer with a self-attention mechanism; The node information update layer includes a PNA layer, which is used to aggregate the paired edge features and geometric information into a node representation. Based on the geometric information and the edge features, the node coordinates are updated using the rEGNN update mechanism in the position information EGNN update layer; Based on the geometric information, the edge features, and the node features, a global feature update layer is used to perform graph-level aggregation in order to update the global features.

5. The method according to claim 1, characterized in that, Both the 3D molecular graph and the predicted molecular graph include atomic feature matrices corresponding to atom types, formal charge matrices corresponding to formal charges, edge feature matrices corresponding to bond types, and atomic coordinate matrices corresponding to atomic coordinates. The step of constructing a loss function based on the 3D molecular graph and the predicted molecular graph includes the following steps: Based on the actual values ​​of the atomic coordinates in the three-dimensional molecular diagram and the predicted values ​​of the atomic coordinates in the predicted molecular diagram, a first loss is constructed using mean square error; Based on the actual values ​​of the atom types in the three-dimensional molecular diagram and the predicted values ​​of the atom types in the predicted molecular diagram, a second loss is constructed using cross-entropy loss. Based on the actual value of the formal charge in the three-dimensional molecular diagram and the predicted value of the formal charge in the predicted molecular diagram, a third loss is constructed using cross-entropy loss; Based on the actual values ​​of the bond types in the three-dimensional molecular diagram and the predicted values ​​of the bond types in the predicted molecular diagram, a fourth loss is constructed using cross-entropy loss. The first loss, the second loss, the third loss, and the fourth loss are weighted and summed to obtain the loss function; The expression for the loss function is as follows: In the formula, Represents the loss function; , , and Indicates the weight of each loss; Represents cross-entropy loss; Indicates mean square error; , , and These represent the actual values ​​of atom type, formal charge, bond type, and atom coordinates, respectively. , , and The predicted values ​​are represented in order by atom type, formal charge, bond type, and atom coordinates.

6. The method according to claim 1, characterized in that, Before the step of training the pharmacophore control module, the method further includes the following steps: The pharmacophore control strategy of the pharmacophore control module is configured based on a progressive control strategy; The pharmacophore control strategy includes unambiguous level, fuzzy connection information, centroidal atoms, fuzzy partial element level, and fuzzy bond level. The denoising process of the uncontrolled molecular generation model adopts an encoding / decoding architecture. The pharmacophore control module is used to input pharmacophore control information in the decoding stage of the molecular generation module based on the three-dimensional molecular diagram and the pharmacophore diagram, through the pharmacophore control strategy.

7. The method according to claim 1, characterized in that, The process of generating three-dimensional molecules for the target drug effect using the target model includes the following steps: Obtain the pharmacophore model corresponding to the reference molecule; The pharmacophore model and the sampled noise molecule data are converted into a graph structure and input into the target model. The reference molecule is denoised using the pre-trained uncontrolled molecule generation model to generate a new molecule; wherein the denoising process includes an encoding stage and a decoding stage, and the pharmacophore control module is used to input pharmacophore control information in the decoding stage based on the three-dimensional molecular map and the pharmacophore map to guide the generation of the new molecule through a preset pharmacophore control strategy; Candidate molecules are obtained from the new molecules based on matching scores and screening scores.

8. The method according to claim 7, characterized in that, The method further includes the following steps: In response to the retention instruction of the target object, a portion of atoms in the reference molecule are marked as substructures by a mask; The substructure representation must retain the structure, and the remaining parts other than the substructure are regions to be generated; Obtain intermediate states from a forward-noised distribution; The first region corresponding to the substructure is extracted from the intermediate state based on the mask; The region is generated by sampling from the denoising process; The second region corresponding to the region to be generated is extracted from the generated region based on the inversion code of the mask; The first region is fused with the second region to generate a new molecule with a fixed substructure.

9. A three-dimensional molecular generation device, characterized in that, The device includes: The data acquisition module is used to acquire datasets of drug molecules; A data construction module is used to construct a three-dimensional molecular map and a pharmacophore map for each drug molecule based on the dataset. The model training module is used to optimize and train the preset model based on the three-dimensional molecular diagram and the pharmacophore diagram to obtain the target model; The preset model includes a pre-trained uncontrolled molecule generation model and a pharmacophore control module; the optimization training of the preset model based on the three-dimensional molecular map and the pharmacophore map to obtain the target model includes the following steps: Based on the aforementioned three-dimensional molecular map, the uncontrolled molecular generation model is pre-trained through diffusion and denoising processes; The pre-trained uncontrolled molecular generation model is parameter-frozen in the structure of the denoising process as a molecular generation module; Based on the three-dimensional molecular map and the pharmacophore map, the pharmacophore control module is trained in conjunction with the molecular generation module; The target model is obtained based on the pre-trained uncontrolled molecule generation model and the trained pharmacophore control module; The pre-training of the uncontrolled molecule generation model based on the three-dimensional molecular map through diffusion and denoising processes includes the following steps: Input the three-dimensional molecular graph into the uncontrolled molecular generation model; During the diffusion process, different properties of the three-dimensional molecular map are subjected to independent perturbation and noise addition processing. In the denoising process, based on the result of the noise addition process, a predicted molecular map is output using an encoding / decoding architecture; A loss function is constructed based on the three-dimensional molecular map and the predicted molecular map, and the parameters of the uncontrolled molecular generation model are adjusted using the loss function. The model application module is used to generate three-dimensional molecules of the target drug effect using the target model.

10. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 8.

11. A computer storage medium storing a processor-executable program, which, when executed by a processor, is used to implement the method as described in any one of claims 1 to 8.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 8.

Images