Molecular generation method and device

Abstract

The embodiments of this specification disclose a method and device for molecule generation. The method includes: obtaining protein monomer information and protein pocket information of a target protein monomer; inputting the protein monomer information and protein pocket information of the target protein monomer into a molecule generation model, and the molecule generation model outputs a candidate molecule; the molecule generation model includes a graph neural network, a first predictor, a second predictor, and a discriminator; the graph neural network is used to embed the interaction information between the protein pocket and the molecular fragment, the first predictor is used to predict the atom type, the second predictor is used to predict the molecular bond, and the discriminator is used to judge whether the molecule generation is completed. The final model outputs a candidate molecule that has good binding ability with the protein pocket of the protein monomer. The process is limited to the space composed of the entire protein pocket, and atoms are added one by one in the space to generate new compound molecules. The device is implemented based on the method.

Description

Technical Field

[0001] One or more embodiments of this specification relate to the technical field of computer-aided drug design, and in particular, to a molecule generation method and apparatus. Background Art

[0002] Computer-aided drug discovery is a prominent approach in modern preclinical drug discovery and is becoming a key driver in the pharmaceutical field. Advanced computer technologies, such as virtual screening, molecular modeling and simulation, machine learning, deep learning, and the application of bioinformatics methods, have provided new tools and methods for drug design, discovery, and optimization.

[0003] De novo drug design is one of the main research directions of computer-aided drug design, aiming to generate new compound molecules with the intended target. De novo drug design can be divided into ligand-based de novo drug design and receptor-based de novo drug design. Ligand-based de novo drug design relies on prior knowledge of a set of ligands with a given structure to generate new compounds. However, this type of method only relies on information about the compound itself to generate new compound molecules. Therefore, it is difficult to ensure that the generated compounds meet the expected properties, or even lack pharmaceutical activity, which makes ligand-based methods difficult to use directly in drug development. Receptor-based de novo drug design, also known as structure-based de novo drug design, mostly utilizes the sequence or structural information of the target protein, especially the active pocket information of the target protein, to generate new compound molecules with good binding ability to the target protein. Therefore, compared with ligand-based de novo drug design, the compound molecules obtained by this method are more likely to have better biological activity.

[0004] Currently, most methods for generating molecules based on protein pocket information primarily employ a strategy of implicitly encoding the protein pocket information and then decoding the encoded information to produce a series of compounds with the desired properties. Although this method of generating molecules based on pocket information has been widely studied, it also has some drawbacks:

[0005] 1. Most existing methods utilize relatively simple protein pocket information, generally considering only the sequence information of the protein pocket while ignoring its structural information. Although some studies have attempted to encode pocket information as a graph consisting of nodes and edges to better utilize the pocket's structural information during molecule generation, encoding pockets as a graph also results in the loss of the original pocket's structural information.

[0006] 2. In the process of generating molecules, most of the existing molecule generation methods based on protein pocket information do not take into account the interaction information between molecules and protein pockets. Summary of the Invention

[0007] One or more embodiments of this specification describe a method and apparatus for generating molecules, which fully utilize the structural information of protein pockets and the interaction information between the amino acid atoms in the protein pockets and the atoms of the generated molecular fragments to generate molecules that have good binding ability to the protein pockets. The molecules can even be used as candidate compounds for disease treatment.

[0008] In a first aspect, the embodiments of this specification provide a method for generating a molecule, comprising:

[0009] Obtaining protein monomer information and protein pocket information of the target protein monomer; wherein the protein monomer information includes the protein monomer sequence and the protein monomer atom set, wherein the protein monomer atom set is the set of atoms constituting the protein monomer and their spatial positions; the protein pocket information includes the protein pocket center position and protein pocket size;

[0010] The protein monomer information and protein pocket information of the target protein monomer are input into the molecular generation model, and the molecular generation model outputs a candidate molecule; the molecular generation model includes a graph neural network, a first predictor, a second predictor, and a discriminator; the specific process is:

[0011] The graph neural network is used to embed information about interactions between protein pockets and molecular fragments;

[0012] The first predictor predicts atoms at a certain spatial position based on the interaction information between protein pockets and molecular fragments embedded in the graph neural network;

[0013] The second predictor predicts the bond connecting the atom and the previously predicted molecular fragment based on the previously predicted molecular fragment information and the atom at a certain spatial position predicted by the first predictor;

[0014] The discriminator is used to determine that when molecule generation is not completed, it returns the predicted atoms at a certain spatial position and the bonds connecting them to the molecular fragments to the graph neural network. The graph neural network embeds the interaction information between the new protein pocket and the molecular fragment, and then uses the first predictor and the second predictor to predict the atoms at a new spatial position and the bonds connecting them to the molecular fragments. This process repeats until the molecule generation is completed, and the final molecule is output as a candidate molecule.

[0015] Among them, during the first prediction, the protein pocket and molecular fragment interaction information embedded by the graph neural network based on the protein monomer information and protein pocket information of the target protein monomer does not contain molecular fragment information; the first predictor predicts the first atom in a certain spatial position that constitutes the molecular fragment based on the protein pocket and molecular fragment interaction information first embedded by the graph neural network, and returns it to the graph neural network.

[0016] In some embodiments, obtaining protein monomer information and protein pocket information of a target protein monomer includes:

[0017] Obtaining a target protein monomer, wherein the target protein monomer includes protein monomer information;

[0018] Determining a protein pocket based on the protein monomer information;

[0019] Based on the protein pocket, the center position and size of the protein pocket are determined.

[0020] In some embodiments, obtaining a target protein monomer comprises:

[0021] Obtaining protein and ligand complexes;

[0022] The protein and its ligand complex are separated to obtain target protein monomers.

[0023] In some embodiments, when the first predictor predicts and outputs an atom at a certain spatial position with the highest probability each time, the molecule generation model completes a cycle and outputs a candidate molecule; when the first predictor predicts and outputs the N atoms at a certain spatial position with the top N probabilities each time, the molecule generation model completes a cycle and outputs N M candidate molecules, and M is the number of cycles greater than 0.

[0024] In some embodiments, the training process of the graph neural network, the first predictor, and the second predictor in the molecular generative model is as follows:

[0025] Obtaining a training set, the training set comprising a plurality of training subsets of protein and ligand complexes, each training subset comprising protein monomer sample information, protein pocket sample information, and ligand sample information; the ligand sample information is divided into masked molecular fragment sample information, atomic sample information of masked atoms, and bond sample information of masked bonds after atom masking processing; the atomic sample information comprises atoms and their position information;

[0026] The protein monomer sample information, protein pocket sample information, atom sample information, and bond sample information of the training subset are input into the graph neural network for training, and the predicted information of the interaction between protein pockets and molecular fragments is output;

[0027] The predicted information of the interaction between protein pockets and molecular fragments is input into the first predictor, which outputs the predicted atoms at a certain spatial position;

[0028] The molecular fragment sample information and the predicted atom at a certain spatial position are input into the second predictor, which outputs the predicted bond;

[0029] The first loss function is used to judge the convergence of predicted atoms and sample atoms; the second loss function is used to judge the convergence of predicted keys and sample keys; when both the first loss function and the second loss function meet the convergence conditions, the training of the graph neural network, the first predictor, and the second predictor is completed, otherwise, the training continues until the convergence conditions are met.

[0030] In some embodiments, after the prediction is completed, the method further includes:

[0031] The performance of the predicted molecules is evaluated and the molecules are output together with the evaluation results.

[0032] In some embodiments, the performance evaluation includes: drug-likeness evaluation, synthetic accessibility evaluation, and molecule-protein pocket docking score evaluation.

[0033] In a second aspect, the embodiments of this specification provide a molecule generation device, comprising:

[0034] An acquisition module is used to obtain protein monomer information and protein pocket information of a target protein monomer; wherein the protein monomer information includes a protein monomer sequence and a protein monomer atom set, wherein the protein monomer atom set is a set of atoms constituting the protein monomer and their spatial positions; and the protein pocket information includes a protein pocket center position and a protein pocket size;

[0035] The molecule generation module is used to input the protein monomer information and protein pocket information of the target protein monomer into the molecule generation model, and the molecule generation model outputs the candidate molecule; the molecule generation model includes a graph neural network, a first predictor, a second predictor, and a discriminator; the specific process is as follows:

[0036] The graph neural network is used to embed information about interactions between protein pockets and molecular fragments;

[0037] The first predictor predicts atoms at a certain spatial position based on the interaction information between protein pockets and molecular fragments embedded in the graph neural network;

[0038] The second predictor predicts the bond connecting the atom and the previously predicted molecular fragment based on the previously predicted molecular fragment information and the atom at a certain spatial position predicted by the first predictor;

[0039] The discriminator is used to determine that when the molecule generation is not completed, the atoms at a certain spatial position predicted this time and the bonds connecting them to the molecular fragments are returned to the graph neural network. The graph neural network embeds the interaction information between the new protein pocket and the molecular fragment, and then uses the first predictor and the second predictor to predict the atoms at a new spatial position and the bonds connecting them to the molecular fragments predicted previously. This process is repeated until the prediction is completed, and the final molecule is output as a candidate molecule.

[0040] Among them, during the first prediction, the protein pocket structure and molecular fragment interaction information embedded by the graph neural network based on the protein monomer information and protein pocket information of the target protein monomer does not contain molecular fragment information. The first predictor predicts the first atom in a certain spatial position that constitutes the molecular fragment based on the protein pocket and molecular fragment interaction information first embedded by the graph neural network, and returns it to the graph neural network.

[0041] In some embodiments, the acquisition module includes:

[0042] an acquisition unit, configured to acquire a target protein monomer, wherein the target protein monomer includes protein monomer information;

[0043] The extraction unit is used to determine the protein pocket based on the protein monomer information; and based on the protein pocket, determine the center position and size of the protein pocket.

[0044] In some embodiments, the device further comprises an evaluation module for performing a performance evaluation on the predicted molecules after the prediction is completed to obtain an evaluation result of the molecules.

[0045] The beneficial effects of the technical solutions provided by some embodiments of this specification include at least:

[0046] In one or more embodiments of this specification, the method of the embodiment of this specification makes full use of the structural information of the protein pocket, by clarifying the spatial coordinates of all atoms contained in the protein pocket, with the space composed of the entire pocket as the limit, adding atoms one by one in the space to generate new compound molecules. In the process of generating molecules, the structural information of the protein pocket and the interaction information between the protein pocket and the atoms of the generated molecular fragments can be fully utilized to generate new compound molecules with high binding ability to the protein pocket. Further, the generated compound molecules are likely to become candidate compounds for the treatment of diseases. The method of the embodiment of this specification improves the efficiency of the drug development process, reduces the cost of the drug development process, and also provides more reference examples for pharmaceutical experts. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of this specification, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0048] Figure 1 A flow chart of a method for generating molecules provided in an embodiment of this specification;

[0049] Figure 2A structural block diagram of a molecule generation device provided in an embodiment of this specification;

[0050] Figure 3a Schematic diagram of the molecular structure of the original ligand in the example of molecule generation based on the L1 / S3 pocket of the p53 protein;

[0051] Figure 3b Schematic diagrams of the molecular structures of multiple candidate molecules in the example of molecular generation based on the L1 / S3 pocket of the p53 protein;

[0052] Figure 4a Generate drug-likeness evaluation charts for multiple candidate molecules based on the L1 / S3 pocket of p53 protein;

[0053] Figure 4b Synthetic accessibility evaluation chart for multiple candidate molecules in the example of molecular generation based on the L1 / S3 pocket of the p53 protein

[0054] Figure 4c A chart showing the evaluation scores of the molecule-protein pocket docking for multiple candidate molecules in the example of molecule generation based on the L1 / S3 pocket of the p53 protein. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of this specification will be described clearly and completely below in conjunction with the drawings in the embodiments of this specification.

[0056] Throughout this specification, the claims, and the accompanying drawings, the terms "first," "second," "third," and the like are used to distinguish between different items, not to describe a particular order. Furthermore, the terms "including," "having," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements but may optionally include steps or elements not listed, or may include other steps or elements inherent to the process, method, product, or apparatus.

[0057] like Figure 1 The method of the embodiment of this specification provides a method for generating a molecule, comprising:

[0058] Step 102: Obtain protein monomer information and protein pocket information of the target protein monomer; wherein the protein monomer information includes the protein monomer sequence and the protein monomer atom set, wherein the protein monomer atom set is the set of atoms constituting the protein monomer and their spatial positions; and the protein pocket information includes the protein pocket center position and the protein pocket size;

[0059] Step 104: Input the protein monomer information and protein pocket information of the target protein monomer into a molecular generation model, which outputs candidate molecules. The molecular generation model includes a graph neural network, a first predictor, a second predictor, and a discriminator. The specific process is as follows:

[0060] The graph neural network is used to embed information about interactions between protein pockets and molecular fragments;

[0061] The first predictor predicts atoms at a certain spatial position based on the interaction information between protein pockets and molecular fragments embedded in the graph neural network;

[0062] The second predictor predicts the bond connecting the atom and the previously predicted molecular fragment based on the previously predicted molecular fragment information and the atom at a certain spatial position predicted by the first predictor;

[0063] The discriminator is used to determine that when molecule generation is not completed, it returns the predicted atoms at a certain spatial position and the bonds connecting them to the molecular fragments to the graph neural network. The graph neural network embeds the interaction information between the new protein pocket and the molecular fragment, and then uses the first predictor and the second predictor to predict the atoms at a new spatial position and the bonds connecting them to the molecular fragments. This process repeats until the molecule generation is completed, and the final molecule is output as a candidate molecule.

[0064] Among them, during the first prediction, the protein pocket and molecular fragment interaction information embedded by the graph neural network based on the protein monomer information and protein pocket information of the target protein monomer does not contain molecular fragment information; the first predictor predicts the first atom in a certain spatial position that constitutes the molecular fragment based on the protein pocket and molecular fragment interaction information first embedded by the graph neural network, and returns it to the graph neural network.

[0065] The method of the embodiment of this specification is performed by a computer device, using a molecular generation model, based on protein monomer information and protein pocket information as input, and can output a candidate molecule. The whole process uses the protein pocket as a constraint condition and generates candidate molecules by adding atoms one by one in the pocket. This process reduces the complexity of the molecular search and can obtain candidate molecules more quickly using the molecular generation model. In addition, this process also makes full use of the interaction information between the protein pocket and the molecular fragment, so that the generated molecules can better adapt to the shape of the protein pocket and can be better combined with the protein pocket. Further, after the performance evaluation of the generated molecules, they can also be used as candidate compounds for disease treatment.

[0066] The computer device may be a mobile terminal device, a desktop computer, a notebook, etc.

[0067] Next, each step will be described in detail.

[0068] In step 102, in one example, the protein monomer information and protein pocket information of the target protein monomer can be obtained by a user inputting the protein monomer information and protein pocket information in an interface.

[0069] In another example, the protein monomer information and protein pocket information of the target protein monomer can be obtained at the system level.

[0070] For example, protein monomer information and protein pocket information of a target protein monomer are obtained from a storage list of multiple protein monomers.

[0071] For another example, the obtained protein monomer is analyzed and processed to obtain protein monomer information and protein pocket information. Specifically, the protein monomer information and protein pocket information of the target protein monomer are obtained, including:

[0072] Step a1, obtaining a target protein monomer, wherein the target protein monomer includes protein monomer information;

[0073] Step a2, determining the protein pocket based on the protein monomer information;

[0074] Step a3: Based on the protein pocket, determine the center position and size of the protein pocket.

[0075] In step a1, the target protein monomer can be a protein monomer selected from a plurality of protein monomer storage lists, or can be obtained by separating the obtained protein and its ligand complex. The protein and its ligand complex can be obtained from a public database or experimental data, and is generally saved as a PDB file. Molecular dynamics simulation is performed on the protein and its ligand complex, and then the protein and its ligand complex is separated and the separated protein monomer is saved. The separated protein monomer also needs to be hydrogenated to complete the protein sequence, and water molecules need to be removed.

[0076] In step a2, after the protein pocket is determined, it is necessary to make the protein pocket contain the main interacting residues with the known ligand as much as possible. For example, the protein pocket is determined by limiting the distance between the residues that interact with the ligand and the center of the pocket to be less than the pocket space. In addition, the residues can also be determined manually, such as the residues that form hydrogen bonds with the ligand reported in the literature, or the main amino acid residues that the user believes are related to the protein function. The manually determined residues are pre-stored before the protein pocket is determined. When the protein pocket is determined, the relevant residues can be identified based on the pre-stored information.

[0077] In step a3, the protein pocket is determined as a set of amino acid atoms contained in the pocket and their positions There are forms. Indicates the information of the i-th non-hydrogen atom in the protein pocket (including element type, amino acid, and whether it is the main chain). represents the three-dimensional coordinates of the i-th non-hydrogen atom in the protein pocket, and N represents the number of non-hydrogen atoms contained in the protein pocket.

[0078] Obtain the atomic coordinates of the protein pocket circumference, for example, by using AutoDock Tools software to calculate the coordinates of the center of the protein pocket, as well as the size of the protein pocket. The size of the protein pocket refers to the side length of a cube box that contains the pocket.

[0079] In step 104, the molecular generation model includes a graph neural network, a first predictor, a second predictor, and a discriminator.

[0080] The graph neural network can be a rotation and translation invariant graph neural network, which is used to embed the interaction information between protein pockets and molecular fragments.

[0081] During the initial prediction, the graph neural network embeds information about protein pockets and molecular fragment interactions that lack molecular fragment information based on the input protein monomer and protein pocket information of the target protein. Since no atoms constituting the molecule have yet been predicted, the molecular fragment information is empty.

[0082] The protein pocket and molecular fragment interaction information that does not have molecular fragment information is input into the first predictor, and the first predictor outputs the predicted atom at a certain spatial position. The atom predicted this time is the first atom that makes up the molecule, which can be understood as the molecular fragment predicted for the first time. The first predictor can be a multi-layer perceptron, which is used to predict the probability of which type of atom appears at a certain coordinate point in the protein cavity. In one example, the atom with the highest probability of occurrence can be determined as the predicted atom. For example, (carbon C: probability P1) appears under (x1, y1, z1). In another example, the top N atoms with the highest probability of occurrence can be determined as the predicted atoms. For example, (carbon C: probability P1; oxygen O: probability P2; nitrogen N: probability P3) appear under (x1, y1, z1).

[0083] Since only the first atom constituting the molecule is predicted during the first prediction, the second predictor cannot predict a bond based on one atom.

[0084] The discriminator is used to determine whether molecule generation has ended. If the discriminator outputs a termination signal, the molecule generation is confirmed to be complete. Otherwise, the molecule generation is not complete and the next prediction needs to be entered. During the first prediction, the discriminator does not output a termination signal and the next prediction needs to be entered.

[0085] The atoms output by the first predictor are fed back into the graph neural network, which embeds information about the interactions between the new protein pocket and the molecular fragment. This is when the second prediction process begins.

[0086] The protein pocket and the molecular fragment interaction information is input into the first predictor, which outputs the predicted atom at a certain spatial position. In this case, the first predictor performs predictions based on the spatial range of the protein pocket and the molecular fragment interaction information.

[0087] The first predictor inputs the predicted atoms into a second predictor, which then predicts the bonds connecting the atoms to the previously predicted molecular fragments. The bonds predicted for the second time are the molecular bonds between the atoms predicted for the second time and the atoms predicted for the first time. The second predictor can be a feedforward neural network that predicts the molecular bonds between the newly added atoms and the existing molecular fragments. When the first predictor inputs N atoms into the second predictor, the second predictor outputs N bonds corresponding to the atoms.

[0088] When the discriminator determines that the molecule generation is not completed, it continues to make predictions.

[0089] The atoms output by the first predictor and the bonds output by the second predictor are fed back into the graph neural network, which embeds information about the interactions between the new protein pocket and the molecular fragment. This is when the third prediction process begins.

[0090] This process repeats until the molecule generation is completed.

[0091] The molecular generation model stores the information of each prediction and, after the molecular generation is completed, generates molecules based on the atoms and bonds predicted multiple times. The generated molecules are finally determined as candidate molecules.

[0092] The molecular fragment generated during the molecule generation process is represented as a collection of atoms, bonds, and atomic coordinates contained in the fragment. in Represents the information of the i-th atom in the molecular fragment (atoms include element type, valence, hydrogen bond donor and hydrogen bond acceptor), Indicates the bond information formed between the i-th atom and other atoms, represents the three-dimensional coordinates of the i-th atom in the molecular fragment, and N represents the number of atoms contained in the molecular fragment.

[0093] At the end of the prediction, when the first predictor predicts and outputs the atom at a certain spatial position with the highest probability each time, the molecule generation model completes a cycle and outputs a candidate molecule; when the first predictor predicts and outputs the N atoms at a certain spatial position with the top N probabilities each time, the molecule generation model completes a cycle and outputs N M candidate molecules, and M is the number of cycles greater than 0.

[0094] If the discriminator doesn't output a termination signal during the above process, it could potentially fall into an infinite loop. In this case, the discriminator must also determine the number of atoms generated. If the number of atoms generated falls below a threshold, molecule generation continues; otherwise, generation ceases. This atomic number determination prevents the molecule generation process from falling into an infinite loop and also prevents the generation of very large molecules containing a large number of atoms.

[0095] The number threshold is determined based on the usual number of molecules, for example, 50. When larger molecules need to be generated, the number threshold can be adjusted as needed. The method of the embodiment of this specification, before step 102, further includes: training a molecule generation model.

[0096] The training process of the graph neural network, the first predictor, and the second predictor in the molecular generative model is as follows:

[0097] Step 201: Acquire a training set, wherein the training set comprises a plurality of training subsets of protein and ligand complexes, each training subset comprising protein monomer sample information, protein pocket sample information, and ligand sample information; the ligand sample information is divided into masked molecular fragment sample information, masked atom atomic sample information, and masked bond bond sample information after atom masking processing; the atomic sample information comprises atom and position information;

[0098] Step 203: Input the protein monomer sample information, protein pocket sample information, atom sample information, and bond sample information of the training subset into a graph neural network for training, and output prediction information on the interaction between protein pockets and molecular fragments;

[0099] The predicted information of the interaction between protein pockets and molecular fragments is input into the first predictor, which outputs the predicted atoms at a certain spatial position;

[0100] The molecular fragment sample information and the predicted atom at a certain spatial position are input into the second predictor, which outputs the predicted bond;

[0101] The first loss function is used to judge the convergence of the predicted atoms and the sample atoms; the second loss function is used to judge the convergence of the predicted keys and the sample keys; when the first loss function and the second loss function both meet the convergence conditions, the training of the graph neural network, the first predictor, and the second predictor is completed; otherwise, step 203 is repeated to continue training until the convergence conditions are met.

[0102] In step 201, the training set is obtained from a dataset. The dataset is a dataset of a protein and its ligand complex, which includes a protein monomer pdb file, a ligand sdf file, and the RMSD of the docking pose of the ligand and the protein. Using the RMSD of the docking pose of the ligand and the protein, the RMSD of the docking pose of the ligand and the protein is screened and retained. Protein pocket and corresponding ligand files.

[0103] Each protein monomer pdb file stores protein monomer sample information. Protein pockets are determined for the protein monomer sample information, and the center position and size of the protein pocket are calculated, i.e., the protein pocket sample information.

[0104] Each ligand sdf file stores the molecular information of the ligand. Partial masking is performed on the molecule. For example, a single atom in the molecule is masked, and the masked molecule becomes a molecular fragment. The masked atom and the bonds connecting the fragment are also masked. The masked atoms and bonds are then the atoms and bonds to be predicted, while the unmasked molecule is the molecule to be predicted.

[0105] After masking, the randomly masked atoms constitute atomic sample information, the bonds corresponding to the masked atoms constitute bond sample information, and the masked molecules constitute molecular fragment sample information. Based on the masked molecular fragment sample information, atomic sample information for the masked atoms, and bond sample information for the masked bonds, the information of the molecules or molecular fragments before masking can be determined.

[0106] Afterwards, the graph neural network, the first predictor, and the second predictor are jointly trained using multiple training subsets.

[0107] The first and second loss functions can be distance loss functions, which are used to calculate the degree of approximation between the predicted value and the true value. When the predicted value is close to the true value, the convergence condition is reached and the training is completed. During the training process, the parameters of the graph neural network, the first predictor, and the second predictor are adjusted to obtain better prediction results.

[0108] The discriminator is trained separately, and a binary classifier is trained to determine whether the generation is completed.

[0109] The method of the embodiment of this specification, after the prediction is completed, further includes:

[0110] The performance of the predicted molecules is evaluated and the molecules are output together with the evaluation results.

[0111] This performance evaluation includes assessments of the docking score between the molecule and the protein pocket, synthetic accessibility, and other aspects. For example, if the generated molecule is intended for pharmaceutical use, the drug-likeness of the molecule can be assessed.

[0112] The drug-likeness and synthetic accessibility of the molecule can be evaluated using existing Python package code packages. The docking score evaluation of the molecule to the target pocket can be achieved using the open source software VINA.

[0113] Furthermore, the method of the embodiment of this specification also includes: outputting the optimal candidate molecule based on the evaluation results and the screening conditions.

[0114] The screening criteria are set based on common experience and design needs. For example, in the field of molecular generation, it is common practice to determine if a molecule with a QED score greater than 0.5 is more likely to be drug-like, while a molecule with a SA score less than 5 is less complex to synthesize. If the user desires a more drug-like molecule or one with greater synthetic accessibility, the corresponding screening values can be adjusted appropriately.

[0115] Next, we will explain the molecule generation project based on the L1 / S3 pocket of the p53 protein as an example.

[0116] The pdb file for the p53 protein R175H mutant (PDB ID: 2xwr) was obtained from the PDB database. Molecular docking simulations of the p53 protein L1 / S3 pocket and the UCI-L0023 molecule were then performed on a computer using AutoDock Vina software. Molecular dynamics simulations of the docking results of the p53 protein L1 / S3 pocket and the UCI-L0023 molecule were then performed using Gromacs software to obtain a more reasonable binding mode of the p53 protein L1 / S3 pocket and the UCI-L0023 molecule. PyMol software was then used to preprocess the p53 protein L1 / S3 pocket and the UCI-L0023 molecule complex: obtain the protein-ligand complex; separate the protein-ligand complex; complete the protein monomer sequence; and remove water molecules.

[0117] Afterwards, the structure-preprocessed p53 protein monomer was processed using AutoDockTools software to obtain the pocket center coordinates of the p53 protein L1 / S3 pocket and the size of the pocket space.

[0118] The pre-processed p53 protein monomer and the center coordinates of the L1 / S3 pocket, as well as the size of the pocket space, are fed into the molecule generation model. After multiple predictions, the generated molecules are output. In this example, a total of 1000 molecules are generated. Figure 3a Schematic diagram of the structure of the UCI-L0023 molecule; Figure 3b Schematic diagram of the structures of 6 molecules among the 1000 predicted molecules.

[0119] Afterwards, all generated molecules were evaluated for drug-likeness, synthetic accessibility, and molecule-protein pocket docking scores. Figure 4a 、 Figure 4b 、 Figure 4c These are the drug-likeness (QED) evaluation chart, synthetic accessibility (SA) evaluation chart, and molecule-protein pocket docking score evaluation chart of multiple candidate molecules in the molecule generation example based on the L1 / S3 pocket of the p53 protein. Figure 4a The QED score ranges from 0 to 1, and the higher the QED score, the better the drug properties of the molecule. Figure 4b The SA score ranges from 0 to 10, and the lower the SA score, the easier the molecule is to synthesize. Figure 4c The Vina docking score distribution of the generated molecules with the L1 / S3 pocket of the p53 protein. The more negative the dock score, the better the binding between the ligand and the receptor. The dotted line in the figure is the docking score of the target ligand UCI-L0023 with the L1 / S3 pocket of the p53 protein.

[0120] in accordance with Figure 4a-4c As shown, the QED scores of the generated molecules are mostly higher than 0.5, which indicates that most of the generated molecules exhibit good drug-like properties; the SA scores of the generated molecules are mostly lower than 4, which indicates that the synthesis complexity of most of the generated molecules is low; the dock scores of the generated molecules are mostly lower than the dock scores of UCI-L0023 and the L1 / S3 pocket of p53 protein, which indicates that the binding ability of the generated molecules to the L1 / S3 pocket of p53 protein is better than that of the target ligand.

[0121] contrast Figure 3a and Figure 3b , the bold part in the figure is the structure that both the generated molecule and the target molecule have. The embodiments of this specification can effectively learn the intrinsic relationship between the protein pocket structure and the molecular structure. The method of the embodiments of this specification utilizes the structural information of the protein pocket, takes the cavity of the protein pocket as a restriction, and generates a series of drug-like molecules in the protein cavity by adding atoms one by one. These molecules generally have good binding ability with the target protein pocket and have the potential to become candidate drugs. This improves the efficiency of the drug development process, reduces the cost of the drug development process, and also provides more reference examples for pharmaceutical experts.

[0122] like Figure 2 The embodiment of this specification also provides a molecule generation device, which may include at least: an acquisition module 1001 and a molecule generation module 1002, wherein:

[0123] The acquisition module 1001 is used to obtain protein monomer information and protein pocket information of the target protein monomer;

[0124] The protein monomer information includes the protein monomer sequence and the protein monomer atom set, wherein the protein monomer atom set is the set of atoms constituting the protein monomer and their spatial positions; the protein pocket information includes the protein pocket center position and protein pocket size;

[0125] The molecule generation module 1002 is used to input the protein monomer information and protein pocket information of the target protein monomer into the molecule generation model, and the molecule generation model outputs the candidate molecule; the molecule generation model includes a graph neural network, a first predictor, a second predictor, and a discriminator; the specific process is:

[0126] The graph neural network is used to embed information about interactions between protein pockets and molecular fragments;

[0127] The first predictor predicts atoms at a certain spatial position based on the interaction information between protein pockets and molecular fragments embedded in the graph neural network;

[0128] The second predictor predicts the bond connecting the atom and the previously predicted molecular fragment based on the previously predicted molecular fragment information and the atom at a certain spatial position predicted by the first predictor;

[0129] The discriminator is used to determine that when the molecule generation is not completed, the atoms at a certain spatial position predicted this time and the bonds connecting them to the molecular fragments are returned to the graph neural network. The graph neural network embeds the interaction information between the new protein pocket and the molecular fragment, and then uses the first predictor and the second predictor to predict the atoms at a new spatial position and the bonds connecting them to the molecular fragments predicted previously. This process is repeated until the prediction is completed, and the final molecule is output as a candidate molecule.

[0130] Among them, during the first prediction, the protein pocket structure and molecular fragment interaction information embedded by the graph neural network based on the protein monomer information and protein pocket information of the target protein monomer does not contain molecular fragment information. The first predictor predicts the first atom in a certain spatial position that constitutes the molecular fragment based on the protein pocket and molecular fragment interaction information first embedded by the graph neural network, and returns it to the graph neural network.

[0131] The device of the embodiment of this specification is applied to a computer device, and a molecular generation model is used to output candidate molecules based on protein monomer information and protein pocket information as input. The entire process uses the protein pocket as a constraint and generates candidate molecules by adding atoms one by one in the pocket. This process reduces the complexity of the molecular search and can obtain candidate molecules more quickly using the molecular generation model. In addition, this process also makes full use of the interaction information between the protein pocket and the molecular fragment, so that the generated molecules can better adapt to the shape of the protein pocket and can be better combined with the protein pocket. Further, after the performance evaluation of the generated molecules, they can also be used as candidate compounds for disease treatment.

[0132] Next, each module will be described in detail.

[0133] In one example, the acquisition module includes an acquisition unit for acquiring protein monomer information and protein pocket information of the target protein monomer.

[0134] The above information can be obtained by the user inputting protein monomer information and protein pocket information in the interface.

[0135] Alternatively, the above information can be obtained at the system level to obtain protein monomer information and protein pocket information. For example, the protein monomer information and protein pocket information of the target protein monomer can be obtained from a storage list of multiple protein monomers. For another example, the obtained protein monomer is analyzed and processed to obtain the protein monomer information and protein pocket information. Specifically, the acquisition module includes:

[0136] an acquisition unit, configured to acquire a target protein monomer, wherein the target protein monomer includes protein monomer information;

[0137] The target protein monomer can be a protein monomer selected from a plurality of protein monomer storage lists, or can be obtained by separating an obtained protein and its ligand complex. The protein and its ligand complex can be obtained from a public database or experimental data, and is generally saved as a PDB file. Molecular dynamics simulation is performed on the protein and its ligand complex, and then the protein and its ligand complex is separated and the separated protein monomer is saved. The separated protein monomer also needs to be hydrogenated to complete the protein sequence, and also needs to be processed such as removing water molecules.

[0138] The extraction unit is used to determine the protein pocket based on the protein monomer information; and based on the protein pocket, determine the center position and size of the protein pocket.

[0139] Once a protein pocket is identified, it is necessary to ensure that the pocket contains as many residues as possible that are primarily interacting with the known ligand. For example, this can be determined by limiting the distance between the ligand-interacting residues and the pocket center to less than the pocket volume. Alternatively, residues can be manually identified, such as those reported in the literature to form hydrogen bonds with the ligand, or residues identified by the user as being primarily relevant to protein function. These manually identified residues are pre-stored before protein pocket identification. Once the protein pocket is identified, relevant residues can be identified based on this pre-stored information.

[0140] Obtain the atomic coordinates of the protein pocket circumference, calculate the coordinates of the protein pocket center, and the size of the protein pocket.

[0141] The apparatus of the embodiment of this specification further includes a training module for training the molecular generation model.

[0142] The training module includes:

[0143] The first training unit is used to train the graph neural network, the first predictor, and the second predictor. The training process is as follows:

[0144] Obtaining a training set, the training set comprising a plurality of training subsets of protein and ligand complexes, each training subset comprising protein monomer sample information, protein pocket sample information, and ligand sample information; the ligand sample information is divided into masked molecular fragment sample information, atomic sample information of masked atoms, and bond sample information of masked bonds after atom masking processing; the atomic sample information comprises atoms and their position information;

[0145] The protein monomer sample information, protein pocket sample information, atom sample information, and bond sample information of the training subset are input into the graph neural network for training, and the predicted information of the interaction between protein pockets and molecular fragments is output;

[0146] The predicted information of the interaction between protein pockets and molecular fragments is input into the first predictor, which outputs the predicted atoms at a certain spatial position;

[0147] The molecular fragment sample information and the predicted atom at a certain spatial position are input into the second predictor, which outputs the predicted bond;

[0148] The first loss function is used to judge the convergence of predicted atoms and sample atoms; the second loss function is used to judge the convergence of predicted keys and sample keys; when both the first loss function and the second loss function meet the convergence conditions, the training of the graph neural network, the first predictor, and the second predictor is completed, otherwise, the training continues until the convergence conditions are met.

[0149] The second training unit is used to train a discriminator. The discriminator is trained separately and determines whether the generation is completed by training a binary classifier.

[0150] The device of the embodiment of this specification further includes an evaluation module, which is used to perform performance evaluation on the predicted molecules after the prediction is completed to obtain evaluation results of the molecules.

[0151] This performance evaluation includes assessments of the docking score between the molecule and the protein pocket, synthetic accessibility, and other aspects. For example, if the generated molecule is intended for pharmaceutical use, the drug-likeness of the molecule can be assessed.

[0152] The evaluation results are output together with the generated molecules. Users can filter and obtain the desired molecules as needed.

[0153] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of this specification is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted via the computer-readable storage medium. The computer instructions can be transmitted from a website, computer, server or data center to another website, computer, server or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrations. The available medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a digital versatile disc (DVD)), or a semiconductor medium (eg, a solid state disk (SSD)).

[0154] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When executed, the program can include the processes of the above-described method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks. The technical features of this embodiment and the implementation scheme can be combined in any manner unless they conflict.

[0155] The embodiments described above are merely preferred embodiments of this specification and are not intended to limit the scope of this specification. Without departing from the design spirit of this specification, various modifications and improvements made to the technical solutions of this specification by ordinary technicians in this field should fall within the scope of protection determined by the claims of this specification.

Claims

1. Molecular generation methods, including: Obtaining protein monomer information and protein pocket information of the target protein monomer; wherein the protein monomer information includes the protein monomer sequence and the protein monomer atom set, wherein the protein monomer atom set is the set of atoms constituting the protein monomer and their spatial positions; the protein pocket information includes the protein pocket center position and protein pocket size; The protein monomer information and protein pocket information of the target protein monomer are input into the molecular generation model, and the molecular generation model outputs a candidate molecule; the molecular generation model includes a graph neural network, a first predictor, a second predictor, and a discriminator; the specific process is: The graph neural network is used to embed information about interactions between protein pockets and molecular fragments; The first predictor predicts atoms at a certain spatial position based on the interaction information between protein pockets and molecular fragments embedded in the graph neural network; The second predictor predicts the bond connecting the atom and the previously predicted molecular fragment based on the previously predicted molecular fragment information and the atom at a certain spatial position predicted by the first predictor; The discriminator is used to determine that when molecule generation is not completed, it returns the predicted atoms at a certain spatial position and the bonds connecting them to the molecular fragments to the graph neural network. The graph neural network embeds the interaction information between the new protein pocket and the molecular fragment, and then uses the first predictor and the second predictor to predict the atoms at a new spatial position and the bonds connecting them to the molecular fragments. This process repeats until the molecule generation is completed, and the final molecule is output as a candidate molecule. Among them, during the first prediction, the protein pocket and molecular fragment interaction information embedded by the graph neural network based on the protein monomer information and protein pocket information of the target protein monomer does not contain molecular fragment information; the first predictor predicts the first atom in a certain spatial position that constitutes the molecular fragment based on the protein pocket and molecular fragment interaction information first embedded by the graph neural network, and returns it to the graph neural network.

2. The method according to claim 1, wherein obtaining protein monomer information and protein pocket information of the target protein monomer comprises: Obtaining a target protein monomer, wherein the target protein monomer includes protein monomer information; Determining a protein pocket based on the protein monomer information; Based on the protein pocket, the center position and size of the protein pocket are determined.

3. The method according to claim 2, wherein obtaining the target protein monomer comprises: Obtaining protein and ligand complexes; The protein and its ligand complex are separated to obtain target protein monomers.

4. The method according to claim 1, when the first predictor predicts and outputs an atom at a certain spatial position with the highest probability each time, the molecule generation model completes a cycle and outputs a candidate molecule; when the first predictor predicts and outputs the N atoms at a certain spatial position with the top N probabilities each time, the molecule generation model completes a cycle and outputs N M candidate molecules, and M is the number of cycles greater than 0.

5. The method according to claim 1, wherein the training process of the graph neural network, the first predictor, and the second predictor in the molecular generative model is as follows: Obtaining a training set, wherein the training set comprises a plurality of training subsets of proteins and their ligand complexes, each training subset comprising protein monomer sample information, protein pocket sample information, and ligand sample information; After the atom masking process, the ligand sample information is divided into masked molecular fragment sample information, masked atom atomic sample information, and masked bond bond sample information; The atomic sample information includes information about atoms and their positions; The protein monomer sample information, protein pocket sample information, atom sample information, and bond sample information of the training subset are input into the graph neural network for training, and the predicted information of the interaction between protein pockets and molecular fragments is output; the predicted information of the interaction between protein pockets and molecular fragments is input into the first predictor, and the predicted atom at a certain spatial position is output; The molecular fragment sample information and the predicted atom at a certain spatial position are input into the second predictor, which outputs the predicted bond; The first loss function is used to judge the convergence of predicted atoms and sample atoms; the second loss function is used to judge the convergence of predicted keys and sample keys; when both the first loss function and the second loss function meet the convergence conditions, the training of the graph neural network, the first predictor, and the second predictor is completed, otherwise, the training continues until the convergence conditions are met.

6. The method according to claim 1, further comprising: The performance of the predicted molecules is evaluated and the molecules are output together with the evaluation results.

7. The method according to claim 6, wherein the performance evaluation comprises: Drug-likeness assessment, synthetic accessibility assessment, and molecule and protein pocket docking score assessment.

8. A molecular generation device comprising: An acquisition module is used to obtain protein monomer information and protein pocket information of a target protein monomer; wherein the protein monomer information includes a protein monomer sequence and a protein monomer atom set, wherein the protein monomer atom set is a set of atoms constituting the protein monomer and their spatial positions; and the protein pocket information includes a protein pocket center position and a protein pocket size; The molecule generation module is used to input the protein monomer information and protein pocket information of the target protein monomer into the molecule generation model, and the molecule generation model outputs the candidate molecule; the molecule generation model includes a graph neural network, a first predictor, a second predictor, and a discriminator; the specific process is as follows: The graph neural network is used to embed information about interactions between protein pockets and molecular fragments; The first predictor predicts atoms at a certain spatial position based on the interaction information between protein pockets and molecular fragments embedded in the graph neural network; The second predictor predicts the bond connecting the atom and the previously predicted molecular fragment based on the previously predicted molecular fragment information and the atom at a certain spatial position predicted by the first predictor; The discriminator is used to determine that when the molecule generation is not completed, the atoms at a certain spatial position predicted this time and the bonds connecting them to the molecular fragments are returned to the graph neural network. The graph neural network embeds the interaction information between the new protein pocket and the molecular fragment, and then uses the first predictor and the second predictor to predict the atoms at a new spatial position and the bonds connecting them to the molecular fragments predicted previously. This process is repeated until the prediction is completed, and the final molecule is output as a candidate molecule. Among them, during the first prediction, the protein pocket structure and molecular fragment interaction information embedded by the graph neural network based on the protein monomer information and protein pocket information of the target protein monomer does not contain molecular fragment information. The first predictor predicts the first atom in a certain spatial position that constitutes the molecular fragment based on the protein pocket and molecular fragment interaction information first embedded by the graph neural network, and returns it to the graph neural network.

9. The apparatus according to claim 8, wherein the acquisition module comprises: an acquisition unit, configured to acquire a target protein monomer, wherein the target protein monomer includes protein monomer information; The extraction unit is used to determine the protein pocket based on the protein monomer information; and based on the protein pocket, determine the center position and size of the protein pocket.

10. The device according to claim 8, further comprising an evaluation module, configured to perform a performance evaluation on the predicted molecules after the prediction is completed to obtain an evaluation result of the molecules.

Images