Method and electronic device for generating ligand

Abstract

The embodiments of the present disclosure relate to a method for generating a ligand and an electronic device. The method includes: obtaining a trained ligand generation model, the trained ligand generation model is generated based on the decomposition of the ligand by modeling the atomic position, atomic type and chemical bond; and obtaining a target ligand molecule corresponding to the target protein based on the target protein by using the trained ligand generation model. In this way, the ligand generation model in the embodiments of the present disclosure takes into account the decomposition of the ligand and models the chemical bond at the same time, and can generate a target ligand molecule with stronger affinity.

Description

Technical Field

[0001] The present disclosure relates generally to the fields of computers and bioinformatics, and more particularly to methods and electronic devices for generating ligands. Background Art

[0002] The interaction between biological molecules is an important basis for achieving their biological activity. For example, the human body can produce antibody proteins that bind to invading viruses to inhibit diseases. In biopharmaceutical research, the physical and chemical mechanisms of molecular interactions can be understood by analyzing molecules that are known to bind to each other, thereby helping to design new drug molecules that can bind to certain specific targets. For example, ligand molecules that bind to target proteins can be identified, and then drug design can be performed based on ligand molecules. Therefore, how to efficiently identify ligand molecules that can bind to target proteins is one of the problems that need to be solved at present. Summary of the invention

[0003] According to an exemplary embodiment of the present disclosure, a method for generating a ligand is provided, which obtains a target ligand molecule corresponding to a target protein based on a trained ligand generation model, and can be applied to drug design.

[0004] In a first aspect of an embodiment of the present disclosure, a method for ligand generation is provided, comprising: obtaining a trained ligand generation model, wherein the trained ligand generation model is generated based on the decomposition of the ligand by modeling atomic positions, atomic types, and chemical bonds; and obtaining a target ligand molecule corresponding to the target protein based on the target protein by using the trained ligand generation model.

[0005] In a second aspect of an embodiment of the present disclosure, an electronic device is provided, comprising: at least one processing unit; and at least one memory, wherein the at least one memory is coupled to the at least one processing unit and stores instructions for execution by the at least one processing unit, and when the instructions are executed by the at least one processing unit, the electronic device executes the method described in the first aspect of the present disclosure.

[0006] In a third aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, which has machine-executable instructions stored thereon, and when the machine-executable instructions are executed by a device, the device performs the method described in the first aspect of the present disclosure.

[0007] In a fourth aspect of an embodiment of the present disclosure, a computer program product is provided, comprising computer executable instructions, wherein the computer executable instructions implement the method described according to the first aspect of the present disclosure when executed by a processor.

[0008] In a fifth aspect of an embodiment of the present disclosure, an electronic device is provided, comprising: a processing circuit configured to execute the method described according to the first aspect of the present disclosure.

[0009] The invention summary is provided to introduce a series of concepts in a simplified form, which will be further described in the specific embodiments below. The invention summary is not intended to identify the key features or essential features of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easy to understand through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, wherein:

[0011] Figure 1 A schematic diagram showing the decomposition of a ligand molecule;

[0012] Figure 2 A schematic flow chart of an example training process according to some embodiments of the present disclosure is shown;

[0013] Figure 3 A schematic flow chart showing an example usage process according to some embodiments of the present disclosure;

[0014] Figure 4 shows a schematic diagram of generating a target ligand molecule according to some embodiments of the present disclosure;

[0015] Figure 5 A block diagram showing an example apparatus according to some embodiments of the present disclosure; and

[0016] Figure 6 A block diagram of an example device that may be used to implement embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0017] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as being limited to the embodiments described herein, which are instead provided for a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not intended to limit the scope of protection of the present disclosure.

[0018] In biopharmaceutical research, drug design methods can include ligand-based drug design methods and structure-based drug design methods, depending on the type of data they rely on.

[0019] Ligand-based drug design methods collect a large amount of existing ligand information for a specific target, and based on the structure of the ligand, they can explore the molecular rules of binding to the specific target, thereby designing or optimizing the ligand structure. Ligand-based drug design methods do not rely on the structural data of the target, but only require ligand information, so they can only be applied to targets that already have a large amount of ligand information. Ligand-based drug design models can be constructed based on deep learning, for example, they can be used to generate a one-dimensional simplified molecular linear system (SMILES) or a two-dimensional molecular graph of a molecule.

[0020] The structure-based drug design method can design a 3D ligand structure that can form a tight interaction with the 3D structure according to the three-dimensional (3D) structure of the target. Specifically, the interaction force or degree of adaptation between the ligand and the target can be estimated by the relative position of the atoms between them, so as to design and optimize the 3D ligand structure. The structure-based drug design method takes into account the 3D structure of both the target and the ligand, and realizes the drug molecule design by meeting the appropriate interaction force target, so it can be applied to any target with a 3D structure. A structure-based drug design model can be constructed based on deep learning, for example, using the 3D structure of the target-ligand as training data, and the structure-based drug design model aims to generate a 3D ligand structure within the target binding pocket.

[0021] Structure-based drug design models based on deep learning usually treat the ligand molecule as a whole, and predict the type and position of ligand atoms (such as expressed as coordinates) through generation techniques such as conditional variational autoencoders, autoregression, and diffusion models, thereby completing the design task of 3D ligand molecules. One of the existing structure-based drug design models is a 3D drug design model based on a diffusion model, which uses a diffusion model to model the distribution of atomic types and atomic coordinates in ligand molecules. The 3D drug design model based on the diffusion model has a lower risk of cumulative error generation, and can better utilize the interactions between atoms to generate 3D ligand structures with stronger affinity to the target.

[0022] However, the current scheme still needs to be further improved. For example, the affinity between the generated ligand structure and the target is insufficient and needs to be further enhanced. For example, the 3D drug design model based on the diffusion model assumes that the prior distribution of ligand molecules obeys the standard Gaussian distribution, which leads to insufficient affinity between the generated ligand structure and the target. Due to the lack of direct constraints on the positional relationship between ligand atoms and targets, the sampled ligand molecules may be too close to the target or even collide with it, and the mutual repulsion caused by it will lead to insufficient affinity.

[0023] In view of this, the embodiments of the present disclosure provide a solution for ligand generation to solve the above problems and potential other problems. Specifically, a target ligand molecule corresponding to a target protein can be obtained based on a trained ligand generation model, which can then be applied to drug design.

[0024] As considered in traditional drug design, a ligand molecule can be divided into a skeleton structure and multiple functional groups connected to the skeleton. Exemplarily, the skeleton structure can also be referred to as a skeleton or scaffold, and the functional groups can also be referred to as arms. From the relative position of the ligand binding to the target, the functional groups are usually in close contact with the target surface, which plays a role in improving the activity, while the skeleton supports the functional groups and places the functional groups in the appropriate binding area. Figure 1 A schematic diagram of the decomposition 100 of a ligand molecule is shown. Figure 1 As shown, the ligand molecule is divided into a backbone 120 and arms 140 .

[0025] In an embodiment of the present disclosure, a ligand generation model based on a diffusion model is provided, which can simultaneously generate the atomic type, atomic position, and chemical bond of the ligand molecule, thereby improving the affinity between the ligand and the target.

[0026] The term "target" in the embodiments of the present disclosure may also be referred to as target protein, protein target, receptor protein, receptor, protein, etc. The term "ligand" may also be referred to as ligand molecule, ligand small molecule, etc.

[0027] In the embodiments of the present disclosure, a ligand generation model can be used to generate a target ligand molecule corresponding to a target protein. The ligand generation model is a ligand generation model based on a diffusion model, which is generated based on the decomposition of atoms in the ligand, and thus can also be referred to as a decomposable diffusion (DecompDiff) model in the present disclosure.

[0028] Figure 2 A schematic flow chart of an example training process 200 according to some embodiments of the present disclosure is shown. In block 210, a training data set including a plurality of training data items is obtained, each of the plurality of training data items including a target sample and a corresponding ligand sample. In block 220, a trained ligand generation model is generated through training based on the training data set.

[0029] The training data set may be referred to as a training set or a data set, which includes multiple data items, each of which includes a target sample and a corresponding ligand sample, that is, a pair of target samples and ligand samples, which may also be referred to as a target-ligand pair.

[0030] In some embodiments of the present disclosure, a decomposition prior distribution can be determined based on target samples and ligand samples; an atomic neighbor graph is constructed, which characterizes the relationship between multiple adjacent atoms; a fully connected graph is constructed, which characterizes the relationship between different atoms of the ligand sample; and a trained ligand generation model is generated through training based on the decomposition prior distribution, the atomic neighbor graph and the fully connected graph.

[0031] The process of determining the decomposition prior distribution can be called the data preparation stage, for example, the decomposition prior of the backbone-arm can be obtained. For example, the position of the local binding pocket on the target sample can be determined; based on the binding of the ligand sample to the local binding pocket on the target sample, multiple atoms of the ligand sample are divided into multiple atomic clusters; and the mean matrix and covariance matrix of the atomic positions in the multiple atomic clusters are further determined to obtain the decomposition prior distribution.

[0032] Alternatively, an existing target binding site prediction method can be used to determine the location of the local binding pocket on the target sample. Exemplarily, the local binding pocket on the target sample is determined based on geometric information (such as curvature, etc.), chemical information (such as affinity, etc.), etc. For example, the location of the local binding pocket on the target sample can be extracted using target binding site prediction software such as AlphaSpace.

[0033] Optionally, the atoms of the ligand sample can be divided by a clustering algorithm according to the binding between each atom in the ligand sample and the local binding pocket on the target sample. For example, atoms occupying the same local binding pocket can be divided into the same arm atom cluster, and atoms not occupying the local binding pocket can be divided into skeleton atom clusters. For example, the divided multiple atom clusters include a skeleton atom cluster and multiple arm atom clusters.

[0034] Optionally, the mean and covariance matrix of the atomic coordinate distribution in each atomic cluster can be obtained based on the positions (e.g., coordinates) of the atoms in each atomic cluster by methods such as maximum likelihood estimation. In this way, the combination of the mean matrix and covariance matrix of multiple atomic clusters can be understood as a decomposed prior distribution. Or it can be understood that the decomposed prior distribution of the ligand sample in a certain data item includes the mean matrix and covariance matrix of each atomic cluster in the decomposed multiple atomic clusters, such as multiple prior distributions corresponding to multiple atomic clusters, including a skeleton prior distribution and multiple arm prior distributions. Assuming that the total number of multiple atomic clusters is a positive integer K, the mean matrix of the kth atomic cluster in the multiple atomic clusters can be expressed as μ k , the covariance matrix of the kth atomic cluster is expressed as ∑ k , where k is a positive integer between 1 and K.

[0035] The process of constructing a neighbor graph and a fully connected graph can be referred to as the stage of extracting data features. Exemplarily, the nodes of the atomic neighbor graph may represent atoms of a target sample or atoms of a ligand sample. Optionally, the features of the nodes corresponding to the atoms of the target sample in the atomic neighbor graph include at least one of the following items of the atom: atom type, atom position, amino acid type, whether it is located on the main chain, and within the predetermined range of which arms. Optionally, the features of the nodes corresponding to the atoms of the ligand sample in the atomic neighbor graph include at least one of the following items of the atom: atom type, atom position, which arm it belongs to, or which skeleton it belongs to. Optionally, the features of the edges of the atomic neighbor graph include the distance between the two atoms connected by the edge and the type of edge, wherein the type of edge includes any of the following items: target and ligand edge, target and target edge, ligand and target edge, or ligand and ligand edge. Exemplarily, the features of a node of a fully connected graph include at least one of the following items of an atom: atom type, atom position, whether it belongs to a skeleton, or whether it belongs to an arm, and the features of an edge of a fully connected graph include a chemical bond type and an indication of whether the two atoms connected by the edge belong to the same prior distribution.

[0036] Optionally, the characteristics of the atoms in the target sample can be determined, expressed as It represents one or more of the following: atom type (e.g., one of H, C, N, O, S, Se, F, P, Cl, etc.), atom position (e.g., expressed as coordinates), amino acid type (e.g., one of 20 types), whether the atom is located on the main chain, and the predetermined range of which arms the atom is located (e.g., ) etc.

[0037] Alternatively, the characteristics of the atoms in the ligand sample can be determined, expressed as It represents one or more of the following: atom type (e.g., one of H, C, N, O, S, Se, F, P, Cl, etc.), atom position (e.g., expressed as coordinates), whether the atom belongs to an arm or a backbone, and the arm or backbone to which it belongs. For example, if an atom is aligned with the center of any arm prior distribution (μ k ) is lower than the threshold (such as ), then it can be determined that the atom belongs to the arm, otherwise it can be determined that the atom belongs to the skeleton.

[0038] Based on the characteristics of atoms in the target sample and the characteristics of the atoms in the ligand sample To generate the nearest neighbor graph and the fully connected graph. Optionally, a k-nearest neighbor graph can be constructed on the atoms in the target sample and the atoms in the ligand sample, expressed as For example, taking a certain atom (target atom or ligand atom) as the center, determine the k atoms that are the neighbors of the atom at the center, and construct a k-nearest neighbor graph. In the example, nodes are atoms, and node features are the aforementioned atomic features. The edge between two nodes can be characterized by the joint embedding of the distance between the two atoms represented by the two nodes and the edge type, where the edge type can be one of the four types: target-ligand edge, target-target edge, ligand-target edge, and ligand-ligand edge. Optionally, a fully connected graph can be constructed on the atoms in the ligand sample, represented as In the figure In the example, nodes are atoms in the ligand sample, and node features are the aforementioned atomic features. The characteristics of the edge between two nodes include the chemical bond type and whether the atoms come from the same prior distribution, where the chemical bond type can be one of the five types: single bond, double bond, triple bond, aromatic bond, and no chemical bond.

[0039] Alternatively, the total number of atom types can be expressed as K a , the total number of chemical bonds is represented by K b It is understood that the types listed above are only for illustration, and in actual scenarios, the total number of types may be more or less. For example, in some examples, the set of atom types may include 7 types of atoms (i.e., K a =7). The present disclosure is not limited thereto.

[0040] The ligand generation model in the embodiment of the present disclosure is constructed based on the diffusion model, which involves a forward denoising stage and a reverse denoising node. In the model training stage in the present disclosure, that is, based on the decomposition of the prior distribution, the atomic neighbor graph and the fully connected graph, the process of generating a trained ligand generation model through training may include a denoising stage and a denoising stage. Both the denoising stage and the denoising stage include multiple steps, assuming that one of the steps is the tth step. And in the denoising stage, the tth step is determined based on the t-1th step; and in the denoising stage, the t-1th step is determined based on the tth step.

[0041] In some embodiments, noisy atomic positions, noisy atomic types, and noisy chemical bonds can be obtained through a noisy process based on atomic positions, atomic types, and chemical bonds in the ligand sample.

[0042] Assume that the atomic position, atomic type, and chemical bond at the tth step of the noise addition phase are represented by x t 、v t and b t Optionally, corresponding Gaussian noise can be applied to the atomic position of the t-1th step, and corresponding white noise can be applied to the atomic type and chemical bond of the t-1th step, respectively. For example, the manner of applying noise to each atom in the ligand sample can be expressed as the following formulas (1)-(3):

[0043]

[0044]

[0045]

[0046] In equations (1)-(3), q represents the noise addition function; Represents the atomic positions, atomic types, and chemical bonds of the target sample; represents Gaussian distribution; represents the atomic position of the decentralized ligand, i.e. η k represents the prior distribution to which the atom belongs, that is, C represents discrete category distribution; β t represents the amplitude of the noise applied by the noise adding process. Optionally, β t is a known constant set in advance. Optionally, although the noise amplitude in equations (1)-(3) is β t , but in practical applications, different loading amplitudes can be used, such as β in formula (1) t Not equal to β in equation (2) or (3) t .

[0047] If the total number of steps is assumed to be T, that is, the maximum value of t is T, such as T = 1000 or other values. Then the noisy atomic positions, noisy atomic types and noisy chemical bonds obtained by the noisy stage can be represented as x T 、v T and b T .

[0048] In some embodiments, the atomic neighbor graph and the fully connected graph are input into a group equivariant graph neural network, and the noisy atomic positions, noisy atomic types, and noisy chemical bonds are updated to obtain denoised atomic positions, denoised atomic types, and denoised chemical bonds.

[0049] Exemplarily, the group equivariant graph neural network may include a plurality of modules (e.g., 6) stacked in sequence, each of the plurality of modules including a first neural network, a second neural network, and a third neural network, wherein the first neural network is used to update features corresponding to atomic types, the second neural network is used to update features corresponding to chemical bonds, and the third neural network is used to update features corresponding to atomic positions.

[0050] The group equivariant graph neural network may be an SE(3) equivariant graph neural network, and the first neural network, the second neural network and the third neural network may be represented in sequence as: and For example, the neighbor graph and the fully connected graph The SE(3) equivariant graph neural network is input to alternately update the features corresponding to the atomic type (e.g., called atomic features), the features corresponding to the chemical bonds (e.g., called chemical bond features), and the features corresponding to the atomic positions (i.e., atomic positions).

[0051] For example, the neighbor graph and the fully connected graph The update process in a module may include: (1) using a first neural network to update the node features and edge features; wherein the node features may be represented as scalar features and vector features of the node (i.e., atom). Integrate each feature (including the scalar feature of the atom, the vector feature, and the feature of the edge) and update the atomic feature; (2) through the second neural network Integrate each feature (including updated atomic features (scalar features, vector features) and edge features) to update the chemical bond features; and (3) through a third neural network Integrate each feature (including updated atomic features (scalar features, vector features) and edge features) to update the atomic position of the ligand. For example, in (1), the scalar features and vector features corresponding to the atoms in the target and ligand can be extracted through a neural network, and in (2), the chemical bond features between them can be updated based on the features of each atom in the ligand.

[0052] Optionally, the first neural network Second neural network And the third neural network Each may include an attention mechanism (such as the attention mechanism in the figure) for the above-mentioned feature integration.

[0053] Optionally, the first neural network Second neural network And the third neural network It can also be referred to as the three layers of the module, for example, respectively referred to as the atomic feature update layer, the chemical bond feature update layer, and the position update layer. For example, the dimension of the feature vector in the layer can be set to 128 or other values, and the number of attention heads in the layer can be set to 16 or other values.

[0054] In this way, after passing through multiple (for example, 6) modules in sequence, updated atomic positions, updated atomic types, and updated chemical bonds can be obtained. For example, the updated atomic positions after multiple modules are the denoised atomic positions, which are expressed as For example, the updated atom types after multiple modules are passed through several (e.g., 2) multilayer perceptrons to obtain the denoised atom types, which can be expressed as For example, the updated chemical bonds after multiple modules are passed through several (e.g., 2) multi-layer perceptrons to obtain denoised chemical bonds, which can be expressed as

[0055] It can be understood that the denoising stage involves multiple steps (such as T steps), each of which can include the update process of the SE (3) equivariant graph neural network combined with multiple modules, for example, the output of the t-th step is represented by the input of the t-1-th step. Accordingly, after T steps, the denoised atomic positions can be obtained. Denoised atom types and denoised chemical bonds

[0056] In some embodiments, a loss function may be constructed during the training process, and whether the training process is completed is determined based on the loss function. Exemplarily, the loss function may include a first loss function, a second loss function, and a third loss function. Optionally, the first loss function, the second loss function, and the third loss function may be represented as L x , L v and L b The first loss function can be constructed based on the atomic positions in the ligand sample and the denoised atomic positions, the second loss function can be constructed based on the atomic types in the ligand sample and the denoised atomic types, and the third loss function can be constructed based on the chemical bonds in the ligand sample and the denoised chemical bonds. For example, the first loss function, the second loss function and the third loss function can be obtained by the following formulas (4)-(6), respectively:

[0057]

[0058]

[0059]

[0060] In formula (5), c(v t ,v0) represents the posterior probability of the atom type. In formula (6), c(b t , b0) represents the posterior probability of the chemical bond. Optionally, the loss function used for training can be determined by the following formula (7):

[0061] L=L x +γ a L v +γ b L b (7)

[0062] In formula (7), γ a and γ bFor weights, illustratively, the weights can be updated by a gradient descent method during the training process until convergence. For example, the gradient descent method may include an Adam adaptive learning rate gradient descent algorithm.

[0063] References Figure 2 An example training process of a ligand generation model is described. The trained ligand generation model can be used to generate target ligand molecules corresponding to target proteins. In the training process, a prior distribution based on the decomposition characteristics of the ligand molecules is introduced to explicitly model chemical bonds. Specifically, in the embodiments of the present disclosure, three types of information, namely atomic position, atom type and chemical bond, are generated and modeled based on a diffusion model. By decomposing into multiple atomic clusters and considering the type of chemical bonds between atoms, the affinity between the ligand and the target can be more explicitly enhanced to alleviate the problem of insufficient rationality. In the following, the following will be combined with the above examples. Figure 3 and Figure 4 Describe the schematic usage process of the ligand generation model.

[0064] Figure 3 A schematic flow chart of an example usage process 300 according to some embodiments of the present disclosure is shown. In block 310, a trained ligand generation model is obtained, where the trained ligand generation model is generated by modeling atomic positions, atomic types, and chemical bonds based on the decomposition of the ligand. In block 320, a target ligand molecule corresponding to the target protein is obtained based on the target protein by using the trained ligand generation model.

[0065] It is understood that the trained ligand generation model can be obtained by Figure 2 In the embodiments of the present disclosure, the target protein may be input by a user, or may be obtained from a storage device.

[0066] Exemplarily, the trained ligand generation model can be represented as in represents the initial data input into the model, and M0=(x0,v0,b0) represents the output data output by the model.

[0067] In some embodiments of the present disclosure, initial data can be determined based on the target protein, such as the initial data including initial atomic positions, initial atomic types, and initial chemical bonds. The initial data can then be input into a trained ligand generation model to obtain output data, such as the output data including output atomic positions, output atomic types, and output chemical bonds. Exemplarily, the output data can be a target ligand molecule corresponding to the target protein.

[0068] Exemplarily, the decomposed prior distribution of the backbone and the arms in the initial ligand molecule can be determined based on the target protein; and then the initial data can be determined by sampling based on the decomposed prior distribution. Optionally, the decomposed prior distribution includes a mean matrix and a covariance matrix of multiple atomic clusters. For example, the decomposed prior distribution includes multiple prior distributions corresponding to multiple atomic clusters, each prior distribution including a mean matrix and a covariance matrix of the corresponding atomic cluster.

[0069] In some embodiments, it can be determined whether there is a known ligand molecule for the target protein. If so, the known ligand molecule is used as the initial ligand molecule. Exemplarily, a known ligand molecule bound to the target protein can be obtained; based on the binding between multiple atoms of the known ligand molecule and the binding pocket of the target protein, the multiple atoms are decomposed into multiple atomic clusters, wherein one atomic cluster in the multiple atomic clusters is a skeleton atomic cluster, and the remaining atomic clusters in the multiple atomic clusters are arm atomic clusters; based on the coordinates of the atoms in each atomic cluster of the multiple atomic clusters, the mean matrix and covariance matrix of each atomic cluster of the multiple atomic clusters are determined; and based on the mean matrix and covariance matrix of the multiple atomic clusters, the decomposition prior distribution is determined. It can be understood that the process of determining the decomposition prior distribution based on the target protein and the known ligand molecule is similar to the process of determining the decomposition prior distribution based on the target sample and the ligand sample in the above-mentioned training process. To avoid repetition, it will not be repeated here.

[0070] If there is no known ligand molecule, then the pseudo-ligand molecule can be determined first, and the pseudo-ligand molecule can be used as the initial ligand molecule. Exemplarily, the pseudo-ligand molecule that binds to the target protein can be determined by prediction. For example, the pseudo-ligand molecule can be obtained by a prediction algorithm, such as using AlphaSpace2. Further, the decomposition prior distribution can be determined based on the pseudo-ligand molecule. Exemplarily, based on the binding pocket of the target protein, multiple pseudo-ligand atoms in the pseudo-ligand molecule can be determined, for example, multiple pseudo-ligand atoms are beta atoms; based on the distance between the multiple pseudo-ligand atoms, the multiple pseudo-ligand atoms are decomposed into multiple atomic clusters, wherein one atomic cluster in the multiple atomic clusters is a skeleton atomic cluster, and the remaining atomic clusters in the multiple atomic clusters are arm atomic clusters; based on the coordinates of the atoms in each atomic cluster of the multiple atomic clusters, the mean matrix and covariance matrix of each atomic cluster of the multiple atomic clusters are determined; and based on the mean matrix and covariance matrix of the multiple atomic clusters, the decomposition prior distribution is determined. For example, Beta atoms can be retrieved from the vicinity of the binding pocket of the target protein, and the Beta score can be calculated to indicate the ability of the corresponding Beta atom to pair with the binding pocket. Beta atoms can be divided into multiple atomic clusters by clustering based on the distance between the Beta atoms, for example, if the distance is less than a certain threshold, The two Beta atoms belong to the same atomic cluster. And the mean of the Beta scores of all Beta atoms in the same atomic cluster can be the Beta score of the atomic cluster. Subsequently, multiple atomic clusters can be divided into skeleton atomic clusters and arm atomic clusters by filtering. For each atomic cluster, the respective mean matrix and covariance matrix are determined to obtain the decomposed prior distribution.

[0071] In some embodiments, the initial data may be determined by sampling. For example, the decomposed prior distribution may be sampled to determine the initial atomic position; a set of multiple atomic types may be uniformly sampled to determine the initial atomic type; and a set of multiple chemical bonds may be uniformly sampled to determine the initial chemical bond. For example, the initial atomic position, the initial atomic type, and the initial chemical bond are represented as x, respectively. T 、v T and b T , and satisfies v T ~Uniform(K a ) and b T ~Uniform(K b ), where “Uniform” represents a function that is randomly and uniformly sampled.

[0072] In some embodiments, the process of obtaining output data may include: inputting initial data into a trained ligand generation model, and obtaining output data through a multi-step denoising process. Exemplarily, in each step of the multi-step denoising process: inputting the atomic position, atomic type, and chemical bond obtained in the previous step into a group equivariant graph neural network, and obtaining intermediate atomic positions, intermediate atomic types, and intermediate chemical bonds through denoising; and applying position gradient guidance to the intermediate atomic position, obtaining the atomic position, atomic type, and chemical bond for the next step based on the intermediate atomic position, intermediate atomic type, and intermediate chemical bond.

[0073] Assuming that the multi-step denoising process involves T steps, then based on the output of step t, the input of step t-1 can be determined. Assuming that the output of step t includes x t ,v t ,b t , which can be output to the SE(3) equivariant graph neural network to denoise the atomic positions, atomic types, and chemical bonds, and obtain and As mentioned above combined Figure 2 As shown in the training process, the SE(3) equivariant graph neural network includes multiple modules, each module includes three layers, namely the first neural network Second neural network And the third neural network Get here and The process is similar to the corresponding process in the aforementioned training process, and will not be described here to avoid repetition.

[0074] For example, in the multiple modules of SE(3) equivariant graph neural network, we obtain and Based on the position guidance, position guidance can be further applied. Exemplarily, position gradient guidance is used to apply position constraints based on the positions of the intermediate atoms so that the distance between the bonded backbone atoms and the arm atoms is between the minimum chemical bond length and the maximum chemical bond length, and does not collide with the atoms of the target protein.

[0075] Position guidance can also be called rational guidance. By introducing rational guidance, the quality and accuracy of the generated target ligand molecules can be improved. In some examples, a gradient guidance as shown in the following formula (8) can be applied to the atomic position based on the chemical bond length:

[0076]

[0077] In formula (8) It is the following formula (9):

[0078]

[0079] In equations (8) and (9), Indicates x t The gradient of represents the distance of the atom at step t, represents the set of arm atoms, represents the set of ligand atoms included in the nth arm atom cluster, represents the set of backbone atoms, ρ min and ρ max They represent the minimum and maximum reasonable chemical bond lengths respectively, and ξ1 and ξ2 are hyperparameters greater than 0.

[0080] Thus, by imposing constraints on the atomic positions as shown in formula (8), the distance between the arm atoms and the backbone atoms in the ligand molecule can be made approximately a reasonable chemical bond length, thereby encouraging the arm atoms and the backbone atoms to connect to form a complete molecule.

[0081] In some examples, a gradient guidance as shown in the following formula (10) can be applied to the atomic positions of the ligand based on the atoms of the target protein:

[0082]

[0083] In formula (10), S is as follows:

[0084]

[0085] In equations (10) and (11), σ, γ, and ξ3 are hyperparameters. is the atomic position of the jth target protein, N P Represents the total number of atoms on the target protein, N M Represents the total number of atoms on the ligand molecule.

[0086] In this way, by imposing constraints on the atomic positions as shown in formula (10), spatial collisions between atoms in the generated ligand molecule and the target protein can be avoided.

[0087] In some examples, at step t out of T steps, and Based on this, by imposing constraints on the atomic positions, we can obtain x for the t-1 step t-1 ,v t-1 ,b t-1 .

[0088] Accordingly, it can be understood that similar operations are performed for each step from step T to step 0, and output data M0 = (x0, v0, b0) can be obtained. Then, a molecule can be constructed based on the output atomic position x0, the output atomic type v0 and the output chemical bond b0 to obtain the target ligand molecule.

[0089] In some examples, the model usage process 300 may also be referred to as a model inference process, or a DecompDiff sampling process, etc., which is not limited in the present disclosure.

[0090] Figure 4 Schematic diagram of generating target ligand molecule 400 according to some embodiments of the present disclosure is shown. Figure 4 As shown, for the target protein 4012, an initial ligand molecule (such as a known ligand molecule or a pseudo-ligand molecule, also referred to as an initial ligand molecule) can be determined, and a decomposed prior distribution 401 of the initial ligand molecule can be determined, for example, including multiple prior distributions corresponding to multiple atomic clusters. Figure 4 In the figure, multiple atomic clusters are circled by curved ellipses, for example, multiple arm atomic clusters 4013-4015 and a skeleton atomic cluster 4016, wherein the five-pointed star located at the center of the ellipse represents the mean (a priori center) of the arm atomic cluster, and the inverted triangle represents the mean (a priori center) of the skeleton atomic cluster.

[0091] Furthermore, the corresponding output data can be obtained through the trained ligand generation model 412. In this process, multiple steps (such as T=1000) are involved, such as the output of the tth step is the input of the t-1th step. As an example, Figure 4 For example, the Tth step 414 is shown in FIG.

[0092] In the Tth step 414, the neighbor graph and the fully connected graph can be determined based on the decomposed prior distribution 401 of the target target protein 4012 and the initial ligand molecule, for example, the nodes of the neighbor graph 415-416 include atoms in the target target protein 4012 and atoms in the initial ligand molecule, for example, the nodes of the fully connected graph 417 are atoms in the initial ligand molecule. And in the Tth step 414, position guidance is also applied to the atomic position. For example, at 422, there may be a situation where the ligand atom and the atom in the target target protein 4012 collide in space, so a position constraint can be applied to keep the ligand atom away from the atom in the target target protein 4012. For example, at 424, the distance between the two ligand atoms may be greater than the maximum chemical bond length, so a position constraint can be applied to reduce the distance between the two atoms.

[0093] After T steps, output data can be obtained, for example, ligand molecule 4024 can be determined. It can be seen that, compared with the initial ligand molecule at 401, the ligand molecule 4024 obtained by the trained ligand generation model of the embodiment of the present disclosure has a stronger affinity with the target protein 4012.

[0094] Table 1

[0095]

[0096]

[0097] Table 1 shows example results of the effect of the scheme using the embodiments of the present disclosure. The baseline model in Table 1 represents the target diffusion (TargetDiff) model, which models the atomic position and atom type based on the diffusion model. The score value in Table 1 is the average score of the Wiener docking score (VinaDocking Score) on the CrossDocked2020 dataset. Since the lower the score value, the better the model, it can be seen that this scheme (score value -8.39) is better than the baseline model (score value -7.8). Therefore, the scheme using the embodiments of the present disclosure can obtain target ligand molecules with stronger affinity.

[0098] In addition, the scheme using the embodiments of the present disclosure has a higher proportion of drug-likeness and synthesizability that meet the requirements of drug-making. For example, on the cross-docking 2020 dataset, an average quantitative estimate of drug-likeness (QED) of 0.45 and a synthetic accessibility (SA) of 0.61 can be achieved. Taking into account the success rate indicators of Wiener docking score, QED and SA, it can reach 24.5%. Therefore, the scheme of the embodiments of the present disclosure has certain rationality.

[0099] In this way, the embodiments of the present disclosure provide a scheme for ligand generation, and based on the trained ligand generation model, a target ligand molecule corresponding to the target protein can be obtained. Exemplarily, in the embodiments of the present disclosure, the input is determined based on the decomposition of the prior distribution, for example, the Wiener docking score can be increased to -9.08. Exemplarily, in the embodiments of the present disclosure, chemical bonds are modeled, which can significantly improve QED and SA, for example, the success rate is increased to 15.38%. Exemplarily, in the embodiments of the present disclosure, guidance / constraints are applied to the atomic positions, for example, the Wiener docking score can be increased to -6.11, and the complete ratio can be increased to 0.94, where the complete ratio represents the probability of connecting all atoms in the generated ligand to form a legal molecule.

[0100] It should be understood that in the embodiments of the present disclosure, "first", "second", "third", etc. are only used to indicate that multiple objects may be different, but at the same time do not exclude that two objects are the same, and should not be interpreted as any limitation on the embodiments of the present disclosure.

[0101] It should also be understood that the methods, situations, categories and divisions of the embodiments in the present disclosure are only for the convenience of description and should not constitute special limitations. The features in various methods, categories, situations and embodiments can be combined with each other when it is logical.

[0102] It should also be understood that the above content is only to help those skilled in the art better understand the embodiments of the present disclosure, rather than to limit the scope of the embodiments of the present disclosure. Those skilled in the art may make various modifications, changes or combinations based on the above content. Such modifications, changes or combinations are also within the scope of the embodiments of the present disclosure.

[0103] It should also be understood that the description of the above content focuses on emphasizing the differences between the various embodiments, and the same or similar points can be referenced or borrowed from each other. For the sake of brevity, they will not be repeated here.

[0104] Figure 51 shows a schematic block diagram of an example device 500 according to some embodiments of the present disclosure. The device 500 may be implemented in software, hardware, or a combination of both. Figure 5 As shown, the device 500 includes a model acquisition unit 510 and a ligand generation unit 520 .

[0105] The model acquisition unit 510 is configured to acquire a trained ligand generation model, which is generated by modeling atomic positions, atomic types, and chemical bonds based on the decomposition of the ligand. The ligand generation unit 520 is configured to obtain a target ligand molecule corresponding to the target protein based on the target protein by using the trained ligand generation model.

[0106] In some embodiments, the ligand generation unit 520 may include a decomposition prior distribution determination unit, an initial data determination unit, an output data determination unit, and a target ligand generation unit. The decomposition prior distribution determination unit is configured to determine the decomposition prior distribution of the skeleton and the arm in the initial ligand molecule based on the target target protein, and the decomposition prior distribution includes the mean matrix and covariance matrix of multiple atomic clusters. The initial data determination unit is configured to determine the initial data based on the decomposition prior distribution by sampling, and the initial data includes the initial atomic position, the initial atomic type, and the initial chemical bond. The output data determination unit is configured to input the initial data into the trained ligand generation model to obtain output data, and the output data includes the output atomic position, the output atomic type, and the output chemical bond. The target ligand generation unit is configured to generate the target ligand molecule based on the output data.

[0107] The initial ligand molecule may include a known ligand molecule that binds to a target protein, and the decomposition prior distribution determination unit may be configured to: obtain a known ligand molecule that binds to a target protein; based on the binding conditions between multiple atoms of the known ligand molecule and a binding pocket of the target protein, decompose the multiple atoms into multiple atomic clusters, wherein one atomic cluster in the multiple atomic clusters is a skeleton, and the remaining atomic clusters in the multiple atomic clusters are arms; based on the coordinates of atoms in each atomic cluster of the multiple atomic clusters, determine the mean matrix and covariance matrix of each atomic cluster of the multiple atomic clusters; and determine the decomposition prior distribution based on the mean matrix and covariance matrix of the multiple atomic clusters.

[0108] The initial ligand molecule may include a pseudo-ligand molecule that binds to a target protein, and the decomposition prior distribution determination unit may be configured to: determine the pseudo-ligand molecule that binds to the target protein by prediction; determine multiple pseudo-ligand atoms in the pseudo-ligand molecule based on the binding pocket of the target protein; decompose the multiple pseudo-ligand atoms into multiple atomic clusters based on the distances between each of the multiple pseudo-ligand atoms, wherein one atomic cluster in the multiple atomic clusters is a skeleton and the remaining atomic clusters in the multiple atomic clusters are arms; determine the mean matrix and covariance matrix of each atomic cluster of the multiple atomic clusters based on the coordinates of atoms in each atomic cluster of the multiple atomic clusters; and determine the decomposition prior distribution based on the mean matrix and covariance matrix of the multiple atomic clusters.

[0109] Exemplarily, the initial data determination unit can be configured to: sample the decomposed prior distribution to determine the initial atomic position; uniformly sample a set of multiple atomic types to determine the initial atomic type; and uniformly sample a set of multiple chemical bonds to determine the initial chemical bond.

[0110] Exemplarily, the output data determination unit may be configured to: input the initial data into a trained ligand generation model, and obtain the output data through a multi-step denoising process.

[0111] The trained ligand generation model includes a group equivariant graph neural network, and the output data determination unit can be configured to: in each step of the multi-step denoising process: input the atomic position, atom type and chemical bond obtained in the previous step into the group equivariant graph neural network, and obtain the intermediate atomic position, intermediate atom type and intermediate chemical bond through denoising; and obtain the atomic position, atom type and chemical bond for the next step based on the intermediate atomic position, intermediate atom type and intermediate chemical bond by applying position gradient guidance to the intermediate atomic position.

[0112] Optionally, position gradient guidance is used to impose position constraints based on the intermediate atom positions so that the distance between the bonded backbone atoms and the arm atoms is between the minimum chemical bond length and the maximum chemical bond length and does not collide with the atoms of the target protein.

[0113] like Figure 5 As shown, the device 500 may further include a training unit 505, which is configured to obtain a training data set including multiple training data items, each of the multiple training data items including a target sample and a corresponding ligand sample; and generate a trained ligand generation model through training based on the training data set.

[0114] In some embodiments, the training unit 505 can be specifically configured to: determine the decomposition prior distribution based on the target sample and the ligand sample; construct an atomic neighbor graph, which characterizes the relationship between multiple neighboring atoms; construct a fully connected graph, which characterizes the relationship between different atoms of the ligand sample; and generate a trained ligand generation model through training based on the decomposition prior distribution, the atomic neighbor graph and the fully connected graph.

[0115] The trained ligand generation model includes a group equivariant graph neural network, and the training unit 505 can be specifically configured as follows: based on the atomic positions, atomic types and chemical bonds in the ligand sample, through a noisy process, to obtain noisy atomic positions, noisy atomic types and noisy chemical bonds; and inputting the atomic neighbor graph and the fully connected graph into the group equivariant graph neural network, and updating the noisy atomic positions, noisy atomic types and noisy chemical bonds to obtain denoised atomic positions, denoised atomic types and denoised chemical bonds.

[0116] Exemplarily, the group equivariant graph neural network includes a plurality of modules stacked in sequence, each of the plurality of modules includes a first neural network, a second neural network and a third neural network, and wherein the first neural network is used to update features corresponding to atomic types, the second neural network is used to update features corresponding to chemical bonds, and the third neural network is used to update features corresponding to atomic positions.

[0117] Optionally, the features of the nodes corresponding to the atoms of the target sample in the atom neighbor graph include at least one of the following items of the atom: atom type, atom position, amino acid type, whether it is located on the main chain, and within a predetermined range of which arms; wherein the features of the nodes corresponding to the atoms of the ligand sample in the atom neighbor graph include at least one of the following items of the atom: atom type, atom position, which arm it belongs to, or which skeleton it belongs to; wherein the features of the edges of the atom neighbor graph include the distance between the two atoms connected by the edge and the type of the edge; wherein the type of the edge includes any one of the following items: an edge between target and ligand, an edge between target and target, an edge between ligand and target, or an edge between ligand and ligand.

[0118] Optionally, the characteristics of the nodes of the fully connected graph include at least one of the following items of the atom: atom type, atom position, whether it belongs to the skeleton, or whether it belongs to the arm, and the characteristics of the edges of the fully connected graph include the chemical bond type and an indication of whether the two atoms connected by the edge belong to the same decomposition prior distribution.

[0119] Exemplarily, the training unit 505 is also configured to determine whether the training process is completed based on a loss function, wherein the loss function includes a first loss function constructed based on the atomic positions in the ligand sample and the denoised atomic positions, a second loss function constructed based on the atomic types in the ligand sample and the denoised atomic types, and a third loss function constructed based on the chemical bonds in the ligand sample and the denoised chemical bonds.

[0120] Figure 5 The device 500 can be used to achieve the above combination Figures 2 to 4 For the sake of brevity, the process will not be described in detail here.

[0121] The division of modules or units in the embodiments of the present disclosure is schematic and is only a logical function division. There may be other division methods in actual implementation. In addition, each functional unit in the disclosed embodiments may be integrated into one unit, or may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0122] Figure 6 6 is a block diagram of an example device 600 that can be used to implement embodiments of the present disclosure. It should be understood that Figure 6 The device 600 shown is merely exemplary and should not be construed as limiting the functionality and scope of the implementations described herein. Figures 2 to 4 For example, the device 600 may be implemented as a classical computer and / or a quantum computer.

[0123] like Figure 6 As shown, device 600 is in the form of a general-purpose computing device. Components of computing device 600 may include, but are not limited to, one or more processors or processing units 610, memory 620, storage device 630, one or more communication units 640, one or more input devices 650, and one or more output devices 660. Processing unit 610 may be an actual or virtual processor and is capable of performing various processes according to a program stored in memory 620. In a multi-processor system, multiple processing units execute computer executable instructions in parallel to increase the parallel processing capabilities of computing device 600.

[0124] The computing device 600 typically includes a plurality of computer storage media. Such media may be any available media accessible to the computing device 600, including but not limited to volatile and non-volatile media, removable and non-removable media. The memory 620 may be a volatile memory (e.g., registers, caches, random access memory (RAM)), a non-volatile memory (e.g., read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory) or some combination thereof. The storage device 630 may be a removable or non-removable medium, and may include a machine-readable medium, such as a flash drive, a disk, or any other medium, which may be capable of being used to store information and / or data (e.g., training data for training) and may be accessed within the computing device 600.

[0125] The computing device 600 may further include additional removable / non-removable, volatile / non-volatile storage media. Figure 6 As shown in , a disk drive for reading or writing from a removable, non-volatile disk (e.g., a "floppy disk") and an optical drive for reading or writing from a removable, non-volatile optical disk may be provided. In these cases, each drive may be connected to the bus (not shown) by one or more data media interfaces. The memory 620 may include a computer program product 625 having one or more program modules that are configured to perform various methods or actions of various implementations of the present disclosure.

[0126] The communication unit 640 enables communication with other computing devices via a communication medium. Additionally, the functions of the components of the computing device 600 can be implemented in a single computing cluster or multiple computing machines that can communicate via a communication connection. Therefore, the computing device 600 can operate in a networked environment using a logical connection with one or more other servers, a network personal computer (PC), or another network node.

[0127] Input device 650 may be one or more input devices, such as a mouse, keyboard, tracking ball, etc. Output device 660 may be one or more output devices, such as a display, a speaker, a printer, etc. Computing device 600 may also communicate with one or more external devices (not shown) through communication unit 640 as needed, such as storage devices, display devices, etc., communicate with one or more devices that allow a user to interact with computing device 600, or communicate with any device that allows computing device 600 to communicate with one or more other computing devices (e.g., a network card, a modem, etc.). Such communication may be performed via an input / output (I / O) interface (not shown).

[0128] According to an exemplary implementation of the present disclosure, a computer-readable storage medium is provided, on which computer-executable instructions are stored, wherein the computer-executable instructions are executed by a processor to implement the method described above. According to an exemplary implementation of the present disclosure, a computer program product is also provided, which is tangibly stored on a non-transitory computer-readable medium and includes computer-executable instructions, and the computer-executable instructions are executed by a processor to implement the method described above. According to an exemplary implementation of the present disclosure, a computer program product is provided, on which a computer program is stored, and when the program is executed by a processor, the method described above is implemented.

[0129] Various aspects of the present disclosure are described herein with reference to the flowcharts and / or block diagrams of the methods, devices, equipment, and computer program products implemented according to the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer-readable program instructions.

[0130] These computer-readable program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processing unit of the computer or other programmable data processing device, a device that implements the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and / or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0131] Computer-readable program instructions can be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operational steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, so that the instructions executed on the computer, other programmable data processing apparatus, or other device implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0132] The flow chart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to multiple implementations of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and a part of a module, program segment or instruction includes one or more executable instructions for realizing the logical function of the specification. In some implementations as replacements, the function marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous square boxes can actually be executed substantially in parallel, and they can sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be realized by a special hardware-based system that performs the function or action of the specification, or can be realized by a combination of special hardware and computer instructions.

[0133] The above descriptions of various implementations of the present disclosure are exemplary, non-exhaustive, and not limited to the disclosed implementations. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described implementations. The selection of terms used herein is intended to best explain the principles of the implementations, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the various implementations disclosed herein.

Claims

1. A method for generating a ligand, comprising: Obtaining a trained ligand generation model, wherein the trained ligand generation model is generated by modeling atomic positions, atom types, and chemical bonds based on decomposition of the ligand; as well as By using the trained ligand generation model, a target ligand molecule corresponding to the target protein is obtained based on the target protein, The trained ligand generation model is generated by the following process: Determine a decomposed prior distribution based on target samples and corresponding ligand samples in the training data items; Constructing an atomic neighbor graph, wherein the atomic neighbor graph represents the relationship between multiple neighboring atoms; constructing a fully connected graph, wherein the fully connected graph represents the relationship between different atoms of the ligand sample; as well as The trained ligand generation model is generated through training based on the decomposed prior distribution, the atomic neighbor graph and the fully connected graph.

2. The method according to claim 1, wherein obtaining a target ligand molecule corresponding to the target protein based on the target protein by using the trained ligand generation model comprises: Based on the target protein, determining the decomposition prior distribution of the skeleton and the arm in the initial ligand molecule, wherein the decomposition prior distribution of the skeleton and the arm in the initial ligand molecule includes the mean matrix and the covariance matrix of multiple atomic clusters of the initial ligand molecule; Determining initial data based on the decomposed prior distribution of the backbone and the arms in the initial ligand molecule by sampling, wherein the initial data includes initial atomic positions, initial atomic types, and initial chemical bonds; Inputting the initial data into the trained ligand generation model to obtain output data, wherein the output data includes output atom positions, output atom types, and output chemical bonds; as well as The target ligand molecule is generated based on the output data.

3. The method of claim 2, wherein the initial ligand molecule comprises a known ligand molecule that binds to the target protein of interest, and wherein determining the decomposed prior distribution of the backbone and arms in the initial ligand molecule based on the target protein of interest comprises: Obtaining the known ligand molecule that binds to the target protein; Based on the binding conditions between the multiple atoms of the known ligand molecule and the binding pocket of the target protein, decomposing the multiple atoms into multiple atomic clusters, wherein one atomic cluster in the multiple atomic clusters is a skeleton, and the remaining atomic clusters in the multiple atomic clusters are arms; determining a mean matrix and a covariance matrix for each of the plurality of atomic clusters based on coordinates of atoms in each of the plurality of atomic clusters; as well as Based on the mean matrix and the covariance matrix of the plurality of atomic clusters, the decomposed prior distribution of the backbone and the arms in the initial ligand molecule is determined.

4. The method of claim 2, wherein the initial ligand molecule comprises a pseudo-ligand molecule that binds to the target protein of interest, and wherein determining a decomposed prior distribution of the backbone and arms in the initial ligand molecule based on the target protein of interest comprises: Determining the pseudo-ligand molecule that binds to the target protein by prediction; Based on the binding pocket of the target protein, determining multiple pseudo-ligand atoms in the pseudo-ligand molecule; Based on the distances between the multiple pseudo-ligand atoms, decomposing the multiple pseudo-ligand atoms into multiple atomic clusters, wherein one atomic cluster in the multiple atomic clusters is a skeleton, and the remaining atomic clusters in the multiple atomic clusters are arms; determining a mean matrix and a covariance matrix for each of the plurality of atomic clusters based on coordinates of atoms in each of the plurality of atomic clusters; as well as Based on the mean matrix and the covariance matrix of the plurality of atomic clusters, the decomposed prior distribution of the backbone and the arms in the initial ligand molecule is determined.

5. The method according to claim 2, wherein determining the initial data by sampling based on the decomposed prior distribution of the backbone and the arms in the initial ligand molecule comprises: Sampling the decomposed prior distribution of the backbone and the arms in the initial ligand molecule to determine the initial atomic positions; uniformly sampling a set of multiple atomic types to determine the initial atomic type; as well as A collection of multiple chemical bonds is uniformly sampled to determine the initial chemical bond.

6. The method of claim 2, wherein inputting the initial data into the trained ligand generation model to obtain output data comprises: The initial data is input into the trained ligand generation model, and the output data is obtained through a multi-step denoising process.

7. The method of claim 6, wherein the trained ligand generation model comprises a group equivariant graph neural network, and at each step of the multi-step denoising process: Inputting the atomic positions, atomic types and chemical bonds obtained in the previous step into the group equivariant graph neural network, and obtaining intermediate atomic positions, intermediate atomic types and intermediate chemical bonds by denoising; and By applying a position gradient guidance to the intermediate atomic position, the atomic position, atom type and chemical bond for the next step are obtained based on the intermediate atomic position, the intermediate atomic type and the intermediate chemical bond.

8. The method according to claim 7, wherein the position gradient guidance is used to impose position constraints based on the position of the intermediate atom so that the distance between the bonded backbone atoms and the arm atoms is between the minimum chemical bond length and the maximum chemical bond length, and does not collide with the atoms of the target protein.

9. The method according to claim 1, further comprising: Acquire a training data set including a plurality of training data items, each of the plurality of training data items including a target sample and a corresponding ligand sample; as well as The trained ligand generation model is generated by training based on the training data set.

10. The method according to claim 1, wherein the trained ligand generation model comprises a group equivariant graph neural network, and wherein generating the trained ligand generation model through training based on the decomposed prior distribution, the atomic neighbor graph, and the fully connected graph comprises: Based on the atomic positions, atomic types and chemical bonds in the ligand sample, a noisy atomic positions, noisy atomic types and noisy chemical bonds are obtained through a noisy process; as well as The atomic neighbor graph and the fully connected graph are input into the group equivariant graph neural network, and the noisy atomic positions, the noisy atomic types and the noisy chemical bonds are updated to obtain denoised atomic positions, denoised atomic types and denoised chemical bonds.

11. The method according to claim 10, wherein the group equivariant graph neural network comprises a plurality of modules stacked in sequence, each of the plurality of modules comprises a first neural network, a second neural network and a third neural network, and wherein the first neural network is used to update features corresponding to atomic types, the second neural network is used to update features corresponding to chemical bonds, and the third neural network is used to update features corresponding to atomic positions.

12. The method according to claim 1, wherein the characteristics of the nodes corresponding to the atoms of the target sample in the atom neighbor graph include at least one of the following items of the atom: atom type, atom position, amino acid type, whether it is located on the main chain, and within a predetermined range of which arms; wherein the characteristics of the nodes corresponding to the atoms of the ligand sample in the atom neighbor graph include at least one of the following items of the atom: atom type, atom position, which arm it belongs to, or which skeleton it belongs to; wherein the characteristics of the edges of the atom neighbor graph include the distance between the two atoms connected by the edge and the type of the edge; wherein the type of the edge includes any one of the following items: an edge between target and ligand, an edge between target and target, an edge between ligand and target, or an edge between ligand and ligand.

13. A method according to claim 1, wherein the characteristics of the nodes of the fully connected graph include at least one of the following items of the atom: atom type, atom position, whether it belongs to a skeleton, or whether it belongs to an arm, and the characteristics of the edges of the fully connected graph include a chemical bond type and an indication of whether the two atoms connected by the edge belong to the same decomposition prior distribution.

14. The method according to claim 10, further comprising: Whether the training process is completed is determined based on a loss function, wherein the loss function includes a first loss function constructed based on the atomic positions in the ligand sample and the denoised atomic positions, a second loss function constructed based on the atomic types in the ligand sample and the denoised atomic types, and a third loss function constructed based on the chemical bonds in the ligand sample and the denoised chemical bonds.

15. An electronic device, comprising: at least one processing unit; at least one memory coupled to the at least one processing unit and storing instructions for execution by the at least one processing unit, the instructions, when executed by the at least one processing unit, causing the electronic device to perform actions, the actions comprising: obtaining a trained ligand generation model, the trained ligand generation model being generated by modeling atomic positions, atom types, and chemical bonds based on decomposition of the ligand; and By using the trained ligand generation model, a target ligand molecule corresponding to the target protein is obtained based on the target protein, The trained ligand generation model is generated by the following process: Determine a decomposed prior distribution based on target samples and corresponding ligand samples in the training data items; Constructing an atomic neighbor graph, wherein the atomic neighbor graph represents the relationship between multiple neighboring atoms; constructing a fully connected graph, wherein the fully connected graph represents the relationship between different atoms of the ligand sample; and The trained ligand generation model is generated through training based on the decomposed prior distribution, the atomic neighbor graph and the fully connected graph.

16. A device for generating a ligand, comprising: a model acquisition unit configured to acquire a trained ligand generation model, wherein the trained ligand generation model is generated by modeling atomic positions, atomic types, and chemical bonds based on decomposition of the ligand; as well as A ligand generation unit is configured to obtain a target ligand molecule corresponding to the target protein based on the target protein by using the trained ligand generation model. The trained ligand generation model is generated by the following process: Determine a decomposed prior distribution based on target samples and corresponding ligand samples in the training data items; Constructing an atomic neighbor graph, wherein the atomic neighbor graph represents the relationship between multiple neighboring atoms; constructing a fully connected graph, wherein the fully connected graph represents the relationship between different atoms of the ligand sample; as well as The trained ligand generation model is generated through training based on the decomposed prior distribution, the atomic neighbor graph and the fully connected graph.

17. A computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method according to any one of claims 1 to 14.

Images