A method and system for generating fusion lipophilic drug molecules based on a molecular generation model

Abstract

This invention discloses a method and system for generating fusion lipophilic drug molecules based on a molecular generation model. The invention first determines the lipophilicity (c) of the desired molecule and the size (M) of the generated molecule; then, it uses a normal distribution N... xh An initial pure noise z is obtained by sampling. T Then, an iterative loop of number T is performed, with each loop eliminating the noise predicted at the current time step. This noise is represented by the concatenation of the atomic coordinates and atomic features at the current time step. t The time t and the desired lipophilicity value c of the molecule are input into the molecular generation model to obtain z0, which is the concatenation of the corresponding atomic coordinates and atomic features. Finally, atomic bonding processing is performed on the obtained atomic coordinates and atomic features to obtain a set of generated molecules. This invention proposes a drug generation method that integrates lipophilicity based on a diffusion model. That is, based on the diffusion model, conditional information—namely, molecular lipophilicity—is injected into the network, and this conditional information will be used to guide the generation of drug molecules.

Description

Technical Field

[0001] This invention belongs to the fields of computer science and drug discovery technology, and relates to a method and system for generating drug molecules, specifically a method and system for generating fused lipophilic drug molecules based on a drug molecule generation model. Background Technology

[0002] Drug discovery has always faced numerous obstacles, requiring substantial human, material, and financial resources. The difficulty lies in the vast and discrete search space of chemical molecules. In recent years, with the widespread application of deep learning and the continuous development of generative models, molecular generation has not only reduced the search space for chemical molecules but also decreased the time required for drug discovery. In particular, de novo molecular design—the automatic generation of effective chemical structures with ideal properties—has become a key task in the field of drug discovery. However, existing generative models still have limitations in using drug properties as conditions to guide de novo molecular generation. Therefore, integrating drug properties into molecular generation models is of great significance.

[0003] Drug properties are one of the fundamental issues in cheminformatics research, playing a crucial role in many applications of chemistry and drug discovery. Lipophilicity, as a drug property, refers to the ability of a compound to dissolve in fats, oils, lipids, or nonpolar solvents. Lipophilicity reflects the ability of drug molecules to penetrate cell membranes and, to some extent, reflects the cell's absorption capacity for this compound. Furthermore, lipophilicity is one of the five rules of drug-likeness and a major factor influencing drug-likeness; therefore, it is a critical physical property for developing small molecule drugs. This invention focuses on incorporating lipophilicity into molecular generation models and employing appropriate methods to evaluate the effectiveness of this incorporation. Of course, many other important chemical properties of molecules exist; this invention only explores lipophilicity to provide insights for further research on the integration of other properties.

[0004] Currently, widely used molecular generation models include variational autoencoders, generative adversarial networks, autoregressive models, and diffusion models, which have recently achieved great success in various tasks. Therefore, this invention will incorporate lipophilicity into the molecular generation model, based on the most advanced diffusion model. The diffusion model, also a type of generative model, is inspired by the physical process of gas diffusion and guided by non-equilibrium thermodynamics. It can be represented as two Markov chains: a diffusion process and a reverse process (i.e., a denoising process). In the forward diffusion stage, molecular information data is gradually contaminated with added noise until the input molecular information data becomes completely random noise. The reverse process is a gradual removal of prediction noise, implemented using a series of Markov chains to recover data from Gaussian noise. In the diffusion stage, like a physical phenomenon, the forward diffusion process gradually diffuses the data into a predefined noise distribution. Setting the time from 1 to T, the forward diffusion process is from Z0 to Z... T During the process, noise is gradually added to the data representing the molecular geometry, and Z t+1 It is obtained by adding noise to Zt, and it is only affected by Z. t The diffusion process is a Markov process because of the influence of time steps (t). As the number of steps (t) increases, the final distribution will be closer to a standard Gaussian distribution. The reverse process eliminates the noise added to the data at each time step during the diffusion process. Since this noise is unknown, a network is needed to predict the noise that should be eliminated at each step in the reverse process, and this noise should be as consistent as possible with the noise added in the corresponding step during the forward diffusion process. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and system for generating fused lipophilic drug molecules based on a drug molecule generation model. This method can generate corresponding molecules based on the input expected lipophilicity, and the actual lipophilicity of the generated molecules will be close to the input expected lipophilicity.

[0006] The technical solution adopted by the method of the present invention is: a method for generating fusion lipophilic drug molecules based on a molecular generation model, comprising the following steps:

[0007] Step 1: Determine the lipophilicity c of the desired molecule and the size M of the resulting molecule;

[0008] Step 2: From the normal distribution N xh An initial pure noise z is obtained by sampling. T ;

[0009] Step 3: Concatenate the atom coordinates and atom features at time t (z) tThe time t and the desired lipophilicity value c of the molecule are input into the molecular generation model to predict the noise that should be eliminated, and the noise that should be eliminated at the current time is output.

[0010] Perform T iterations until z0 is obtained, which is the concatenation of the corresponding atomic coordinates and atomic features;

[0011] Step 4: Perform atomic bonding processing on the obtained atomic coordinates and atomic features to form molecules based on these two data; calculate the distance between all pairs of atoms based on the generated atomic coordinates, and then predict whether there are bonds between pairs of atoms and the type of bonds based on these calculated distances and the generated atomic features; finally, a set of generated molecules is obtained.

[0012] Preferably, in step 3, the splicing z-axis of the atomic coordinates and atomic features from the previous moment is calculated. s The formula details are as follows:

[0013]

[0014] Where α in the formula t|s =α t / α s ,and And σ t→s The meaning is σ t|s σ s / σ t ; For molecular generation model; α t σ represents the degree to which the original data is preserved at each step of the diffusion process. t denoted by α, where ∈ represents the degree of noise added at each step of the diffusion process; ∈ represents the noise that needs to be added at time t from the normal distribution; s represents the noise added with respect to α. t The final value that affects the precision, α s σ represents the degree to which the original data is preserved in the s-th step of the diffusion process. s This indicates the degree of noise introduced during the s-th step of the diffusion process.

[0015] Preferably, the molecular generation model is a trained molecular generation model;

[0016] The training process includes the following steps:

[0017] (1) Preparation of training dataset;

[0018] First, the data in the dataset is standardized using SMILES; then, duplicate data is removed, and for data with multiple lipophilicity values, the lipophilicity value that appears most frequently for that molecule is selected; finally, the three-dimensional information of the molecule is constructed, including atomic coordinates and atom types.

[0019] (2) Model training;

[0020] (2.1) Prepare training data, including the atomic coordinates x of the molecule, atomic features h, molecular lipophilicity c, and an isovariant graph neural network used to predict noise during diffusion. Atomic characteristics refer to the type of atoms;

[0021] (2.2) Randomly select time t from the time points (0, ..., T), and sample the noise ∈ to be added at time t from the normal distribution; where T is a preset value;

[0022] (2.3) The splicing z of atomic coordinates and atomic features at time t during diffusion is calculated. t ;

[0023] z t =α t [x,h]+σ t ∈;

[0024] Where, α t σ represents the degree to which the original data is preserved at each step of the diffusion process. t This indicates the degree of noise added at each step in the diffusion process, and sets...

[0025] α t = (1-2s)·f(t)+s;

[0026] f(t) = (1 - (t / T)) 2 );

[0027] In the above formula, s represents the relationship between α and α. t The precision value that ultimately affects the final value;

[0028] (2.4) The z obtained from (2.3) t The current time t and the lipophilicity c of the input data molecules are fed into the equivariant graph neural network as inputs. In this context, the output is the predicted noise at time t.

[0029] During training, the loss function is calculated based on the predicted noise and the actual noise, i.e., minimizing... The gradient of the loss is backpropagated to update the equivariant graph neural network. The possible values ​​of the parameters;

[0030] (2.5) Repeat (2.2) to (2.4) until the molecular generation model converges.

[0031] As a preferred option, the generated molecules obtained in step 3 are further evaluated by using a lipophilicity prediction network to predict the lipophilicity of the generated molecules. The error between the predicted lipophilicity of the generated molecules and the expected lipophilicity of the generated molecules, which is the input condition, is calculated. The smaller the error value, the better. The smaller the error value, the closer the actual lipophilicity of the molecule generated based on the given expected lipophilicity is to the expected value.

[0032] Preferably, the lipophilicity prediction network consists of three parallel encoders: a 3D information encoder, a substructure encoder, and a directional MPNN encoder. The 3D information encoder extracts the 3D geometric information of the molecule, encoding the 3D information given by the pairwise Euclidean distances of all atoms, thereby generating an embedded 3D representation. The substructure encoder uses the generalized features of the extracted substructure to generate an embedded representation through a feedforward neural network, and then performs an averaging operation to obtain the current encoder's embedded representation. The directional MPNN encoder creates a molecular vector embedded representation by passing edge focusing information.

[0033] The lipophilicity prediction network embeds the final molecular vector into a neural network consisting of two fully connected layers, and then outputs the final molecular lipophilicity prediction result.

[0034] The technical solution adopted by the system of the present invention is: a fusion lipophilic drug molecule generation system based on a molecular generation model, characterized in that it comprises:

[0035] One or more processors;

[0036] A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method for generating fusion lipophilic drug molecules based on a molecular generation model.

[0037] Compared to existing technologies, the beneficial effects of this invention are as follows: Based on a diffusion model, this invention injects conditional information—namely, molecular lipophilicity—into the network. This conditional information guides the generation of drug molecules. This method successfully integrates lipophilicity into the molecule generation process, effectively incorporating it into the generated molecular structure. That is, it can generate corresponding molecules based on the input expected lipophilicity, and the actual lipophilicity of the generated molecule will be close to the input expected lipophilicity. This can generate numerous molecules with specific lipophilic properties. In drug development and new drug discovery, this effectively reduces the search space for chemical molecules and the time consumed in drug discovery, significantly contributing to saving human, material, and financial resources in drug discovery. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention.

[0039] Figure 2 This is a structural diagram of the molecular generation model according to an embodiment of the present invention;

[0040] Figure 3 This is a diagram of the lipophilicity prediction network structure according to an embodiment of the present invention;

[0041] Figure 4 This is a two-dimensional structural diagram of a molecule in an embodiment of the present invention.

[0042] Figure 5 This is a three-dimensional conformation diagram of a molecule in an embodiment of the present invention. Detailed Implementation

[0043] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0044] Drug development has always faced numerous obstacles, requiring substantial human, material, and financial resources. The difficulty in drug discovery lies in the vast and discrete search space of chemical molecules. In recent years, with the widespread application of deep learning and the continuous development of generative models, molecular generation has not only reduced the search space of chemical molecules but also decreased the time required for drug discovery. However, existing generative models still have limitations in using drug properties as conditions to guide de novo molecular generation. Therefore, integrating drug properties into molecular generation models is of great significance. Lipophilicity, as a drug property, refers to the ability of a compound to dissolve in fats, oils, lipids, or nonpolar solvents. Lipophilicity reflects the ability of drug molecules to penetrate cell membranes and, to some extent, reflects the cell's absorption capacity for this compound. Furthermore, lipophilicity, as one of the five rules of drug-likeness, is a major factor influencing drug-likeness, making it a crucial physical property for developing small molecule drugs. To address these issues, this invention proposes a drug generation method that integrates lipophilicity based on a diffusion model. Specifically, it injects conditional information—molecular lipophilicity—into the network based on the diffusion model, which guides the generation of drug molecules. Furthermore, a suitable evaluation baseline is proposed that can be used to evaluate the quality of the method of the present invention.

[0045] Please see Figure 1 The present invention provides a method for generating fusion lipophilic drug molecules based on a molecular generation model, comprising the following steps:

[0046] Step 1: Determine the lipophilicity c of the desired molecule and the size M of the resulting molecule;

[0047] Step 2: From the normal distribution N xh An initial pure noise z is obtained by sampling. T ;

[0048] Step 3: Concatenate the atom coordinates and atom features at time t (z) t The time t and the desired lipophilicity value c of the molecule are input into the molecular generation model to predict the noise that should be eliminated, and the noise that should be eliminated at the current time is output.

[0049] Perform T iterations until z0 is obtained, which is the concatenation of the corresponding atomic coordinates and atomic features;

[0050] Step 4: Perform atomic bonding processing on the atomic coordinates and atomic features obtained in Step 3 to form molecules. Calculate the distances between all paired atoms based on the generated atomic coordinates, and then predict the existence and type of bonds between paired atoms based on these calculated distances and the generated atomic features (i.e., atom types). This invention only considers single, double, and triple bonds. Furthermore, a reasonable deviation value is set for the bond lengths of single, double, and triple bonds, allowing the bond lengths to better fit the lipophilic dataset used in the experiment. Finally, a set of generated molecules is obtained.

[0051] In one implementation, step 3 calculates the concatenation z of the atomic coordinates and atomic features from the previous time step. s The formula details are as follows:

[0052]

[0053] Where α in the formula t|s =α t / α s ,and And σ t→s The meaning is σ t|s σ s / σ t ; For molecular generation model; α t σ represents the degree to which the original data is preserved at each step of the diffusion process. t denoted by α, where ∈ represents the degree of noise added at each step of the diffusion process; ∈ represents the noise that needs to be added at time t from the normal distribution; s represents the noise added with respect to α. t The final value is the precision value that affects the outcome. s σ represents the degree to which the original data is preserved in the s-th step of the diffusion process. s This indicates the degree of noise introduced during the s-th step of the diffusion process.

[0054] The molecular generation model in this embodiment has the following overall network architecture: Figure 2 As shown, Figure 2 The image below illustrates a reverse denoising process for generating molecules. To generate molecules, noise z is sampled from a standard normal distribution. t The denoising process begins, ultimately generating the atomic coordinates x and features h of the molecule, which is achieved by analyzing the distribution p(z). t-1 |z t It is achieved through iterative sampling in ). Figure 2 The diagram above illustrates the training and learning process of the model. At a randomly selected time step t, noise is added to the data x and h, and then the network learns to denoise them, i.e., it uses an isotropic graph neural network. (Noise Prediction Network) predicts the noise added at time step t based on the input information, minimizing the difference between the predicted noise and the actual noise added at time step t.

[0055] In one implementation, the molecular generation model is a trained molecular generation model; the training process includes the following steps:

[0056] Step 1: Data preparation. This involves pre-processing the dataset with lipophilic labels used to train the drug molecule generation model. The specific implementation process includes the following sub-steps:

[0057] Step 1.1: SMILES Standardization. The existing dataset consists of a series of SMILES strings and their corresponding lipophilic labels. SMILES is a method of representing chemical molecules using strings. However, because the "order of appearance" of atoms in a molecule can be arbitrarily modified, the same chemical molecule may have multiple different SMILES forms, which greatly complicates molecule matching and deduplication. SMILES standardization ensures that each chemical molecule has only one form of SMILES, guaranteeing SMILES uniqueness.

[0058] Step 1.2: Removal of duplicate data. For datasets that have already been SMILES-normalized, remove duplicate data. For data with multiple lipophilicity values, select the most common lipophilicity value for that molecule.

[0059] Step 1.3: Constructing the 3D Information of the Molecules. The existing dataset lacks 3D information of the molecules. Therefore, a conformation generation algorithm is needed to generate the 3D information of the molecules. This 3D information is not just a set of coordinates. For a given molecular diagram, there can be multiple conformations, meaning there are multiple possible arrangements of atoms, which may lead to different chemical properties. Here, this invention only considers conformations with lower energy levels, because conformations with lower energy levels are naturally more likely to exist. There are various tools that can be used to calculate conformations, from classical force field-based methods to slower but more accurate molecular dynamics simulations. This invention uses the ETKDG algorithm of RDKit to generate the required conformations. The parameters are set to generate 100 corresponding conformations for each molecule, and then the 30 conformations with the lowest energy levels are selected from them. Finally, the required 3D information of the molecules (including atomic coordinates and atom types) is obtained.

[0060] For example, a SMILES is

[0061] The molecule C1CC(C(N([H])[H])(=O)=O)C(N([H])[H])C(CC(OCC)=O)C1CC(O[H])=O has the following two-dimensional structure diagram: Figure 4As shown, one of the three-dimensional configurations generated according to the ETKDG algorithm is as follows: Figure 5 As shown.

[0062] Step 2: Based on the dataset prepared in Step 1, train a drug molecule generation model that incorporates lipophilicity. The specific implementation process is as follows:

[0063] (1) Input the data prepared for training the model, including the atomic coordinates x of the molecule, atomic features h, molecular lipophilicity c, and the isovariant graph neural network used to predict noise during diffusion. Atomic characteristics refer to the type of atoms.

[0064] (2) Randomly select time t from the time intervals (0, ..., T), and sample the noise ∈ to be added at time t from the normal distribution, where T is set to 1000 in this invention.

[0065] Step 2.3: Calculate the splicing z of atomic coordinates and atomic features at diffusion step t using the following formula. t The specific content of the formula is as follows:

[0066] z t =α t [x,h]+σ t ∈

[0067] because Therefore, we only need to define α. t That's sufficient. x, h represent the atomic coordinates and atomic features input in step 2.1, and ∈ represents the noise to be added at time t, sampled from the normal distribution in step 2.2. Furthermore, it's necessary to define α for t = 1, ..., T. t and σ t The values ​​of these two are due to Therefore, we only need to define α. t That's sufficient. Furthermore, during the diffusion process, the principle should be followed: retain more original data information in the early stages of diffusion, and retain less original data information as diffusion progresses. Therefore, α... t The value of should be monotonically decreasing, that is, from α0≈1 to α. T ≈0. Therefore, it is defined as follows in this invention:

[0068] α t = (1-2s)·f(t)+s;

[0069] f(t) = (1 - (t / T)) 2 );

[0070] In the above formula, s represents the relationship between α and α. tThe final value chosen is a precision value that minimizes the impact of numerical instability; in this invention, the value is set to 10. -5 .

[0071] Step 3: Based on the drug molecule generation model that has been trained in Step 2 and incorporates lipophilicity, input the desired lipophilicity and molecular size of the generated molecule. According to the input condition information, the trained drug molecule generation model outputs the generated molecule.

[0072] (1) Input the equivariant graph neural network for noise prediction trained in step 2. And the lipophilicity c of the desired molecule and the size M of the resulting molecule.

[0073] (2) From the normal distribution N xh A pure noise z is obtained by sampling. T .

[0074] (3) Perform an iterative loop T times (t ranges from T to 0). In each iteration, the iterative graph neural network obtained during the training process in step 2 will be used as a reference. To predict the noise that should be eliminated, this type of graph neural network... The concatenation of the atomic coordinates and atomic features at time t is used to construct z. t Taking time t and the desired lipophilicity value c as inputs, the output is the noise portion that should be eliminated at the current time. The concatenation z of the atomic coordinates and atomic features from the previous time step is calculated using the following formula. s The formula details are as follows:

[0075]

[0076] Repeat this iterative process until z0 is obtained, thus obtaining the corresponding atomic coordinates and atomic features.

[0077] (4) The atomic coordinates and atomic features obtained in step 3.3 are processed for atomic bonding, and molecules are formed based on these two factors. The distances between all paired atoms are calculated based on the generated atomic coordinates. Then, based on these calculated distances and the generated atomic features (i.e., atom types), the existence and type of bonds between paired atoms are predicted. In this invention, only single bonds, double bonds, and triple bonds are considered, and all possible single bond lengths, double bond lengths, and triple bond lengths are shown in Tables 1, 2, and 3, respectively, which are derived from common distances in encyclopedias. In addition, a reasonable deviation value is set for the single bond, double bond, and triple bond length, which allows the bond lengths to better fit the lipophilic dataset used in the experiment. In this invention, these deviation values ​​are set to 12, 3, and 3, respectively. Finally, a set of generated molecules is obtained.

[0078] Table 1

[0079]

[0080] Table 2

[0081]

[0082] Table 3

[0083]

[0084] In one implementation, the generated molecule is evaluated to assess the difference between its lipophilicity and the expected lipophilicity, which is used as conditional input. A trained lipophilicity prediction network is used as the evaluation tool to predict the lipophilicity of the generated molecule. The error between the predicted lipophilicity and the expected lipophilicity of the generated molecule, which is used as conditional input, is calculated. A smaller error value is better, indicating that the actual lipophilicity of the molecule generated based on the given expected lipophilicity is closer to the expected value. Therefore, this error value is used to evaluate the drug molecule generation method for fusion lipophilicity.

[0085] In one implementation, please see Figure 3 The lipophilicity prediction network model consists of three parallel encoders. The first encoder extracts the 3D geometry of the molecule, encoding the 3D information given by the pairwise Euclidean distances of all atoms, thus generating an embedded 3D representation. The second encoder uses generalized features of the extracted substructure to generate an embedded representation through a feedforward neural network (FFNN), and then performs an averaging operation to obtain the current encoder's embedded representation. The third encoder is a D-MPNN encoder, which creates a molecular embedded representation by passing edge focusing information.

[0086] The proposed lipophilicity prediction model (hereinafter referred to as the 3D lipophilicity prediction model) connects the embedding representations generated by three parallel encoders to produce the final molecular vector embedding. This final molecular vector embedding integrates the 3D geometric information of the molecule, combines the 2D structural information of the molecule, and also takes into account the substructure effects of the molecule. This vector is used to predict the lipophilicity of the molecule. Finally, in this 3D lipophilicity prediction model, the obtained final molecular vector embedding is input into a neural network consisting of two fully connected layers, and the final molecular lipophilicity prediction result is output.

[0087] In summary, the drug molecule generation method with lipophilicity proposed in this invention can output corresponding molecules based on the expected lipophilicity of the input, which can effectively reduce the search space of chemical molecules in drug discovery and save material, financial and human resources.

[0088] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for generating fusion lipophilic drug molecules based on a molecular generation model, characterized in that, Includes the following steps: Step 1: Determine the lipophilicity c of the desired molecule and the size M of the resulting molecule; Step 2: From the normal distribution Sampling is performed to obtain an initial pure noise. ; Step 3: Concatenate the atomic coordinates and atomic features at the current time t. The time t and the desired lipophilicity value c of the molecule are input into the molecular generation model to predict the noise that should be eliminated, and the noise that should be eliminated at the current time is output. Perform an iterative loop T times until you obtain... , That is, the splicing of the corresponding atomic coordinates and atomic features; This involves calculating and splicing the atomic coordinates and atomic features from the previous moment. The formula is as follows: In the formula ,and ,and The meaning is ; For molecular generation models; This represents the degree to which the original data is preserved at each step of the diffusion process. This indicates the degree of noise added at each step of the diffusion process; To sample the noise needed at time t from a normal distribution; s represents the noise to be added at time t. The final value is the precision value that affects the outcome. This represents the degree to which the original data is retained in the s-th step of the diffusion process. This indicates the degree of noise introduced during the s-th step of the diffusion process; Step 4: Perform atomic bonding processing on the obtained atomic coordinates and atomic features to form molecules based on these two data; calculate the distance between all pairs of atoms based on the generated atomic coordinates, and then predict whether there are bonds between pairs of atoms and the type of bonds based on these calculated distances and the generated atomic features; finally, a set of generated molecules is obtained.

2. The method for generating fusion lipophilic drug molecules based on a molecular generation model according to claim 1, characterized in that: The molecular generation model is a trained molecular generation model; The training process includes the following steps: (1) Preparation of training dataset; First, the data in the dataset is standardized using SMILES; then, duplicate data is removed, and for data with multiple lipophilicity values, the lipophilicity value that appears most frequently for that molecule is selected; finally, the three-dimensional information of the molecule is constructed, including atomic coordinates and atom types. (2) Model training; (2.1) Prepare training data, including the atomic coordinates of molecules. atomic characteristics lipophilicity of molecules And an equivariant graphical neural network used to predict noise during diffusion. Atomic characteristics refer to the type of atoms. (2.2) Randomly select time t from the time intervals (0, ..., T), and sample the noise to be added at time t from the normal distribution. Where T is a preset value; (2.3) The atomic coordinates and atomic features at time t during diffusion are calculated and spliced ​​together. ; in, This represents the degree to which the original data is preserved at each step of the diffusion process, while This indicates the degree of noise added at each step in the diffusion process, and sets... ; ； ； In the above formula, s represents the pair of... The precision value that ultimately affects the final value; (2.4) The result obtained from (2.3) The current time t and the lipophilicity c of the input data molecules are fed into the equivariant graph neural network as inputs. In this context, the output is the predicted noise at time t. During training, the loss function is calculated based on the predicted noise and the actual noise, i.e., minimizing... The gradient of the loss is backpropagated to update the equivariant graph neural network. The possible values ​​of the parameters; (2.5) Repeat (2.2) to (2.4) until the molecular generation model converges.

3. The method for generating fusion lipophilic drug molecules based on a molecular generation model according to claim 2, characterized in that: The generated molecules obtained in step 3 are further evaluated using a lipophilicity prediction network, which predicts the lipophilicity of the generated molecules. The error between the predicted lipophilicity of the generated molecules and the expected lipophilicity of the generated molecules, which is the input condition, is calculated. The smaller the error value, the better, because a smaller error value indicates that the actual lipophilicity of the molecule generated based on the given expected lipophilicity is closer to the expected value.

4. The method for generating fusion lipophilic drug molecules based on a molecular generation model according to claim 3, characterized in that: The lipophilicity prediction network consists of three parallel encoders: a 3D information encoder, a substructure encoder, and a directional MPNN encoder. The 3D information encoder extracts the 3D geometric information of the molecule, encoding the 3D information given by the pairwise Euclidean distances of all atoms, thereby generating an embedded 3D representation. The substructure encoder uses the generalized features of the extracted substructure to generate an embedded representation through a feedforward neural network, and then performs an averaging operation to obtain the current encoder's embedded representation. The directional MPNN encoder creates a molecular vector embedded representation by passing edge focusing information. The lipophilicity prediction network embeds the final molecular vector into a neural network consisting of two fully connected layers, and then outputs the final molecular lipophilicity prediction result.

5. A fusion-based lipophilic drug molecule generation system based on a molecular generation model, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method for generating fusion lipophilic drug molecules based on a molecular generation model as described in any one of claims 1 to 4.

Images