Drug generation method based on state space model and diffusion model

Abstract

The present invention relates to a drug generation method based on a state space model and a diffusion model, and belongs to the field of medical information technology. The method first performs three-dimensional modeling on target protein molecules and drug molecules, and uses the state space model to extract more discriminative molecular structure features. Subsequently, the diffusion model is used to generate drugs to generate drug molecules with more realistic structures. The present invention effectively improves the generation quality of drug molecules by adding noise during the training process and denoising through a prediction network. The drug molecules finally generated have better drug-like properties, easy synthesis, and are more easily combined with specific target proteins, thereby providing an efficient and intelligent method for drug research and development.

Description

Technical Field

[0001] The invention belongs to the field of medical informationization and relates to a drug generation method based on a state space model and a diffusion model. Background Art

[0002] Artificial intelligence algorithms can process and analyze massive amounts of data, thereby revealing the patterns and laws hidden behind the data. Through automated and intelligent data mining and analysis, artificial intelligence algorithms have greatly shortened the time consumed by scientific research, allowing experiments and analysis processes that previously took years to be completed in a shorter time. This not only improves research efficiency, but also promotes the evolution and innovation of scientific theories. The interdisciplinary application of artificial intelligence algorithms has also promoted exchanges and cooperation between different fields. In the biomedical field, the application of artificial intelligence algorithms has accelerated the process of drug development, but designing drug molecules that can bind to specific target proteins is still a basic and challenging problem in drug discovery.

[0003] Therefore, the present invention relates to a drug generation method based on a state space model and a diffusion model, exploring the use of a state space model (mamba) to model target protein molecules and drug molecules to obtain more discriminative molecular structural features, in addition, the present invention generates drugs through a diffusion model (diffusion model). The drug molecules finally generated have better drug-likeness (QED, Quantitative Estimate of Drug-likeness), easy synthesis (SA, Synthetic Accessibility score), and are easier to bind to specific target proteins.

[0004] As more and more scholars begin to study the application of artificial intelligence algorithms in traditional fields such as life sciences, widely recognized public datasets have emerged, such as PDBbind[1] and CrossDocked2020[2]. The PDBbind dataset is the first dataset that systematically annotates protein-ligand (target protein-drug molecule, drug-like molecule or compound molecule) complexes in the PDB protein database with experimental binding data. The 2016 version of the dataset provides binding data for 16,179 biomolecular complexes in the PDB protein database. The dataset provided by PDBbind has been applied to many computational and statistical studies on protein-ligand interactions and various disciplines, and has become the main data resource for research scoring indicators. The CrossDocked2020 dataset contains more than 22 million docked protein-ligand crystal structures for evaluating structure-based drug design methods. After LiGAN[3], many scholars ignored molecular structures with RMSD greater than , thereby obtaining a dataset of approximately 500,000 protein-ligand complexes. Among them, the total number of atom types in the ligand and protein binding sites is 27 and 19, respectively.

[0005] In recent years, artificial intelligence algorithms have been proven to be effective in extracting rich biochemical structural features and modeling target protein molecules and drug molecules. AlphaFold[4] is an advanced protein structure prediction algorithm. It uses deep neural networks and machine learning techniques to predict the three-dimensional structure of proteins by analyzing amino acid sequences. AlphaFold has learned the structural data of millions of known proteins by training the model, enabling it to predict the structure of new proteins with unprecedented accuracy and speed. This technology is of great significance for understanding how proteins function in organisms, and helps to accelerate drug development, design new enzymes and understand disease mechanisms. The success of AlphaFold has laid a solid foundation for future research and application. Pocket2Mol[5] and GraphBP[6] are both graph neural network-based methods that generate drug molecules with three-dimensional structures by sequentially placing atoms into protein binding pockets. They predict the probability of adding new atoms at different positions based on the protein pockets and the atoms already placed. The difference is that Pocket2Mol directly predicts the generation process of new atoms in the sequence, while GraphBP also considers the relative position relative to the coordinate system when placing new atoms. Compared with previous methods, Pocket2Mol not only models the three-dimensional coordinates of atoms, but also pays attention to chemical structure information such as bond types and functional group characteristics. The graph neural network designed by it can efficiently capture spatial features and the bonding relationship between protein pockets and atoms. The drug molecules generated by Pocket2Mol have better binding affinity and other drug properties, such as drug-likeness (QED) and ease of synthesis (SA). GraphBP captures three-dimensional structural features through graph neural networks, models the chemical interactions between protein-ligand complexes, and maintains isovariance during the generation process. TargetDiff[7] is a three-dimensional isovariant diffusion model for target-aware molecule generation and enhanced binding affinity prediction. It solves the problems of violating rotational isovariance, violating geometric constraints, and generating unrealistic structures caused by the three-dimensional target-aware model relying on voxelized atomic density or relying on autoregressive sampling processes. This method can generate molecules with more realistic three-dimensional structures and stronger binding affinity to specific target proteins.

[0006] [1]Liu, Zhihai, et al. "Forging the basis for developing protein–ligandinteraction scoring functions." Accounts of chemical research 50.2(2017):302-309.

[0007] [2]Francoeur,Paul G.,et al."Three-dimensional convolutional neuralnetworks and across-docked data set for structure-based drug design."Journalof chemical information and modeling 60.9(2020):4200-4215.

[0008] [3]Ragoza,Matthew,Tomohide Masuda,and David Ryan Koes."Generating 3Dmolecules conditional on receptor binding sites with deep generative models."Chemical science 13.9(2022):2701-2713.

[0009] [4]Jumper,John,et al."Highly accurate protein structure predictionwith AlphaFold."nature596.7873(2021):583-589.

[0010] [5]Peng,Xingang,et al."Pocket2mol:Efficient molecular sampling basedon 3d protein pockets."International Conference on Machine Learning.PMLR,2022.

[0011] [6]Liu,Meng,et al."Generating 3d molecules for target proteinbinding."arXiv preprint arXiv:2204.09410(2022).

[0012] [7]Guan, Jiaqi, et al. "3d equivariant diffusion for target-awaremolecule generation and affinity prediction." arXiv preprint arXiv:2303.03543(2023). Summary of the invention

[0013] In view of this, the purpose of the present invention is to provide a drug generation method based on a state-space model and a diffusion model to design drug molecules that can bind to specific target proteins. The present invention can be used to generate drug molecules that bind to specific target proteins, and use a state-space model (mamba) to model target protein molecules and drug molecules to obtain more discriminative molecular structural features. The drug molecules finally generated have better drug-likeness (QED, Quantitative Estimate of Drug-likeness), easy synthesis (SA, Synthetic Accessibility score), and are easier to bind to specific target proteins.

[0014] In order to achieve the above object, the present invention provides the following technical solutions:

[0015] The drug generation method based on the state space model and the diffusion model comprises the following steps:

[0016] a) preparing data, including three-dimensional coordinates and elemental characteristic data of target protein molecules, and three-dimensional coordinates and elemental characteristic data of drug molecules;

[0017] b) training a prediction network based on a state-space model, which uses the noisy drug molecule data as input to predict the position information and elemental characteristics of the drug molecules;

[0018] c) a sampling prediction network, which generates the three-dimensional coordinates and elemental characteristic data of the drug molecule using the three-dimensional coordinates and elemental characteristic data of the target protein molecule as control conditions;

[0019] d) Evaluate the generated drug molecules, including drug-likeness, ease of synthesis, and molecular docking scores.

[0020] Further, in a), the three-dimensional coordinates of the target protein molecule are the position information data px0 and element characteristic data pv0 of the target protein molecule, and the three-dimensional coordinates of the real drug molecule are processed into the position information data lx0 and element characteristic data lv0 of the drug molecule; in subsequent steps, px0, pv0, lx0 and lv0 will be used as input data of the deep learning network.

[0021] Further, in b), the prediction network includes a multi-layer perceptron network for feature extraction and a mamba network structure for state space modeling;

[0022] Add noise to the position information data lx0 and element feature data lv0 obtained from the three-dimensional information L0 of the real drug molecule; for the position information lx0, add continuous noise; for the element feature lv0, add uniform discrete noise;

[0023] The training goal is to train the prediction network MPN based on the state space model, and transform the position information lx of the real drug molecules after adding noise t , elemental features of real drug molecules after adding noise t , clear target protein molecule position information px0 and clear target protein molecule element information pv0 as input data, predict the generated drug molecule data and For position information, the mean square error loss MSEloss is used to calculate the distance of continuous data; for element features, the KL divergence KL_loss is used to evaluate the difference of discrete data.

[0024] Further, in c), the sampling process includes a denoising process, and the denoising process uses the generated drug molecule data predicted by the prediction network as the prior probability;

[0025] The position information px0 and element features pv0 are obtained from the three-dimensional information P0 of the target protein, and the generated drug molecules are initialized and represents the position information of the generated drug molecules approximated by noise, represents the elemental features of the generated drug molecules approximated by noise;

[0026] Perform T = 1000 samplings; in each sampling, use the prediction network MPN based on the state space model trained in the second step to predict the position information of the generated drug molecules and element characteristics And use it as the prior probability to finally get the posterior probability: Generate the position information of the drug molecule at step T and element characteristics

[0027] Furthermore, in step d), the drug-likeness is evaluated using the quantitative estimation of drug similarity (QED) method.

[0028] Furthermore, in d), the ease of synthesizing is evaluated using the synthetic accessibility score SA method.

[0029] Furthermore, in said d), the molecular docking score is evaluated using Vina Score, Vina Min and Vina Dock methods.

[0030] Furthermore, the three-dimensional coordinates and elemental characteristic data of the target protein molecule are obtained from the .pdb file.

[0031] Furthermore, the three-dimensional coordinates and elemental characteristic data of the drug molecule are obtained from the .sdf file.

[0032] Furthermore, the dataset is the CrossDocked2020 dataset.

[0033] The beneficial effects of the present invention are:

[0034] (1) Compared with existing methods, the drug molecular structure generated by the present invention is more realistic and more consistent with the characteristics of real-world drug molecules.

[0035] (2) The drug molecules generated by the present invention have better drug-like properties, that is, they are closer to the physical and chemical properties of real drug molecules, such as molecular weight, logP, number of hydrogen bond donors and acceptors, etc.

[0036] (3) The drug molecules generated by the present invention are more easily synthesized, that is, easier to prepare by chemical synthesis methods, thereby reducing the cost and cycle of drug research and development.

[0037] (4) The drug molecules generated by the present invention have a stronger binding force with the specific target protein, that is, have a higher affinity, and can more effectively inhibit the function of the target protein, thereby exerting a therapeutic effect.

[0038] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0040] Figure 1 This is a flow chart of training, sampling and evaluation of the drug generation method based on the state space model and diffusion model proposed in the present invention.

[0041] Figure 2The drug generation method based on the state space model and diffusion model proposed in the present invention is trained on the CrossDocked2020 dataset, and the results of drug molecules generated by selecting the target protein of the CrossDocked2020 dataset as the control condition of the diffusion model are displayed. DETAILED DESCRIPTION

[0042] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0043] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0044] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0045] The present invention is specifically implemented according to the following steps:

[0046] The first step is to prepare the data.

[0047] The present invention relates to a drug generation method based on a state space model and a diffusion model, and the required data are three-dimensional information of target protein molecules and drug molecules. Among them, the three-dimensional information of the target protein molecule is a .pdb file, including the protein pocket name, atomic number, element name, identifier, three-dimensional coordinates [x, y, z] and related properties of the target protein; the three-dimensional information of the drug molecule is a .sdf file, including the chemical structure name, atomic number, chemical bond number, three-dimensional coordinates [x, y, z], atomic information, chemical bond information and chemical properties of the drug molecule. The method proposed by the present invention uses the same data processing method as the method TargetDiff, and the three-dimensional information processing of the target protein molecule is the three-dimensional coordinate [x, y, z] data px0 and element feature data pv0 of the target protein molecule, and the three-dimensional information processing of the real drug molecule is the three-dimensional coordinate [x, y, z] data lx0 and element feature data lv0 of the drug molecule. In the subsequent steps, px0, pv0, lx0 and lv0 will be used as input data of the deep learning network.

[0048] In the second step, a deep learning network based on the state-space model and the diffusion model is trained.

[0049] The training process of the drug generation method based on the state space model and the diffusion model proposed in the present invention is as follows: Figure 1 The upper part of the diagram shows the training process.

[0050] (1) First, noise is added to the position information data lx0 and element characteristic data lv0 obtained from the three-dimensional information L0 of the real drug molecule. For lx0, since it is continuous data representing three-dimensional coordinates, the noise addition formula is:

[0051]

[0052] Among them, lx0 represents the original position information of the real drug molecule, lx t represents the position information of the real drug molecule after adding noise, ∈ represents Gaussian noise that satisfies the normal distribution (mean is 0, variance is 1), A constant that controls the degree of noise addition. The value range is from 0 to 1. For different diffusion step lengths t, As the diffusion step t increases, it decreases. In this invention, the diffusion step t∈[0,1000] is selected; when t=0, Close to 1, the position information after noise addition lx t Approximately the original position information lx0; when t = 1000, Close to 0, the position information lx after adding noise t Approximately Gaussian noise ∈ satisfies the normal distribution (mean is 0 and variance is 1).

[0053] For lv0, since it is discrete data representing element characteristics, the noise addition formula is:

[0054]

[0055] Among them, lv0 represents the original element characteristics of the real drug molecule, lv t represents the element characteristics of the real drug molecule after adding noise, K represents the number of elements, and g represents the Gumbel distribution with a mean of 0 and a variance of 1. argmax() is the maximum value operation, onehot() is the encoding operation, A constant that controls the degree of noise addition. The value range is from 0 to 1. For different diffusion step lengths t, As the diffusion step t increases, it decreases. In this invention, the diffusion step t∈[0,1000] is selected; when t=0, Close to 1, the element feature lv after adding noise t Approximately the original element feature lv0; when t = 1000, Close to 0, the element feature lv after adding noise t Approximately noise.

[0056] (2) The main task of the training process is to train the prediction network (MPN) based on the state space model. In this module, the present invention uses the position information lx of the real drug molecules after adding noise t , elemental features of real drug molecules after adding noise t , clear target protein molecule position information px0 and clear target protein molecule element information pv0 as input data. The processing details are as follows:

[0057]

[0058] Among them, MLPs() represents the multi-layer perceptron network for feature extraction, m() represents the mamba network structure for state space modeling, and A() represents the attention mechanism. When the attention mechanism uses px0 and lx t As Q, use pv0 and lv t As K and V; when predicting the elemental characteristics of generated drug molecules When the attention mechanism uses pv0 and lv t As Q, use px0 and lx t As K and V. In the training process, the loss function is as follows:

[0059]

[0060] Among them, lx0 represents the position information of the real drug molecule, Indicates the location information of generated drug molecules, lv0 indicates the element characteristics of real drug molecules, Represents the element features of the generated drug molecules. MSEloss() represents the mean square error loss, which is used to calculate the distance of continuous data; KL_loss() represents the KL divergence, which is used to evaluate the difference of discrete data.

[0061] In the third step, deep learning networks based on state-space models and diffusion models are sampled.

[0062] The sampling process of the drug generation method based on the state space model and the diffusion model proposed in the present invention is as follows: Figure 1 Middle part: Sampling process shown.

[0063] (1) First, the present invention obtains the position information px0 and element feature pv0 from the three-dimensional information P0 of the target protein, and initializes the generated drug molecule through the following steps:

[0064]

[0065] Among them, Gaussian(0,1) represents Gaussian noise with mean 0 and variance 1, g represents Gumbel distribution with mean 0 and variance 1, argmax() is the maximum value operation, and onehot() is the encoding operation. represents the position information of the generated drug molecules approximated by noise, Represents the elemental signatures of the generated drug molecules approximated by noise.

[0066] (3) In the denoising process, MPN represents the prediction network based on the state space model trained in the second step, and post() represents: the position information of the generated drug molecules predicted by MPN is converted into and element characteristics As a priori probability, the position information of the posterior probability can be obtained and element characteristics The details are as follows:

[0067]

[0068] Among them, px0 and pv0 are the position information and element characteristics of the clear target protein, respectively. and are the position information and element characteristics of the drug molecules generated in step T. The final and Then: Through the sampling method proposed by the present invention, given the drug molecule data generated in the Tth step and Under the premise of , the drug molecule data generated in step T-1 is sampled.

[0069] (3) Specifically, when T = t, that is, the drug molecule data generated in step T is known and Initialized drug molecule data and The drug molecule data generated in step T-1=t-1 is obtained by sampling and for:

[0070]

[0071] When T = 1, that is, the drug molecule data generated in step T is known and The previous step data for the final drug molecule data generated for the present invention and The drug molecule data finally generated by the present invention is obtained by sampling and for:

[0072]

[0073] In the fourth step, we evaluate the deep learning networks based on the state-space model and the diffusion model.

[0074] The evaluation process of the drug generation method based on the state space model and the diffusion model proposed in the present invention is as follows: Figure 1 The evaluation process is shown in the lower part of the figure. In the evaluation process, the drug molecule data finally generated by the present invention obtained by the sampling process in the third step is mainly evaluated. and First, the present invention needs to compare and evaluate the drug molecule data finally generated. and Molecular reconstruction is performed, and the molecular data of failed reconstruction are discarded. Then the integrity of the successfully reconstructed molecules is checked, and incomplete molecules with free groups are discarded. Finally, the molecular data that pass the above two evaluations are evaluated for results, including: drug-likeness, ease of synthesis, and three molecular docking score evaluations (Vina Score, VinaMin, and VinaDock).

[0075] Step 5: Display the results.

[0076] The drug generation method based on the state space model and the diffusion model proposed in the present invention is trained on the CrossDocked2020 dataset, and the target protein of the CrossDocked2020 dataset is selected as the control condition of the diffusion model to generate the drug molecules. Figure 2 As shown. Among them, drug-likeness (QED, Quantitative Estimate of Drug-likeness) is a method to quantify drug similarity, with a value between 0 and 1. The larger the value, the more similar it is to the drug. Drug-likeness is not a characteristic of chemical structure, but an indicator determined by a combination of several physical parameters. It mainly includes molecular weight (MW), logP (ALOGP), number of hydrogen bond donors (HBDs), number of hydrogen bond acceptors (HBAs), polar surface area (PSA), number of rotatable bonds (ROTBs), number of aromatic rings (AROMs) and number of alert structures (ALERTS). Synthetic Accessibility score (SA, Synthetic Accessibility score) is a method to evaluate the difficulty of drug synthesis. The value is between 0 and 1. The larger the value, the easier it is to synthesize the drug. Vina Score, Vina Min and Vina Dock are all molecular docking scores, which indicate the degree of binding between drug molecules and target proteins. The smaller the value, the stronger the binding force with the target protein. Among them, Vina Score is the molecular docking score of the current posture of the molecule, Vina Min is the molecular docking score of the molecule BFGS local optimization, and Vina Dock is the molecular docking score of the global search optimization.

[0077] Each row of experimental results is from left to right: the target protein as the control condition of the diffusion model, the real drug molecules and their evaluation results, and the generated drug molecules and their evaluation results. Figure 2 The results show that the drug generation method based on the state-space model and the diffusion model proposed in the present invention can generate drug molecules that are more consistent with real-world structural information. Compared with real drug molecules, the drug molecules generated by the method proposed in the present invention have better drug-likeness and ease of synthesis, and have stronger binding ability with the target protein.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A drug generation method based on a state space model and a diffusion model, characterized in that: The method comprises the following steps: a) preparing data, including three-dimensional coordinates and elemental characteristic data of target protein molecules, and three-dimensional coordinates and elemental characteristic data of drug molecules; b) Train the prediction network MPN based on the state space model to transform the position information lx of the real drug molecules after adding noise t , elemental features of real drug molecules after adding noise t , clear target protein molecule position information px0 and clear target protein molecule element information pv0 as input data; the processing details are as follows: Among them, MLPs() represents the multi-layer perceptron network for feature extraction, m() represents the mamba network structure for state space modeling, and A() represents the attention mechanism; when predicting the location information of drug molecules When the attention mechanism uses px0 and lx t As Q, use pv0 and lv t As K and V; when predicting the elemental characteristics of generated drug molecules When the attention mechanism uses pv0 and lv t As Q, use px0 and lx t As K and V; in the training process, the loss function is as follows: Among them, lx0 represents the position information of the real drug molecule, Indicates the location information of generated drug molecules, lv0 indicates the element characteristics of real drug molecules, Represents the element features of generated drug molecules; MSEloss() represents mean square error loss, which is used to calculate the distance of continuous data; KL_loss() represents KL divergence, which is used to evaluate the difference of discrete data; c) a sampling prediction network, which generates the three-dimensional coordinates and elemental characteristic data of the drug molecule using the three-dimensional coordinates and elemental characteristic data of the target protein molecule as control conditions; d) Evaluate the generated drug molecules, including drug-likeness, ease of synthesis, and molecular docking scores; In a), the three-dimensional coordinates of the target protein molecule are the position information data px0 and the element characteristic data pv0 of the target protein molecule, and the three-dimensional coordinates of the real drug molecule are processed into the position information data lx0 and the element characteristic data lv0 of the drug molecule; in subsequent steps, px0, pv0, lx0 and lv0 will be used as input data of the deep learning network; In b), the prediction network includes a multi-layer perceptron network for feature extraction and a mamba network structure for state space modeling; Add noise to the position information data lx0 and element feature data lv0 obtained from the three-dimensional information L0 of the real drug molecule; for the position information lx0, add continuous noise; for the element feature lv0, add uniform discrete noise; The training goal is to train the prediction network MPN based on the state space model, and transform the position information lxt of the real drug molecules after noise addition and the element features lv of the real drug molecules after noise addition. t , clear target protein molecule position information px0 and clear target protein molecule element information pv0 as input data, predict the generated drug molecule data and For location information, the mean square error loss MSEloss is used to calculate the distance of continuous data; for element features, the KL divergence KL_loss is used to evaluate the difference of discrete data; In c), the sampling process includes a denoising process, and the denoising process uses the generated drug molecule data predicted by the prediction network as the prior probability; The position information px0 and element features pv0 are obtained from the three-dimensional information P0 of the target protein, and the generated drug molecules are initialized and represents the position information of the generated drug molecules approximated by noise, represents the elemental features of the generated drug molecules approximated by noise; Perform T = 1000 samplings; in each sampling, use the prediction network MPN based on the state space model trained in the second step to predict the position information of the generated drug molecules and element characteristics And use it as the prior probability to finally get the posterior probability: Generate the position information of the drug molecule at step T and element characteristics In step d), the drug-likeness is evaluated using the quantitative estimation of drug similarity (QED) method.

2. The drug generation method based on the state space model and the diffusion model according to claim 1, characterized in that: In the above d), the ease of synthesizing is evaluated using the synthetic accessibility score SA method.

3. The drug generation method based on the state space model and the diffusion model according to claim 1, characterized in that: In the above d), the molecular docking score is evaluated using Vina Score, Vina Min and Vina Dock methods.

4. The drug generation method based on the state space model and the diffusion model according to claim 1, characterized in that: The three-dimensional coordinates and elemental characteristic data of the target protein molecule are obtained from the .pdb file.

5. The drug generation method based on the state space model and the diffusion model according to claim 1, characterized in that: The three-dimensional coordinates and elemental characteristic data of the drug molecule are obtained from the .sdf file.

6. The drug generation method based on the state space model and the diffusion model according to claim 1, characterized in that: The dataset is the CrossDocked2020 dataset.

Images