Ligand molecule generation method based on multi-modal information guidance
The method of generating ligand molecules guided by multimodal information solves the problem of lack of diversity and interaction understanding in the protein target generated molecules in the prior art. The generated molecules show high precision and diversity in biological docking experiments, realizing efficient interaction between proteins and ligand molecules.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-15
- Publication Date
- 2026-03-27
AI Technical Summary
Existing structure-based molecular generation methods neglect the complexity and diversity of protein targets, making it difficult to fully capture the high-precision interactions between protein pockets and ligand molecules. This results in a lack of diversity in the generated molecules, making it difficult to meet the needs of biological docking experiments. Furthermore, they ignore the potential interaction relationships between protein structural features and ligand molecules.
A ligand molecule generation method based on multimodal information is adopted. By acquiring multiple feature information of proteins and ligand molecules, feature interaction is performed using the Sentence-BERT model and cross-attention mechanism, combined with gating mechanism and noisy network, to generate diverse ligand molecules and fully capture the interaction between proteins and ligands.
The generated ligand molecules exhibit greater diversity and precision in biological docking experiments, effectively meeting the needs of subsequent verification, and achieving an excellent balance between binding affinity and molecular properties.
Smart Images
Figure CN121747686A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of protein ligand molecule generation technology, and in particular to a method for ligand molecule generation based on multimodal information guidance. Background Technology
[0002] With the continuous development of modern drug development and molecular design technologies, computer-aided molecular generation methods based on protein structure information have gradually become an important means of new drug discovery. Traditional molecular screening relies on large-scale experimental methods, which are not only time-consuming and costly, but also have low efficiency in screening highly selective and highly active molecules. Therefore, researchers are gradually turning to artificial intelligence models to predict and model the interaction between protein pockets and ligand molecules, in order to rapidly generate candidate molecules in a computer environment and shorten the drug development cycle.
[0003] Most existing structure-based molecular generation methods employ rule-driven or simple generation strategies. While they can generate molecules that meet basic chemical constraints to some extent, they often overlook the complexity and diversity of protein targets, making it difficult to fully capture the high-precision interactions between protein pockets and ligand molecules. This limits their application value in ligand molecule generation tasks.
[0004] In recent years, diffusion models have received widespread attention in molecular generation tasks. The generation process comprises two stages: forward diffusion and backward reconstruction. The former transforms the data into a simple distribution by progressively adding noise, while the latter learns to reconstruct the original data from the noise. This mechanism enables the model to accurately model complex underlying data distributions and generate high-quality molecular structures with high diversity. However, current research still has significant limitations: on the one hand, some methods for generating molecules targeting protein targets only consider the structural information of ligand molecules, lacking sufficient information guidance, resulting in a lack of diversity in the generated molecules and making it difficult to meet the needs of subsequent biological docking experiments for verification; on the other hand, existing methods neglect the potential interaction relationships between protein structural features and ligand molecules. This lack of modal information limits the generative model's deep understanding of the interaction relationship between the target and the molecule; furthermore, how to integrate and balance various modal information in diffusion models to guide molecular generation for different protein targets is an urgent problem to be solved. Summary of the Invention
[0005] Therefore, it is necessary to propose a ligand molecule generation method based on multimodal information guidance to address the above problems.
[0006] A method for generating ligand molecules based on multimodal information guidance, the method comprising:
[0007] S1: Obtain proteins and multiple ligand molecules from the CrossDocked2020 dataset, and add noise to the ligand atom three-dimensional coordinate features and ligand atom type features of the ligand molecules respectively to obtain noisy ligand atom three-dimensional coordinate features and noisy ligand atom type features. The noisy ligand atom three-dimensional coordinate features and noisy ligand atom type features constitute pure noise ligand molecules.
[0008] S2: Obtain the protein text information of the protein and the ligand molecule text information of the multiple ligand molecules; map the protein text information and ligand molecule text information to a d-dimensional vector using the Sentence-BERT model to obtain the protein text information features and ligand molecule text information features.
[0009] S3: Determine the characteristics of protein-protein invariant interactions and ligand-molecule invariant interactions;
[0010] S4: Using the ligand atom type features of ligand molecules at different diffusion steps as query vectors, and using the text information features of the ligand molecules and the invariant interaction features of the ligand molecules as different bonds and values, a cross-attention mechanism is applied to perform feature interaction to obtain the first updated ligand atom type features and the second updated ligand atom type features respectively.
[0011] Using the ligand atom type features of ligand molecules at different diffusion steps as query vectors, and the protein text information features and the protein invariance interaction features as different bonds and values, a cross-attention mechanism is applied to perform feature interaction to obtain the first updated protein atom type features and the second updated protein atom type features respectively.
[0012] S5: By fusing the first updated ligand atom type features and the second updated ligand atom type features respectively through a gating mechanism, the first target ligand atom type features and the second target ligand atom type features of the ligand molecule under different diffusion steps are obtained;
[0013] By fusing the first updated protein atom type features and the second updated protein atom type features respectively using a gating mechanism, the first target protein atom type features and the second target protein atom type features of the protein at different diffusion steps are obtained. ;
[0014] S6: Based on the first target ligand atom type characteristics and the second target ligand atom type characteristics, determine the ligand molecule target type characteristics of the ligand molecules under different diffusion steps, and determine the second updated ligand molecule;
[0015] Based on the first target protein atom type characteristics and the second target protein atom type characteristics, the protein target type characteristics of the protein are determined, and the second updated protein is determined.
[0016] S7: The second updated ligand molecule and the second updated protein constitute a protein-ligand complex. The target ligand molecule is obtained by manipulating the protein-ligand complex through a noise network. This invention fully captures the high-precision interaction between the protein pocket and the ligand molecule. The method for generating molecules targeting protein targets not only considers the structural information of the ligand molecule but also provides sufficient information guidance, enabling the generated molecules to possess diversity and meet the needs of subsequent biological docking experiments. On the other hand, by utilizing the potential interaction relationship between protein structural features and ligand molecules, for different protein targets, the generation of ligand molecules is guided by fusing and balancing various modal information during diffusion. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] in:
[0019] Figure 1 This is a flowchart of a ligand molecule generation method guided by multimodal information in one embodiment;
[0020] Figure 2 This is a structural block diagram of a computer device in one embodiment. Detailed Implementation
[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] Most existing structure-based molecular generation methods employ rule-driven or simple generation strategies. While they can generate molecules that meet basic chemical constraints to some extent, they often neglect the complexity and diversity of protein targets, making it difficult to fully capture the high-precision interactions between protein pockets and ligand molecules. This limits their application value in ligand molecule generation tasks. In recent years, diffusion models have received widespread attention in molecular generation tasks. Their generation process includes two stages: forward diffusion and backward reconstruction. The former transforms the data into a simple distribution by gradually adding noise, while the latter learns to reconstruct the original data from the noise. This mechanism enables the model to accurately model complex underlying data distributions and generate high-quality molecular structures with high diversity. However, current research still has significant limitations: on the one hand, some methods for generating molecules targeting protein targets only consider the structural information of ligand molecules, lacking sufficient information guidance, resulting in a lack of diversity in the generated molecules, making it difficult to meet the needs of subsequent biological docking experiments for verification; on the other hand, existing methods ignore the potential interaction relationship between protein structural features and ligand molecules, and the lack of such modal information limits the generation model's in-depth understanding of the interaction relationship between the target and the molecule; in addition, how to integrate and balance various modal information in diffusion models to guide molecule generation for different protein targets is an urgent problem to be solved.
[0023] To address the aforementioned technical problems, this application provides a method for generating ligand molecules based on multimodal information guidance.
[0024] Proteins and their bound ligand molecules were obtained from the publicly available dataset CrossDocked2020; proteins are composed of multiple protein atoms, and ligand molecules are composed of multiple ligand atoms.
[0025] Proteins can be represented as , = , ,in, and ;in, It is a protein; This is a set of three-dimensional coordinate features of protein atoms, consisting of the three-dimensional coordinate features of multiple atoms of protein P. For protein A set of protein atom type features consisting of multiple atom types of protein atoms; For protein The three-dimensional coordinates of atom i; For protein The type of atom i; and , For protein The number of atoms contained For protein The one-hot encoding dimension of the atomic type features;
[0026] The ligand molecule is represented as , = , , and ;in, It is a ligand molecule; ligand molecules The set of three-dimensional coordinate features of ligand atoms consisting of the set of three-dimensional coordinate features of multiple atoms; ligand molecules The set of ligand atom type features composed of the ligand atom type features of multiple atoms; ligand molecules 3D coordinates; ligand molecules The type of atom i; and , ligand molecules The number of atoms contained ligand molecules The one-hot encoding dimension of atomic type features.
[0027] like Figure 1 As shown, in one embodiment, a method for generating ligand molecules based on multimodal information is provided, specifically including the following steps:
[0028] S1: Obtain proteins and multiple ligand molecules from the CrossDocked2020 dataset, and add noise to the ligand atom three-dimensional coordinate features and ligand atom type features of the ligand molecules respectively to obtain noisy ligand atom three-dimensional coordinate features and noisy ligand atom type features. The noisy ligand atom three-dimensional coordinate features and noisy ligand atom type features constitute pure noise ligand molecules.
[0029] S2: Obtain the protein text information of the protein and the ligand molecule text information of the multiple ligand molecules; map the protein text information and ligand molecule text information to a d-dimensional vector using the Sentence-BERT model to obtain the protein text information features and ligand molecule text information features.
[0030] S3: Determine the characteristics of protein-protein invariant interactions and ligand-molecule invariant interactions;
[0031] S4: Using the ligand atom type features of ligand molecules at different diffusion steps as query vectors, and using the text information features of the ligand molecules and the invariant interaction features of the ligand molecules as different bonds and values, a cross-attention mechanism is applied to perform feature interaction to obtain the first updated ligand atom type features and the second updated ligand atom type features respectively.
[0032] Using the ligand atom type features of ligand molecules at different diffusion steps as query vectors, and the protein text information features and the protein invariance interaction features as different bonds and values, a cross-attention mechanism is applied to perform feature interaction to obtain the first updated protein atom type features and the second updated protein atom type features respectively.
[0033] S5: By fusing the first updated ligand atom type features and the second updated ligand atom type features respectively through a gating mechanism, the first target ligand atom type features and the second target ligand atom type features of the ligand molecule under different diffusion steps are obtained;
[0034] By fusing the first updated protein atom type features and the second updated protein atom type features respectively using a gating mechanism, the first target protein atom type features and the second target protein atom type features of the protein at different diffusion steps are obtained. ;
[0035] S6: Based on the first target ligand atom type characteristics and the second target ligand atom type characteristics, determine the ligand molecule target type characteristics of the ligand molecules under different diffusion steps, and determine the second updated ligand molecule;
[0036] Based on the first target protein atom type characteristics and the second target protein atom type characteristics, the protein target type characteristics of the protein are determined, and the second updated protein is determined.
[0037] S7: The second updated ligand molecule and the second updated protein constitute a protein-ligand complex, and the target ligand molecule is obtained by manipulating the protein-ligand complex through a noise network.
[0038] In one embodiment, the determination of protein-protein invariant interaction characteristics Interaction characteristics with ligand molecule invariance include:
[0039] S3.1: Extracting the proteins using a protein-ligand interaction network. The updated protein atom invariance characteristics are obtained. The ligand molecule is updated to obtain the ligand atom invariance feature. Each layer of the protein-ligand molecule interaction extraction network is an isovariant graph neural network;
[0040] S3.2: Invariant characteristics of the protein atoms Normalization and activation operations were performed sequentially to obtain the ligand atom invariance interaction characteristics. Invariance characteristics of the ligand atoms Normalization and activation operations were performed sequentially to obtain the protein atom-invariant interaction characteristics. Invariant interaction characteristics of multiple protein atoms Features of protein-protein invariance interactions Invariant interaction characteristics of multiple ligand atoms Characterized by the invariance of interactions between ligand molecules .
[0041] In one embodiment, the ligand molecule The ligand atom three-dimensional coordinate features and ligand atom type features are respectively subjected to noise addition processing, including:
[0042] Gradually towards ligand molecules = Gaussian noise and uniform noise were added until the ligand molecule was completely converted into pure noise. This process can be represented as:
[0043] (1)
[0044] in, The number of diffusion steps is The ligand molecule at that time; The number of diffusion steps is Time ligand molecules; For proteins in the dataset; Indicates a Gaussian distribution. Represents category distribution; The number of diffusion steps is The three-dimensional coordinate characteristics of ligand atoms in a ligand molecule at that time; for the three-dimensional coordinate characteristics of ligand atoms in a ligand molecule, Using sigmoid Scheduling scheme, and Regarding the characteristics of ligand atom types in ligand molecules, Using cosine Scheduling scheme ( ); For variance; The number of diffusion steps is Characteristics of ligand atom types in ligand molecules at that time; The one-hot encoding dimension for the ligand atom type characteristics of ligand molecules; , This is the total number of diffusion steps, a hyperparameter used in this scheme. Set to 1000; when for When the noise addition stops, and each addition of noise satisfies the Markov property, a pure noise ligand molecule is obtained. This leads to the acquisition of pure noise ligand molecules. Noisy ligand atom three-dimensional coordinate features and noise-adding processing, noise-adding ligand atom type characteristics ,Right now .
[0045] In one embodiment, the protein text information features and ligand molecule textual information features Obtained through the following methods:
[0046] The large language model GPT-4o is used to generate text guidance; protein is input into GPT-4o. Name: ,protein PDB ID: That is, the unique identifier of a protein structure database, protein Chain ID: That is, the specific binding site number on the protein molecule, the protein Pocket ID: That is, the identifiers of polypeptide chains and ligand molecules in proteins. The SMILES string: And a predefined prompt_t, from which GPT-4o generates text information, including:
[0047] protein Textual information about the structure or properties of a specific binding site:
[0048] (2)
[0049] in, This is the protein text information for a specific binding site of protein P; prompt_t is a predefined prompt.
[0050] ligand molecules Textual information about the structure or properties of ligand molecules:
[0051] (3)
[0052] in, Indicates ligand molecules The ligand molecule text information; prompt_t is a predefined prompt;
[0053] The Sentence-BERT model uses protein text information and ligand molecule text information Mapping to a d-dimensional vector yields protein textual information features. and ligand molecule textual information features .
[0054] In one embodiment, the protein atom invariance feature and the invariance characteristics of the ligand atoms Obtained through the following methods:
[0055] The protein-ligand interaction extraction network consists of multiple layers, each of which is an isovariant graph neural network that independently captures proteins. and ligand molecules at different diffusion steps Geometric structure representation; for ligand molecules with different diffusion steps Construct a fully connected graph , No. The layer updates the ligand molecules at different diffusion steps in the following ways. The invariant and isovariant characteristics of each atom in the material:
[0056] (4)
[0057] (5)
[0058] (6)
[0059] in, Ligand molecules at different diffusion steps The atoms of the ligands In the Invariant features of layered graph neural networks; Represents ligand molecules at different diffusion steps atoms In the Invariant features of layered graph neural networks; Ligand molecules at different diffusion steps Atoms of ligand molecules In the Invariant features of layered graph neural networks; Represents ligand molecules at different diffusion steps Atoms of ligand molecules In the Invariant features of layered graph neural networks; Represents invariant function operations, Atoms representing ligand molecules and atoms of ligand molecules The first Euclidean distance between them; Ligand molecules at different diffusion steps atoms In the Isovariability characteristics of layered isovariable graphical neural networks; Ligand molecules at different diffusion steps atoms In the Isovariability characteristics of layered isovariable graphical neural networks; Ligand molecules at different diffusion steps atoms In the Isovariability characteristics of layered isovariable graphical neural networks; Ligand molecules at different diffusion steps atoms In the Isovariability characteristics of layered isovariable graphical neural networks; Represents isovariant function operations; Represents atoms The set of neighbors;
[0060] implement The next update operation yields the first updated ligand molecule. =[ ];
[0061] For proteins Construct a fully connected graph By performing the same operation, proteins can be obtained. In the Isovariability features of layered equivariant graph neural networks and protein In the Invariant features of layered graph neural networks ;
[0062] implement The second update operation yields the first updated protein. '=[ , ].
[0063] In this way, the first updated ligand molecule is obtained. Invariance of ligand atoms in individual atoms And the first updated protein The invariance of protein atoms at the atomic level. .
[0064] In one embodiment, the protein-protein invariant interaction feature Interaction characteristics with ligand molecule invariance Obtained through the following methods:
[0065] (7)
[0066] (8)
[0067] in, For proteins that involve protein-molecule interactions, the first The invariance of the interaction characteristics of the ligand atom and the atom itself; For the first ligand molecule containing protein-molecule interactions The invariant interaction characteristics of protein atoms at the atomic level; To update the first protein The invariance of proteins at the atomic level; For the first renewal of ligand molecules The invariance characteristics of ligand molecules with individual atoms; This represents the activation function. Atoms for ligands to the atoms of proteins Attention weights; Atoms of proteins Atoms of ligands Attention weights; This is a normalization operation; It is a learnable projection matrix.
[0068] In one embodiment, the first update of the ligand atom type feature Second update of ligand atom type characteristics First update of protein atom type characteristics Second update protein atom type features This can be achieved using the following expression:
[0069] (9)
[0070] (10)
[0071] (11)
[0072] (12)
[0073] in, This indicates the integration of ligand molecule textual information features. Ligand molecules at different diffusion steps The first updated ligand atom type feature; This indicates the integration of ligand molecule invariance interaction characteristics. Ligand molecules at different diffusion steps The second updated ligand atom type feature; , , , , and It is a learnable projection matrix; For feature dimensions; This is a normalization operation; The first update is for the protein atom type characteristics; The first update is the protein atom type feature.
[0074] In one embodiment, for different proteins, the two types of guidance information are not equally important in the process of generating ligand molecules. Overuse of less important guidance information may even reduce the quality of the generated molecule. Therefore, a gating mechanism is employed to enable the model to learn and adaptively adjust the utilization of the two types of guidance information, and the first target ligand atom type feature... Characteristics of the second target ligand atom type First target protein atom type characteristics Second target protein atom type characteristics This can be achieved using the following expression:
[0075] (13)
[0076] (14)
[0077] (15)
[0078] (16)
[0079] in, For features that incorporate ligand molecule textual information through a gating mechanism Ligand molecules at different diffusion steps The first target ligand atom type characteristics, This refers to ligand molecules with different diffusion steps that incorporate protein-molecule interactions through a gating mechanism. The second target ligand atom type characteristics; and This represents the sigmoid activation function used to generate the gated signal; and These are the learnable parameters of the gating unit; Represents the dot product of vectors; This indicates the operation of the self-attention mechanism; The primary target protein atom type characteristics; This refers to the atom type characteristics of the second target protein;
[0080] In one embodiment, the second updated ligand molecule '' and the second updated protein Obtained through the following methods:
[0081] (17)
[0082] (18)
[0083] in, For protein Protein target type characteristics, Ligand molecules at different diffusion steps Ligand molecule target type characteristics; For protein A set of protein atom type features. Ligand molecules at different diffusion steps The set of ligand atom type characteristics; The first target protein atom type characteristics, This refers to the atom type characteristics of the second target protein; ligand molecules atomic type characteristics, Ligand molecules with different diffusion steps for different ligand molecule invariance characteristics Atomic type characteristics; This is a multilayer sensor; at this point, the protein... Updated to the second update protein ''= ; Ligand molecules Updated to the second updated ligand molecule ''= .
[0084] In one embodiment, the second updated protein Obtained through the following methods:
[0085] Through denoising network Protein-ligand complexes The operation yielded a second updated protein. The denoising network For inclusion The protein-ligand complex is composed of an SE(3) isovariant neural network with isovariant layers. The second updated target protein ''= Second updated target ligand molecule ''= Dynamically constructed;
[0086] (19)
[0087] (20)
[0088] in, The third diffusion model Protein-ligand complexes of the layer atoms Type characteristics, The third diffusion model 1 layer of protein-ligand complex atoms Characteristics of protein-ligand complex types; The third diffusion model Protein-ligand complexes of the layer atoms Characteristics of protein-ligand complex types, The third diffusion model protein-ligand complex of the layer atoms Characteristics of protein-ligand complex types; The third diffusion model Protein-ligand complexes of the layer atoms The three-dimensional coordinate features, The third diffusion model Protein-ligand complexes of the layer atoms Three-dimensional coordinate features of protein-ligand complexes; Atoms representing protein-ligand complexes Atoms of protein-ligand complexes The relationship between them, for example, both are protein atoms or ligand atoms, or a protein atom and a ligand atom. It is a ligand molecule mask, which means that only the atomic coordinates of the ligand molecule are updated first, and the atomic coordinates of the protein are not updated first. Indicates protein-ligand complex atoms and protein-ligand complexes atoms The second Euclidean distance between them.
[0089] go through After performing the same operations on each layer, the result is obtained after the denoising network. After one time step update, the atoms of the updated protein-ligand complex are obtained. 3D coordinate features and type features This leads to the acquisition of updated three-dimensional coordinate features of the protein-ligand complex. and complex type characteristics ;
[0090] Using denoising networks Repeat the process for updating the protein-ligand complex obtained at the previous time step. - One operation yields the target protein-ligand complex. = ;in, , ,in, The three-dimensional coordinate features of the protein in the target protein-ligand complex; Three-dimensional coordinate features of ligand molecules in the target protein-ligand complex; The type characteristics of proteins in the target protein-ligand complex; Characterize the type of ligand molecules in the target protein-ligand complex; define the three-dimensional coordinates of the ligand molecules in the target protein-ligand complex. Converted to raw 3D coordinates via MLP layer ; to identify the type characteristics of ligand molecules in the target protein-ligand complex Converted to primitive type via MLP layer The target ligand molecule is composed of the original three-dimensional coordinates. and primitive types constitute.
[0091] When training the diffusion model, the first step of the reverse generation process The loss for reconstructing the atomic 3D coordinates of the step is defined as follows:
[0092]
[0093]
[0094]
[0095]
[0096]
[0097]
[0098] The first step of the reverse generation process The atomic type reconstruction loss for the step is defined as follows:
[0099]
[0100]
[0101]
[0102] The total loss function is achieved through hyperparameters. The combination of weighted atom 3D coordinate reconstruction loss and atom type reconstruction loss, i.e. .
[0103] Training is performed iteratively in a pre-defined number of rounds, with an evaluation every 2000 training steps. Data augmentation is employed during training: Gaussian noise with a standard deviation of 0.1 is added to the three-dimensional coordinates of protein atoms. The learning rate is set to decay exponentially with a coefficient of 0.6, with a minimum value of 1e-6.
[0104] Table 1 analyzes the performance of this invention from two aspects: binding affinity and molecular properties, summarizing the comparison results of key indicators. In most docking-related indicators (Vina Score, Vina Min, and Vina Dock), this invention achieved the best average and median scores, significantly outperforming all non-diffusion methods and improving by approximately 5% compared to other diffusion models. This indicates that this invention can generate molecules with stronger predicted binding potential. Regarding high affinity, this invention leads other benchmark methods with an average of 68.6% and a median of 73.2%, demonstrating that this model effectively captures key protein-ligand interactions. For molecular property indicators such as QED (Quantitative Estimate of Drug-likeness), SA (Synthetic Accessibility Score), and Diversity, this invention maintains a level comparable to other strong benchmark methods. Although some other models achieved the highest scores in QED and SA, they lagged behind in docking indicators, failing to achieve a reasonable trade-off. This invention achieves an excellent balance between binding affinity and reasonable molecular properties by integrating multimodal guidance information.
[0105] Table 1 Comparison Results of Key Indicators
[0106]
[0107] Figure 2 An internal structural diagram of a computer device in one embodiment is shown. This computer device can specifically be a terminal or a server. Figure 2 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a ligand molecule generation method guided by multimodal information. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement a ligand molecule generation method guided by multimodal information. Those skilled in the art will understand that... Figure 2 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0108] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.
[0109] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0110] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.
Claims
1. A method for generating ligand molecules based on multimodal information guidance, characterized in that, The method includes: S1: Obtain proteins and multiple ligand molecules from the CrossDocked2020 dataset, and add noise to the ligand atom three-dimensional coordinate features and ligand atom type features of the ligand molecules respectively to obtain noisy ligand atom three-dimensional coordinate features and noisy ligand atom type features. The noisy ligand atom three-dimensional coordinate features and noisy ligand atom type features constitute pure noise ligand molecules. S2: Obtain the protein text information of the protein and the ligand molecule text information of the multiple ligand molecules; map the protein text information and ligand molecule text information to a d-dimensional vector using the Sentence-BERT model to obtain the protein text information features and ligand molecule text information features. S3: Determine the characteristics of protein-protein invariant interactions and ligand-molecule invariant interactions; S4: Using the ligand atom type features of ligand molecules at different diffusion steps as query vectors, and using the text information features of the ligand molecules and the invariant interaction features of the ligand molecules as different bonds and values, a cross-attention mechanism is applied to perform feature interaction to obtain the first updated ligand atom type features and the second updated ligand atom type features respectively. Using the ligand atom type features of ligand molecules at different diffusion steps as query vectors, and the protein text information features and the protein invariance interaction features as different bonds and values, a cross-attention mechanism is applied to perform feature interaction to obtain the first updated protein atom type features and the second updated protein atom type features respectively. S5: By fusing the first updated ligand atom type features and the second updated ligand atom type features respectively through a gating mechanism, the first target ligand atom type features and the second target ligand atom type features of the ligand molecule under different diffusion steps are obtained; By fusing the first updated protein atom type features and the second updated protein atom type features respectively using a gating mechanism, the first target protein atom type features and the second target protein atom type features of the protein at different diffusion steps are obtained. ; S6: Based on the first target ligand atom type characteristics and the second target ligand atom type characteristics, determine the ligand molecule target type characteristics of the ligand molecules under different diffusion steps, and determine the second updated ligand molecule; Based on the first target protein atom type characteristics and the second target protein atom type characteristics, the protein target type characteristics of the protein are determined, and the second updated protein is determined. S7: The second updated ligand molecule and the second updated protein constitute a protein-ligand complex, and the target ligand molecule is obtained by manipulating the protein-ligand complex through a noise network.
2. The method for generating ligand molecules based on multimodal information guidance according to claim 1, characterized in that, The determination of protein-invariant interaction characteristics and ligand-molecule-invariant interaction characteristics includes: S3.1: The protein atom invariance feature is obtained by updating the protein through the protein-ligand molecule interaction extraction network, and the ligand atom invariance feature is obtained by updating the ligand molecule. Each layer of the protein-ligand molecule interaction extraction network is an isovariant graph neural network. S3.2: Normalize and activate the protein atom invariance features sequentially to obtain ligand atom invariance interaction features; normalize and activate the ligand atom invariance features sequentially to obtain protein atom invariance interaction features; multiple protein atom invariance interaction features constitute protein invariance interaction features; multiple ligand atom invariance interaction features constitute ligand molecule invariance interaction features.
3. The method for generating ligand molecules based on multimodal information guidance according to claim 1, characterized in that, The noise-adding processing of the three-dimensional coordinate features and ligand atom type features of the ligand molecule includes: Gradually adding Gaussian noise and uniform noise to the ligand molecule until it is completely transformed into a pure noise ligand molecule can be represented as: in, The number of diffusion steps is The ligand molecule at that time; The number of diffusion steps is Time ligand molecules; For proteins in the dataset; Indicates a Gaussian distribution. Represents category distribution; The number of diffusion steps is The three-dimensional coordinate characteristics of ligand atoms in a ligand molecule at that time; for the three-dimensional coordinate characteristics of ligand atoms in a ligand molecule, Using sigmoid Scheduling scheme, and Regarding the characteristics of ligand atom types in ligand molecules, Using cosine Scheduling scheme ( ); For variance; The number of diffusion steps is Characteristics of ligand atom types in ligand molecules at that time; The one-hot encoding dimension for the ligand atom type characteristics of ligand molecules; , This is the total number of diffusion steps, a hyperparameter used in this scheme. Set to 1000; when for At this point, noise addition stops, and each addition of noise satisfies the Markov property, resulting in a pure noise ligand molecule. This leads to the acquisition of the three-dimensional coordinate characteristics of the noise-added ligand atoms in the pure noise ligand molecule and the characteristics of the noise-added ligand atom types after noise addition. ,Right now .
4. The method for generating ligand molecules based on multimodal information guidance according to claim 1, characterized in that, The protein textual information features and ligand molecule textual information features are obtained in the following manner: The large language model GPT-4o is used to generate text guidance; the protein name is input into GPT-4o. Protein PDBID: That is, the unique identifier of the protein structure database, the chain ID of the protein: This refers to the specific binding site number on the protein molecule, or the protein's pocket ID. This refers to the identifier of the polypeptide chain in the protein and the SMILES string of the ligand molecule: And a predefined prompt_t, from which GPT-4o generates text information, including: Textual information about the structure or properties of a specific binding site of a protein: in, This provides textual information about a specific protein binding site; `prompt_t` is a predefined prompt. Textual information about the structure or properties of ligand molecules: in, The text information represents the ligand molecule; prompt_t is a predefined prompt. The Sentence-BERT model maps protein text information and ligand molecule text information to d-dimensional vectors to obtain protein text information features and ligand molecule text information features.
5. The method for generating ligand molecules based on multimodal information guidance according to claim 2, characterized in that, The protein atom invariance characteristics and the ligand atom invariance characteristics are obtained in the following manner: The protein-ligand interaction extraction network consists of multiple layers, each being an isotropic graph neural network that independently captures the geometric representations of proteins and ligand molecules at different diffusion steps. For ligand molecules at different diffusion steps, a fully connected graph is constructed. The layer updates the invariant and isovariant characteristics of each atom in the ligand molecule at different diffusion steps in the following manner: in, Ligand molecules at different diffusion steps The atoms of the ligands In the Invariant features of layered graph neural networks; Represents ligand molecules at different diffusion steps atoms In the Invariant features of layered graph neural networks; Ligand molecules at different diffusion steps Atoms of ligand molecules In the Invariant features of layered graph neural networks; Represents ligand molecules at different diffusion steps Atoms of ligand molecules In the Invariant features of layered graph neural networks; Represents invariant function operations, Atoms representing ligand molecules and atoms of ligand molecules The first Euclidean distance between them; Ligand molecules at different diffusion steps atoms In the Isovariability characteristics of layered isovariable graphical neural networks; Ligand molecules at different diffusion steps atoms In the Isovariability characteristics of layered isovariable graphical neural networks; Ligand molecules at different diffusion steps atoms In the Isovariability characteristics of layered isovariable graphical neural networks; Ligand molecules at different diffusion steps atoms In the Isovariability characteristics of layered isovariable graphical neural networks; Represents isovariant function operations; Represents atoms The set of neighbors; implement The next update operation yields the first updated ligand molecule; For proteins, construct fully connected graphs. Performing the same operation, the protein can be obtained in the first stage. The isovariability features of layered isovariant graph neural networks and proteins in the first layer Invariant features of layered graph neural networks; implement The second update operation yields the first updated protein; In this way, the first updated ligand molecule is obtained. The invariance of ligand atoms and the first update of the protein. The characteristic of protein atoms being invariable.
6. The method for generating ligand molecules based on multimodal information guidance according to claim 5, characterized in that, The protein-invariant interaction characteristics and ligand-molecule-invariant interaction characteristics are obtained in the following manner: in, For proteins that involve protein-molecule interactions, the first The invariance of the interaction characteristics of the ligand atom and the atom itself; For the first ligand molecule containing protein-molecule interactions The invariant interaction characteristics of protein atoms at the atomic level; To update the first protein The invariance of proteins at the atomic level; For the first renewal of ligand molecules The invariance of ligand molecules with individual atoms; This represents the activation function. Atoms for ligands to the atoms of proteins Attention weights; Atoms of proteins Atoms of ligands Attention weights; This is a normalization operation; It is a learnable projection matrix.
7. The method for generating ligand molecules based on multimodal information guidance according to claim 1, characterized in that, The first updated ligand atom type feature, the second updated ligand atom type feature, the first updated protein atom type feature, and the second updated protein atom type feature are implemented through the following expressions: in, This indicates the integration of ligand molecule textual information features. Characteristics of the first updated ligand atom type of ligand molecules at different diffusion steps; This indicates the integration of ligand molecule invariance interaction characteristics. Characteristics of the second updated ligand atom types in ligand molecules at different diffusion steps; , , , , and It is a learnable projection matrix; For feature dimensions; This is a normalization operation; The first update is for the protein atom type characteristics; The first update is the protein atom type feature.
8. The method for generating ligand molecules based on multimodal information guidance according to claim 5, characterized in that, The first target ligand atom type feature, the second target ligand atom type feature, the first target protein atom type feature, and the second target protein atom type feature are implemented by the following expressions: in, For features that incorporate ligand molecule textual information through a gating mechanism Ligand molecules at different diffusion steps The first target ligand atom type characteristics, This refers to ligand molecules with different diffusion steps that incorporate protein-molecule interactions through a gating mechanism. The second target ligand atom type characteristics; and This represents the sigmoid activation function used to generate the gated signal; and These are the learnable parameters of the gating unit; Represents the dot product of vectors; This indicates the operation of the self-attention mechanism; The primary target protein atom type characteristics; This refers to the atomic type characteristics of the second target protein.
9. The method for generating ligand molecules based on multimodal information guidance according to claim 5, characterized in that, The second updated ligand molecule '' and the second updated protein Obtained through the following methods: in, For protein Protein target type characteristics, Ligand molecules at different diffusion steps Ligand molecule target type characteristics; For protein A set of protein atom type features. Ligand molecules at different diffusion steps The set of ligand atom type characteristics; The first target protein atom type characteristics, This refers to the atom type characteristics of the second target protein; ligand molecules atomic type characteristics, Ligand molecules with different diffusion steps for different ligand molecule invariance characteristics Atomic type characteristics; This is a multilayer sensor; at this point, the protein... Updated to the second update protein ''= ; Ligand molecules Updated to the second updated ligand molecule ''= .
10. The method for generating ligand molecules based on multimodal information guidance according to claim 5, characterized in that, The second updated protein is obtained in the following manner: A second, updated protein is obtained by manipulating the protein-ligand complex using a denoising network. For inclusion The protein-ligand complex is composed of an SE(3) isovariant neural network with an isovariant layer, and the protein-ligand complex is dynamically constructed from a second updated target protein and a second updated target ligand molecule. go through After performing the same operations on each layer, the updated protein-ligand complex atoms are obtained after one time step update via a denoising network. 3D coordinate features And type characteristics; thereby obtaining updated protein-ligand complex three-dimensional coordinate characteristics and complex type characteristics; The updated protein-ligand complex obtained from the previous time step was repeated using a denoising network. - One operation yields the target protein-ligand complex; The three-dimensional coordinate features of the ligand molecule in the target protein-ligand complex are converted into the original three-dimensional coordinates through an MLP layer; the type features of the ligand molecule in the target protein-ligand complex are converted into the original type through an MLP layer; the target ligand molecule is composed of the original three-dimensional coordinates and the original type.