Multimodal molecular joint generation method, device, equipment and medium based on diffusion model

Abstract

The present invention relates to a method, device, equipment and medium for joint generation of multimodal molecules based on a diffusion model, and to the fields of medicinal chemistry and computer technology. The method integrates discrete and continuous diffusion mechanisms, and is used to learn and generate multimodal molecular representations including atomic properties, two-dimensional molecular structures and three-dimensional molecular coordinates to promote molecular structure-based drug design. Through the diffusion process, the method can effectively capture the probabilistic characteristics of molecular generation and analyze the effects of various factors on molecular structure. In addition, an innovative graph Transformer architecture is used to denoise the diffusion process, which can not only learn stable representations of atoms and bonds, but also learn molecular representations with high robustness to geometric transformations based on invariant constraints such as rotation and translation in three-dimensional space. The present invention also provides a device, computer equipment and medium for running the method. The method has very high molecular validity, uniqueness and stability.

Description

Technical Field

[0001] The present invention relates to the field of medicinal chemistry and computer technology, and in particular to a method, device, equipment and medium for jointly generating multimodal molecules based on a diffusion model. Background Art

[0002] Small molecule drugs are extremely convenient to produce, transport, and store, and they occupy a core position in the pharmaceutical industry. In addition, the high compliance of patients, a wide range of targets, and low immunogenicity play a vital role in patient treatment. However, traditional drug design faces challenges such as long cycles, large investments, high thresholds, and low success rates. In the initial stages of new drug research and development, molecular design is one of the most critical steps. Accurate molecular design is essential to ensure the effective combination of drug molecules with targets, and to enhance wet experiment results and therapeutic potential. Although existing compound databases and molecular screening software are capable of preliminary screening, the drug screening process often requires huge costs.

[0003] In recent years, the application of artificial intelligence technology has brought revolutionary changes to drug research and development, and has gradually become a key breakthrough in the development of new drugs. Deep generative models have been widely used in drug design, and a large number of methods and optimization strategies have been explored. As the main target of drug design, target proteins provide a key foundation for drug design and discovery, and play a core role in the development of safer and more effective treatments. In the field of molecular generation, by accurately identifying the active sites on proteins and incorporating the structural information of target proteins into the process of molecular generation, molecules can be designed more specifically to achieve the regulation of specific targets, thereby improving the efficacy and safety of drugs in vivo. Designing drug molecules that can interact with these targets can not only improve the specificity of drugs and reduce side effects, but also provide more reliable means for treatment. By guiding the model to focus on specific protein structural features, potential drug molecules can be screened more effectively, which can not only speed up the development of new drugs, but also reduce the risk of failure during the development process.

[0004] Existing molecular generation methods can be roughly divided into three categories: based on one-dimensional strings, based on two-dimensional graph structures, and based on three-dimensional geometric structures. Since the one-dimensional representation method of SMILES strings cannot capture molecular similarities and has validity issues in the generation stage, it is not the best choice. As a typical two-dimensional molecular representation, molecular graphs use nodes and edges to represent atoms and bonds, describing the two-dimensional bonding topology of molecules. This intuitive representation is very convenient for chemical synthesis, molecular dynamics simulation, etc. However, molecules exist in three-dimensional physical space, and their geometric shape affects their quantum properties, which are difficult to accurately estimate from low-dimensional representations. Low-dimensional molecular generation models may require additional simulations to find low-energy conformations and may generate unstable molecules. In addition, application scenarios such as structure-based drug design also rely on the three-dimensional geometry of molecules. With the rapid development of deep learning in geometric data processing, generative models for 3D molecules have been proposed and have attracted increasing attention due to their advantages and potential in directly generating 3D molecular conformations and structure-based drug design. However, most current generative models are still limited to single-dimensional modeling, while molecular representations in different dimensions are complementary. The lack of joint modeling capabilities limits the improvement of molecular generation quality and downstream applications. Summary of the invention

[0005] The present invention proposes a method that combines discrete and continuous diffusion mechanisms to learn and generate multimodal molecular representations that include atomic properties, two-dimensional molecular structures, and three-dimensional molecular coordinates to promote molecular structure-based drug design. Through this diffusion process, the method can effectively capture the probabilistic characteristics of molecular generation and analyze the effects of various factors on molecular structure. At the same time, an innovative graph Transformer architecture is introduced in the denoising process, which can not only learn stable representations of atoms and bonds, but also learn molecular representations that are highly robust to geometric transformations based on invariant constraints such as rotation and translation in three-dimensional space.

[0006] The present invention is achieved through the following technical solutions:

[0007] A multimodal molecular joint generation method based on a diffusion model, the method specifically comprising the following steps:

[0008] Step 1. Construct drug molecule datasets containing samples with complete two-dimensional and three-dimensional structures and a subset of samples containing only two-dimensional structures, and perform standardization on the datasets including filling in missing 3D coordinate information;

[0009] Specific implementation method: Obtain high-resolution protein three-dimensional structure from the protein structure database, remove water molecules, ions and irrelevant ligands, repair missing atoms, and construct a target protein data set; use protein pocket identification tools to identify potential binding pockets, minimize the energy of the pocket structure, define the boundaries and characteristics of the pocket and identify key amino acid residues, and extract pocket coordinate information to generate a grid file.

[0010] Step 2. Taking the structural information obtained in step 1 as input and the target structure as background information, noise is applied to the atom and edge representations to enhance the generation process during the forward diffusion process;

[0011] Specifically, continuous noise is applied to atomic representations such as atomic features and coordinates, while discrete noise is introduced into edge representations in the graph structure, capturing both geometric and topological information for a more comprehensive molecular generation. The molecule is represented as M = (A, E, X), where A is the atomic feature, E is the two-dimensional graph structure feature of the chemical bond, one-hot encoding is used to represent the bond type, and X is the three-dimensional coordinate of the atom.

[0012] Furthermore, a time step t is sampled from the uniform distribution to control the noise level of the diffusion process, and Gaussian noise is added to the atomic features and coordinates at each time step t to generate a noise state and The noise added to the atomic feature A at time step t satisfies The noise added to the atomic three-dimensional coordinates X satisfies At the same time, uniform noise is added to the two-dimensional graph structure feature E of the chemical bond to generate a noise state And obtain the atomic characteristics of the noise state at time step t and the three-dimensional coordinates of the atoms This ensures that the perturbations applied to the three-dimensional coordinates do not affect the molecular center of gravity, so that the denoising process remains unchanged during translation, rotation, etc. In addition, at t time steps, the discrete two-dimensional graph structure features E of the chemical bonds are transformed into noisy bond features

[0013] Step 3. Take the atomic features A, the two-dimensional graph structure features E of the chemical bond, and the three-dimensional coordinates X of the atoms in step 2 as input data, obtain the atomic, positional, and structural information through the encoding function, and integrate the information into a unified representation; use the message passing mechanism to merge the paired atomic information into the edge encoding. For each edge e ij ,Use a permutation invariant function to generate the embedded edge representation to ensure consistency;

[0014] Further, from the two-dimensional molecular graph M 2DObtain in-degree and out-degree, incorporate centrality information into atomic encoding, and update node atomic information To obtain local neighborhood information, message passing is used to aggregate the two-dimensional molecular graph M 2D The neighbor information of each atom in the node is expressed as Furthermore, message passing is used to aggregate atomic information near a given atom to obtain a three-dimensional neighborhood code M 3D .

[0015] Step 4. Remove the noise molecules As input, and output the clean molecule M t The estimated value of and the noise of atomic features and coordinate noise In this way, the noise added by the noise process is distinguished from the real atomic features A and the atomic three-dimensional coordinates X, and the estimated noise of the coordinates Subtract the input coordinates from the y-axis to obtain estimates of atomic features and coordinates;

[0016] Step 5. Calculate the loss between the atomic features predicted by the network at the tth time step, the two-dimensional graph structural features of the chemical bonds, the noise estimation value of the atomic three-dimensional coordinates and the real noise, accurately predict the atomic features and the real noise in the atomic three-dimensional coordinates by minimizing the distance between the real noise and the model predicted noise, classify the two-dimensional graph structural features of the chemical bonds and minimize the cross entropy loss of each atom pair, and calculate the gap between the actual bond type and the predicted bond probability distribution;

[0017] Step 6. The model is trained to sample new molecules; clean molecules M are separated from noise molecules by using the posterior distribution; the denoising process iteratively updates the noise molecules, using the predicted probability distribution of the two-dimensional graph structure of the noise and the molecular chemical bond to guide the update, and recovering the clean molecules from the noise molecules; the real sampling process Using an approximation of the complete molecule

[0018] The output module converts the atom and edge embeddings output by the Transformer channel into the final prediction results. Edge Embedding Atomic 3D coordinate embedding Input to the output layer, extract features through SiLU activation function and weight matrix, and get atomic feature prediction A out , edge feature prediction E out and atomic coordinates predicted by X out Finally, the predicted atomic three-dimensional coordinates are added to the input initial atomic three-dimensional coordinates to obtain the final atomic coordinates X out These three parts together constitute the final predicted molecule M out =(Aout ,E out ,X out ), including atomic features, edge features and atomic three-dimensional coordinates.

[0019] The present invention also provides a device for the joint generation of multimodal molecules based on a diffusion model, and the device runs the molecular intelligent generation method based on a diffusion model.

[0020] The present invention also provides a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the multimodal molecular joint generation method based on a diffusion model.

[0021] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor and executing the multimodal molecular joint generation method based on a diffusion model.

[0022] The beneficial effects of the present invention compared with the prior art are as follows:

[0023] Based on the diffusion model and graph transformer architecture, the present invention achieves a complete representation of molecules by learning the connectivity of molecular graphs and the spatial atomic arrangement, while generating multimodal molecular structure information. The present invention performs well in the final generation effect, with very high molecular validity, uniqueness and stability.

[0024] (1) The present invention adopts a diffusion model to add continuous noise to the continuous features of the molecular structure and discrete noise to the discrete features. The clean graph structure is predicted by the denoising model, and the noise of atomic features and coordinates is estimated.

[0025] (2) The present invention can facilitate the interaction between 2D and 3D structural information and calculate graph connectivity and spatial atomic arrangement. When calculating the 3D geometric structure, rotation-translation equivariance constraints are followed to ensure generation accuracy. In addition, it can operate independently even when the 2D or 3D structural information is incomplete, providing a robust and versatile solution for processing incomplete molecular data sets.

[0026] (3) The diffusion model is combined with the graph Transformer architecture, which has better results than previous generative models. Continuous and discrete noise are added to the forward diffusion process to systematically perturb the distribution in the data, and a parameterized neural network is used to learn the reverse diffusion process to restore the distribution of the data. It is highly flexible and easy to calculate, and has better performance in generating and learning comprehensive representations of molecules.

[0027] (4) The uniqueness of the present invention over previous deep learning models is that it simultaneously generates and learns multimodal molecular structures, thereby improving the effectiveness and uniqueness of the generated molecules. By combining the molecular graph connectivity and spatial atomic arrangement information, it is possible to generate unique and accurate molecular structures while ensuring structural integrity. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the multi-channel diffusion model;

[0029] Figure 2 This is a partial diagram of the multimodal molecular structure. DETAILED DESCRIPTION

[0030] The technical solution of the present invention is further explained below through embodiments in combination with the accompanying drawings, but the protection scope of the present invention is not limited in any form by the embodiments.

[0031] Example 1: The main goal of this example is to learn and generate multimodal molecular representations including atomic features, discrete two-dimensional molecular structures and continuous three-dimensional molecular coordinates based on the provided specific protein pocket as fixed background information through a combination of discrete and continuous diffusion processes to achieve structure-based drug design. A new graph Transformer architecture is introduced to denoise the diffusion process. Under the premise of following denaturation constraints such as 3D rotation and translation and maintaining denaturation such as atomic coordinates, invariant atomic and bond representations are learned, thereby learning molecular representations that are highly robust to geometric transformations. Ultimately, drug-like ligands with novel and diverse structures and high binding energy to protein pockets are generated.

[0032] A multimodal molecular joint generation method based on a diffusion model comprises the following steps:

[0033] The data in this example are from (1) QM9 dataset and (2) RCSB PDB protein structure database. The data selected in this example are only used to illustrate the process of the present invention, but are not limited to the datasets described in the example.

[0034] Step 1. Construct a drug dataset based on the QM9 dataset, which contains 130,000 small molecules and their molecular properties and structural information. The dataset is divided into two subsets: one contains samples with complete two-dimensional and three-dimensional structures, and the other contains only two-dimensional structural information to simulate the situation of sparse data, and the original data is standardized by filling missing 3D coordinate information. The dataset is divided into training set, validation set and test set in the ratio of 100,000 / 17,000 / 13,000.

[0035] High-resolution protein structures were obtained from the protein database RCSB PDB, and water molecules, ions and irrelevant ligands were removed, and missing atoms were repaired to construct a target protein data set. The PD Bbind tool was used to identify potential binding pockets based on the geometric features and chemical properties of the protein surface, and the pocket structure was energy minimized to ensure stability. The boundaries and characteristics of the pocket were defined by calculating the volume, surface area and shape descriptors, and key amino acid residues were identified. The pocket coordinate information was extracted, and a grid file containing energy points in the pocket area was generated.

[0036] Before performing structure-based drug design in this embodiment, the RDKit software is used to identify the data in the database in step 1 and convert it into a graph structure encoding form.

[0037] Step 2. Using the structural information obtained in step 1 as input and the target structure as background information, noise is applied to the atomic and edge representations during the forward diffusion process to enhance the generation process. Specifically, continuous noise is applied to representations such as atomic features and atomic 3D coordinates, while discrete noise is introduced into the edge representations in the graph structure to capture both geometric and topological information, achieving more comprehensive molecular generation. The molecule is represented as M = (A, E, X), where n is the number of atoms, d is the characteristic dimension; is the two-dimensional graph structural feature of chemical bonds. The bond type is represented by one-hot encoding, and b is the number of bond types; Furthermore, a time step t is sampled from the uniform distribution to control the noise level of the diffusion process. Gaussian noise is added to the atomic features and atomic 3D coordinates at each time step t to generate a noise state and The noise added to the atomic feature A at time step t satisfies The noise added to the atomic three-dimensional coordinates X satisfies At the same time, uniform noise is added to the two-dimensional graph structure feature E of the chemical bond to generate a noise state And obtain the noisy atomic features at time step t and the three-dimensional coordinates of the atoms

[0038]

[0039] This ensures that the perturbation applied to the three-dimensional coordinates does not affect the molecular center of gravity, making the denoising process invariant during translation, rotation, etc. t is the data retention parameter at time step t, σ t is the parameter for adding Gaussian noise. In addition, at time step t, the discrete two-dimensional graph structure feature E of chemical bonds is transformed into the two-dimensional graph structure feature of chemical bonds with noise in is the transfer matrix:

[0040]

[0041] Step 3. Take the atomic features A, the two-dimensional graph structure features E of the chemical bond, and the three-dimensional coordinates X of the atoms in step 2 as input data, obtain the atomic, positional, and structural information through the encoding function, and integrate the information into a unified representation. Use the message passing mechanism to merge the paired atomic information into the edge encoding. For each edge e ij , a permutation invariant function is used to generate the embedded edge representation to ensure consistency.

[0042] From the two-dimensional molecular graph M 2D Get the in-degree deg - and out-degree deg + , after incorporating the centrality information into the atomic encoding, update the node atomic information

[0043]

[0044] in, is the learnable parameter of the atomic feature, and the in-degree deg - and out-degree deg + To obtain local neighborhood information, message passing is used to aggregate the two-dimensional molecular graph M 2D The neighbor information of each atom in . Specifically, the aggregation of node atoms is expressed as:

[0045]

[0046] in, is the learnable parameter of the atomic feature, N(i) represents the neighborhood of atom i, m j→i Represents a message. Further, the atomic information near a given atom is aggregated using message passing to obtain a three-dimensional neighborhood code M 3D . For atom i:

[0047] d ij =||x i -x j ||2 (Formula 6)

[0048]

[0049] Among them, d ij is the Euclidean distance between atoms i and j, e ij represents the bond characteristics between atoms i and j in the molecule, is the radial basis kernel function, ⊙ is the Hadamard product, f cos (·) is the cosine cutoff function, W edgeis the key feature learnable parameter, W comb To aggregate neighbor information, b comb is the edge embedding bias term, is the learnable parameter of atomic features. Step 4. As input, and output the clean molecule M t and noise in atomic features and coordinates The estimated value of is used to distinguish the noise added by the noise process from the real atomic features A and atomic three-dimensional coordinates X:

[0050]

[0051] ψ θ is a parameterized function that receives the noise molecule and time step t as input, and outputs the predicted probability distribution of noise and bond type in, is a normalization factor that scales the value of time step t to between 0 and 1 to better handle the noise at different time steps. is the estimated value of atomic features and coordinates. t Used to control the amount of noise added to atomic features and coordinates, σ t Used to control the intensity of noise added to atomic features and coordinates. Further, the input coordinates are subtracted from the estimated noise of the coordinates to obtain estimates of atomic features and coordinates.

[0052] Step 5. Calculate the loss between the noise estimate and the actual noise of the atomic features, atomic three-dimensional coordinates, and two-dimensional graph structural features of chemical bonds predicted by the network at the tth time step From the Gaussian distribution N A (0,I),N X Randomly select atomic characteristic noise in (0,I) 3D coordinate noise By minimizing the distance between the true noise and the model's predicted noise To accurately predict the true noise in atomic features and 3D coordinates, classify bond features and minimize the cross entropy loss for each atom pair, and calculate the gap between the actual bond type and the predicted bond probability distribution.

[0053]

[0054] in, Used to balance the importance of atomic features and coordinates, SNT(*) is the signal-to-noise ratio; represents the set of all edges (i, j), E ij represents the edge type connecting atoms i and j, represents the probability of the true bond type.

[0055] Step 6. The model is trained to sample new molecules. Clean molecules M are separated from noise molecules by using the posterior distribution. The denoising process iteratively updates the noise molecules, using the predicted probability distribution of noise and molecular graph structure to guide the update, and recovers clean molecules from the noise molecules. Actual sampling process Using an approximation of the complete molecule The product of the posterior distributions of atomic features, coordinates, and bond features is:

[0056]

[0057] Model schematic diagram Figure 1 As shown in Figure 2, the output module converts the atom and edge embeddings output by the Transformer channel into the final prediction results. Edge Embedding Atomic 3D coordinate embedding Input to the output layer, extract features through SiLU activation function and weight matrix, and get atomic feature prediction A out , edge feature prediction E out and atomic coordinates predicted by X out Finally, the predicted atomic coordinates are added to the input initial atomic coordinates to obtain the final atomic coordinates X out These three parts together constitute the final predicted molecule M out =(A out ,E out ,X out ), including atomic features, edge features, and atomic coordinates.

[0058] The 100,000 training molecules were randomly divided into two groups, 30,000 groups containing 2D and 3D structures, and 70,000 groups containing only 2D structures, and the 3D structure was inferred from the 2D structure. The model was trained on the 30,000 data sets using invariant and equivariant channels, with 1000 time steps in the forward diffusion process, 10,000 rounds of training, and fine-tuned on the 70,000 data sets, and verified and tested.

[0059] The results of the molecular generation in this example are as follows Figure 2 As shown, the similarity and score comparison were performed with the traditional single-modality structure-based designed drug molecules as reference, as shown in Table 1. Some of the drug-like ligands generated by this method are not only novel and diverse, but also show high binding energy when binding to protein pockets;

[0060] Table 1 Comparison of generation effects between traditional generation method and joint generation method

[0061]

[0062] It should be noted that, although the embodiments of the present invention described above are illustrative, they are not intended to limit the present invention, and therefore the present invention is not limited to the above specific embodiments. Without departing from the principles of the present invention, any other embodiments obtained by those skilled in the art under the guidance of the present invention are deemed to be within the protection of the present invention.

Claims

1. A multimodal molecular joint generation method based on a diffusion model, characterized in that: The method specifically comprises the following steps: Step 1. Construct drug molecule datasets containing samples with complete two-dimensional and three-dimensional structures and a subset of samples containing only two-dimensional structures, and perform standardization on the datasets including filling in missing 3D coordinate information; Step 2. Taking the standardized dataset obtained in step 1 as input and the target structure as background information, noise is applied to the atom and edge representations during the forward diffusion process to enhance the generation process; Continuous noise is applied to atomic representations including atomic features and atomic 3D coordinates, while discrete noise is introduced into edge representations in the graph structure, capturing both geometric and topological information to achieve more comprehensive molecule generation; molecules are represented as M = (A, E, X), where A is the atomic feature, E is the 2D graph structure feature of the chemical bond, one-hot encoding is used to represent the bond type, and X is the 3D coordinate of the atom; Step 3. Take the atomic feature A, the two-dimensional graph structure feature E of the chemical bond and the three-dimensional coordinate X of the atom in step 2 as input data, obtain the atomic, position and structure information through the encoding function, and integrate the information into a unified representation; The pairwise atomic information is incorporated into the edge encoding using the message passing mechanism. For each edge e ij ,Use a permutation invariant function to generate the embedded edge representation to ensure consistency; Step 4. Remove the noise molecules As input, and output the clean molecule M t The estimated value of and the noise of atomic features and the estimated noise of the coordinates In this way, the noise added by the noise process is distinguished from the real atomic features A and the atomic three-dimensional coordinates X, and the estimated noise of the coordinates Subtract the input coordinates from the y-axis to obtain estimates of atomic features and coordinates; Step 5. Calculate the loss between the atomic features predicted by the network at the tth time step, the two-dimensional graph structural features of the chemical bonds, the noise estimation value of the atomic three-dimensional coordinates and the real noise, accurately predict the atomic features and the real noise in the atomic three-dimensional coordinates by minimizing the distance between the real noise and the model predicted noise, classify the two-dimensional graph structural features of the chemical bonds and minimize the cross entropy loss of each atom pair, and calculate the gap between the actual bond type and the predicted bond probability distribution; Step 6. The model is trained to sample new molecules; Separate the clean molecules M from the noisy molecules by using the posterior distribution; The denoising process iteratively updates the noise molecules, using the predicted probability distribution of noise and molecular graph structure to guide the update and recover clean molecules from the noise molecules; in the real sampling process The approximation using the complete molecule The output module converts the atom and edge embeddings output by the Transformer channel into the final prediction results.

2. The multimodal molecular joint generation method based on diffusion model according to claim 1, characterized in that: In the step 1, a high-resolution protein three-dimensional structure is obtained from a protein structure database, water molecules, ions and irrelevant ligands are removed, missing atoms are repaired, and a target protein data set is constructed; a protein pocket identification tool is used to identify potential binding pockets, the pocket structure is energy minimized, the boundaries and characteristics of the pocket are defined, key amino acid residues are identified, and pocket coordinate information is extracted to generate a grid file.

3. The multimodal molecular joint generation method based on diffusion model according to claim 1, characterized in that: In step 2, a time step t is first sampled from the uniform distribution to control the noise level of the diffusion process, and Gaussian noise is added to the atomic features and coordinates at each time step t to generate a noise state and The noise added to the atomic feature A at time step t satisfies The noise added to the atomic three-dimensional coordinates X satisfies At the same time, uniform noise is added to the two-dimensional graph structure feature E of the chemical bond to generate a noise state And obtain the atomic characteristics of the noise state at time step t and the three-dimensional coordinates of the atoms To ensure that the perturbation applied to the three-dimensional coordinates does not affect the molecular center of gravity, so that the denoising process remains unchanged during translation, rotation, etc., the discrete two-dimensional graph structure feature E of the chemical bond is transformed into a noisy bond feature at t time steps.

4. The multimodal molecular joint generation method based on diffusion model according to claim 1, characterized in that: Step 3, from the two-dimensional molecular graph M 2D Obtain in-degree and out-degree, incorporate centrality information into atomic encoding, and update node atomic information To obtain local neighborhood information, message passing is used to aggregate the two-dimensional molecular graph M 2D The neighbor information of each atom in the node is represented by the aggregation of node atoms as atom embedding Aggregate the atomic information near a given atom using message passing to obtain the three-dimensional neighborhood code M 3D .

5. The multimodal molecular joint generation method based on diffusion model according to claim 1, characterized in that: Step 6, embedding atoms Edge Embedding Atomic 3D coordinate embedding Input to the output layer, extract features through SiLU activation function and weight matrix, and get atomic feature prediction A out , edge feature prediction E out and atomic coordinates predicted by X out Finally, the predicted atomic three-dimensional coordinates are added to the input initial atomic three-dimensional coordinates to obtain the final atomic coordinates X out These three parts together constitute the final predicted molecule M out =(A out ,E out ,X out ), including atomic features, edge features and atomic three-dimensional coordinates.

6. A device for multimodal molecular joint generation based on a diffusion model, characterized in that: The device runs a multimodal molecular joint generation method based on a diffusion model as described in any one of claims 1-5.

7. A computer device, characterized in that: The device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of a diffusion model-based multimodal molecular joint generation method as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor and executing a multimodal molecular joint generation method based on a diffusion model as described in any one of claims 1 to 5.

Images