A system for handling molecular generation tasks based on autoregressive models

By adopting a system based on autoregressive model in structural drug design, using Pocket2Mol, GraphBP and FLAG models to process molecular generation tasks, the problem of long design cycle and low efficiency of existing SBDD methods is solved, and faster and more efficient drugs are achieved Design process.

CN118866164BActive Publication Date: 2024-12-03BEIJING DP TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411319712.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-23
Publication Date
2024-12-03
Estimated Expiration
2044-09-23

AI Technical Summary

Technical Problem

The existing structural drug design (SBDD) method based on molecular motion simulation and quantum chemistry calculation has a long design cycle and low efficiency.

Method used

The system based on autoregressive model is adopted, including task scheduling module, task training module, task execution module, complex database and autoregressive model library. The molecular generation tasks are processed through Pocket2Mol, GraphBP and FLAG models, and five types of task interfaces ( ligand generation, linker generation, fragment growth, side chain modification and skeleton transition task interfaces) perform molecular generation tasks.

Benefits of technology

The drug design cycle is shortened, the design efficiency is improved, and the processing speed and efficiency of molecular generation tasks are improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118866164B_ABST
    Figure CN118866164B_ABST
Patent Text Reader

Abstract

The embodiment of the present invention relates to a system for processing molecular generation tasks based on an autoregressive model, the system comprising: a task scheduling module, a task training module, a task execution module, a complex database, an autoregressive model library and a task interface library; the task scheduling module prepares reference data for model training to obtain first reference data and sends it to the task training module, and prepares reference data for molecular generation tasks after confirming that the model training is finished to obtain second reference data and sends it to the task execution module, and feeds back the received generated molecular conformation to an external interface; the task training module trains a specified model in the autoregressive model library according to the reference data and the complex database; the task execution module executes the molecular generation task according to the reference data and the task interface library; the structure-based drug design process is improved by the present invention, so that the design cycle can be shortened and the design efficiency can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a system for processing molecule generation tasks based on an autoregressive model. Background Art

[0002] Structure-Based Drug Design (SBDD) is a design method that designs effective ligand molecules by analyzing the three-dimensional structure of the target protein. Conventional SBDD design is mainly achieved through molecular motion simulation and quantum chemical calculations. However, through practical applications, we found that the simulation / calculation cycle of this conventional implementation method is long and the design efficiency is not high. Summary of the invention

[0003] The purpose of the present invention is to provide a system for processing molecular generation tasks based on an autoregressive model in view of the defects of the prior art, the system comprising: a task scheduling module, a task training module, a task execution module, a complex database, an autoregressive model library and a task interface library; wherein the task scheduling module is used to, after receiving the molecular generation task data input from the external interface, first prepare the model training reference data according to the task data to obtain the corresponding reference data and send it to the task training module, then prepare the molecular generation task reference data according to the task data after the task training module confirms that the model training is completed to obtain the corresponding reference data and send it to the task execution module, and finally feed back the generated molecular conformation sent back by the task execution module to the external interface; the task training module is used to train the Pocket2Mol model, GraphBP model and FLAG model of the autoregressive model library according to the reference data and the complex database; the task execution module is used to perform molecular generation task execution processing according to the reference data and five types of task interfaces (ligand generation task interface, linker generation task interface, fragment growth task interface, side chain modification task interface and skeleton transition task interface) of the task interface library. By improving the conventional SBDD design process based on molecular motion simulation and quantum chemical calculation, the present invention can shorten the design cycle and improve the design efficiency.

[0004] To achieve the above-mentioned object, an embodiment of the present invention provides a system for processing molecular generation tasks based on an autoregressive model, the system comprising: a task scheduling module, a task training module, a task execution module, a complex database, an autoregressive model library and a task interface library;

[0005] The task scheduling module is connected to the task training module and the task execution module respectively; the task training module is connected to the compound database and the autoregressive model library respectively; the task execution module is connected to the task interface library; the task interface library is connected to the autoregressive model library;

[0006] The task scheduling module is used to receive the molecular generation task data input by the external interface as the corresponding first task data; and prepare the model training reference data according to the first task data to obtain the corresponding first reference data and send it to the task training module; and when the first training state sent by the task training module is received as the end of model training, prepare the molecular generation task reference data according to the first task data to obtain the corresponding second reference data and send it to the task execution module; and feed back the first generated molecular conformation sent by the task execution module to the external interface; the first generated molecular conformation is a three-dimensional molecular conformation of a generated molecule; the three-dimensional molecular conformation is composed of a corresponding atomic object set and a chemical bond object set; the atomic object set is composed of one or more atomic objects; each of the atomic objects includes an atomic index, an atomic type and an atomic coordinate; the chemical bond object set is composed of one or more chemical bond objects; each of the chemical bond objects includes a chemical bond index, a chemical bond type and a chemical bond atom index group; the chemical bond atom index group is composed of the atomic indexes of the two atomic objects corresponding to the current chemical bond object;

[0007] The task training module is used to train the specified model in the autoregressive model library according to the first reference data and the composite database to obtain the corresponding first training state and send it to the task scheduling module;

[0008] The task execution module is used for performing molecular generation task execution processing according to the second reference data and the task interface library to obtain the corresponding first generated molecular conformation and send it to the task scheduling module;

[0009] The complex database includes a plurality of complex data sets; each of the complex data sets corresponds to a data set index; each of the complex data sets consists of a plurality of complex records; each of the complex records corresponds to a protein-ligand complex, which is formed by the combination of a protein pocket and a ligand molecule; the complex record includes a complex conformation, a protein pocket conformation and a ligand molecule conformation; the complex conformation, the protein pocket conformation and the ligand molecule conformation are respectively the three-dimensional molecular conformations of the protein-ligand complex, the protein pocket and the ligand molecule corresponding to the current record, and the protein pocket conformation and the ligand molecule conformation are each a sub-conformation of the complex conformation;

[0010] The autoregressive model library includes Pocket2Mol model, GraphBP model and FLAG model;

[0011] The task interface library includes five types of molecular generation task interfaces, namely, ligand generation task interface, linker generation task interface, fragment growth task interface, side chain modification task interface and backbone transition task interface.

[0012] Preferably, the first task data includes a first task type, a first model type, a first data set index and a first task information set;

[0013] The first task type includes ligand generation tasks, linker generation tasks, fragment growth tasks, side chain modification tasks and backbone transition tasks;

[0014] The first model type includes a first type, a second type and a third type;

[0015] When the first task type is a ligand generation task, the corresponding first task information set consists of a first protein pocket conformation; the first protein pocket conformation is the three-dimensional molecular conformation of a protein pocket;

[0016] When the first task type is a connector generation task, the corresponding first task information set includes a first protein-ligand complex conformation, a first pocket atom set and a first ligand atom set; the first protein-ligand complex conformation is the three-dimensional molecular conformation of a protein-ligand complex A; the protein-ligand complex A consists of a protein pocket P A and a ligand molecule M A Composition, the ligand molecule M A Composed of at least two or more molecular fragments F A The first pocket atom set and the first ligand atom set are respectively the protein pocket P A and the ligand molecule M A The set of atomic indices of ;

[0017] When the first task type is a fragment growth task, the corresponding first task information set includes a second protein-ligand complex conformation, a second pocket atom set, a second ligand atom set and a first fragment atom set to be grown; the second protein-ligand complex conformation is the three-dimensional molecular conformation of a protein-ligand complex B; the protein-ligand complex B consists of a protein pocket P B and a ligand molecule M B Composition, the ligand molecule M B The molecular fragment that needs to be grown is recorded as the fragment to be grown F B The second pocket atom set, the second ligand atom set and the first to-be-grown fragment atom set are respectively the protein pocket P B , the ligand molecule M B and the fragment to be grown FB The set of atomic indices of ;

[0018] When the first task type is a side chain modification task, the corresponding first task information set includes a third protein-ligand complex conformation, a third pocket atom set and a third ligand atom set; the third protein-ligand complex conformation is the three-dimensional molecular conformation of a protein-ligand complex C; the protein-ligand complex C consists of a protein pocket P C and a ligand molecule M C The third pocket atom set and the third ligand atom set are respectively the protein pocket P C and the ligand molecule M C The set of atomic indices of ;

[0019] When the first task type is a skeleton transition task, the corresponding first task information set includes a fourth protein-ligand complex conformation, a fourth pocket atom set, a fourth ligand atom set and a first to-be-transitioned fragment atom set; the fourth protein-ligand complex conformation is the three-dimensional molecular conformation of a protein-ligand complex D; the protein-ligand complex D consists of a protein pocket P D and a ligand molecule M D Composition, the ligand molecule M D The skeleton fragment that needs to be transitioned on the main skeleton is recorded as the fragment to be transitioned F D The fourth pocket atom set, the fourth ligand atom set and the first to-be-transitioned fragment atom set are respectively the protein pocket P D , the ligand molecule M D and the fragment to be transitioned F D The set of atomic indices.

[0020] Preferably, the task scheduling module is specifically used for sending the first reference data corresponding to the first task type, the first model type and the first data set index of the first task data to the task training module when the first reference data corresponding to the first reference data prepared by performing model training reference data according to the first task data is sent to the task training module;

[0021] The task scheduling module is specifically used to send the corresponding second reference data composed of the first task type, the first model type and the first task information set of the first task data to the task training module when the corresponding second reference data is prepared by performing molecular generation task reference data according to the first task data and sent to the task execution module.

[0022] Preferably, the task training module is specifically used for, when the first training state is obtained by training the designated model in the autoregressive model library according to the first reference data and the composite database and sent to the task scheduling module,

[0023] Extracting the corresponding first task type, first model type and first data set index from the first reference data;

[0024] and taking the composite data set whose data set index in the composite database matches the first data set index as the corresponding first data set;

[0025] and identifying the first task type;

[0026] When the first task type is a ligand generation task, the first model type is identified; if the first model type is the first type, the Pocket2Mol model is trained according to the first data set in a training manner of the ligand generation task; if the first model type is the second type, the GraphBP model is trained according to the first data set in a training manner of the ligand generation task; if the first model type is the third type, the FLAG model is trained according to the first data set in a training manner of the ligand generation task; the Pocket2Mol model, the GraphBP model and the FLAG model trained in the training manner of the ligand generation task are all used to perform corresponding ligand molecule generation processing according to the input protein pocket conformation and ligand context information and output the three-dimensional molecular conformation of a ligand molecule as the corresponding ligand molecular conformation;

[0027] When the first task type is a connector generation task, the first model type is identified; if the first model type is the first type, the Pocket2Mol model is trained according to the first data set in a training manner of the connector generation task; if the first model type is the second type, the GraphBP model is trained according to the first data set in a training manner of the connector generation task; if the first model type is the third type, the FLAG model is trained according to the first data set in a training manner of the connector generation task; the Pocket2Mol model, the GraphBP model and the FLAG model trained according to the training manner of the connector generation task are all used to perform corresponding connector generation processing according to the input protein pocket conformation and ligand context information and output the three-dimensional molecular conformation of a new ligand molecule as the corresponding ligand molecular conformation;

[0028] When the first task type is a fragment growth task, the first model type is identified; if the first model type is the first type, the Pocket2Mol model is trained according to the first data set in a training manner of the fragment growth task; if the first model type is the second type, the GraphBP model is trained according to the first data set in a training manner of the fragment growth task; if the first model type is the third type, the FLAG model is trained according to the first data set in a training manner of the fragment growth task; the Pocket2Mol model, the GraphBP model and the FLAG model trained in the training manner of the fragment growth task are all used to perform corresponding fragment growth processing according to the input protein pocket conformation and ligand context information and output the three-dimensional molecular conformation of a new ligand molecule as the corresponding ligand molecular conformation;

[0029] When the first task type is a side chain modification task, the first model type is identified; if the first model type is the first type, the Pocket2Mol model is trained according to the first data set in a training manner for the side chain modification task; if the first model type is the second type, the GraphBP model is trained according to the first data set in a training manner for the side chain modification task; if the first model type is the third type, the FLAG model is trained according to the first data set in a training manner for the side chain modification task; the Pocket2Mol model, the GraphBP model and the FLAG model trained according to the training manner for the side chain modification task are all used to perform corresponding side chain modification processing according to the input protein pocket conformation and ligand context information and output the three-dimensional molecular conformation of a new ligand molecule as the corresponding ligand molecular conformation;

[0030] When the first task type is a skeleton transition task, the first model type is identified; if the first model type is the first type, the Pocket2Mol model is trained according to the first data set in a training manner of the skeleton transition task; if the first model type is the second type, the GraphBP model is trained according to the first data set in a training manner of the skeleton transition task; if the first model type is the third type, the FLAG model is trained according to the first data set in a training manner of the skeleton transition task; the Pocket2Mol model, the GraphBP model and the FLAG model trained in the training manner of the skeleton transition task are all used to perform corresponding skeleton fragment transition processing according to the input protein pocket conformation and ligand context information and output the three-dimensional molecular conformation of a new ligand molecule as the corresponding ligand molecular conformation;

[0031] And when the current model training is finished, the first training state specifically set as the end of the model training is sent to the task scheduling module.

[0032] Preferably, the task execution module is specifically used for, when the molecular generation task execution processing is performed according to the second reference data and the task interface library to obtain the corresponding first generated molecular conformation and sending it to the task scheduling module,

[0033] identifying the first task type in the second reference data;

[0034] If the first task type is a ligand generation task, sending the second reference data to the ligand generation task interface; and sending the first generated molecular conformation sent back by the ligand generation task interface to the task scheduling module;

[0035] If the first task type is a linker generation task, sending the second reference data to the linker generation task interface; and sending the first generated molecular conformation sent back by the linker generation task interface to the task scheduling module;

[0036] If the first task type is a fragment growth task, sending the second reference data to the fragment growth task interface; and sending the first generated molecular conformation sent back by the fragment growth task interface to the task scheduling module;

[0037] If the first task type is a side chain modification task, sending the second reference data to the side chain modification task interface; and sending the first generated molecular conformation sent back by the side chain modification task interface to the task scheduling module;

[0038] If the first task type is a skeleton transition task, the second reference data is sent to the skeleton transition task interface; and the first generated molecular conformation sent back by the skeleton transition task interface is sent to the task scheduling module.

[0039] Preferably, the ligand generation task interface is used to extract the corresponding first model type and the first protein pocket conformation from the received second reference data;

[0040] and taking the first protein pocket conformation as the corresponding current protein pocket conformation; and setting the corresponding current ligand context information to be empty;

[0041] and identifying the first model type; if the first model type is the first type, calling the Pocket2Mol model to perform corresponding ligand molecule generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding first ligand molecule conformation; if the first model type is the second type, calling the GraphBP model to perform corresponding ligand molecule generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding first ligand molecule conformation; if the first model type is the third type, calling the FLAG model to perform corresponding ligand molecule generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding first ligand molecule conformation;

[0042] And the obtained first ligand molecular conformation is sent to the task execution module as the corresponding first generated molecular conformation.

[0043] Preferably, the connector generation task interface is used to extract the corresponding first model type, the first protein-ligand complex conformation, the first pocket atom set and the first ligand atom set from the received second reference data;

[0044] and extracting the corresponding protein pocket conformation of the first protein-ligand complex conformation based on the first pocket atom set to obtain the corresponding current protein pocket conformation;

[0045] and extracting the corresponding ligand molecular conformation of the first protein-ligand complex conformation based on the first ligand atom set and using the extraction result as the corresponding current ligand context information;

[0046] and identifying the first model type; if the first model type is the first type, calling the Pocket2Mol model to perform corresponding linker generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding second ligand molecular conformation; if the first model type is the second type, calling the GraphBP model to perform corresponding linker generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding second ligand molecular conformation; if the first model type is the third type, calling the FLAG model to perform corresponding linker generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding second ligand molecular conformation;

[0047] And the obtained second ligand molecular conformation is sent to the task execution module as the corresponding first generated molecular conformation.

[0048] Preferably, the fragment growth task interface is used to extract the corresponding first model type, the second protein-ligand complex conformation, the second pocket atom set, the second ligand atom set and the first to-be-grown fragment atom set from the received second reference data;

[0049] and extracting the corresponding protein pocket conformation of the second protein-ligand complex conformation based on the second pocket atom set to obtain the corresponding current protein pocket conformation;

[0050] and extracting the corresponding ligand molecular conformation of the second protein-ligand complex conformation based on the second ligand atom set to obtain the corresponding current ligand molecular conformation;

[0051] Based on the first set of atoms of the fragment to be grown, the fragment conformation corresponding to the fragment to be grown in the current ligand molecular conformation is deleted; and the current ligand molecular conformation after the fragment deletion is completed is used as the corresponding current ligand context information;

[0052] and identifying the first model type; if the first model type is the first type, calling the Pocket2Mol model to perform corresponding fragment growth processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding third ligand molecular conformation; if the first model type is the second type, calling the GraphBP model to perform corresponding fragment growth processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding third ligand molecular conformation; if the first model type is the third type, calling the FLAG model to perform corresponding fragment growth processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding third ligand molecular conformation;

[0053] The obtained third ligand molecular conformation is sent to the task execution module as the corresponding first generated molecular conformation.

[0054] Preferably, the side chain modification task interface is used to extract the corresponding first model type, the third protein-ligand complex conformation, the third pocket atom set and the third ligand atom set from the received second reference data;

[0055] and extracting the corresponding protein pocket conformation of the third protein-ligand complex conformation based on the third pocket atom set to obtain the corresponding current protein pocket conformation;

[0056] and extracting the corresponding ligand molecular conformation of the third protein-ligand complex conformation based on the third ligand atom set and using the extraction result as the corresponding current ligand context information;

[0057] and identifying the first model type; if the first model type is the first type, calling the Pocket2Mol model to perform corresponding side chain modification processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding fourth ligand molecular conformation; if the first model type is the second type, calling the GraphBP model to perform corresponding side chain modification processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding fourth ligand molecular conformation; if the first model type is the third type, calling the FLAG model to perform corresponding side chain modification processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding fourth ligand molecular conformation;

[0058] And the obtained fourth ligand molecular conformation is sent to the task execution module as the corresponding first generated molecular conformation.

[0059] Preferably, the backbone transition task interface is used to extract the corresponding first model type, the fourth protein-ligand complex conformation, the fourth pocket atom set, the fourth ligand atom set and the first to-be-transitioned fragment atom set from the received second reference data;

[0060] and extracting the corresponding protein pocket conformation of the fourth protein-ligand complex conformation based on the fourth pocket atom set to obtain the corresponding current protein pocket conformation;

[0061] and extracting the corresponding ligand molecular conformation of the fourth protein-ligand complex conformation based on the fourth ligand atom set to obtain the corresponding current ligand molecular conformation;

[0062] Based on the first set of atoms of the fragments to be transitioned, the fragment conformation corresponding to the fragment to be transitioned in the current ligand molecular conformation is deleted; and the current ligand molecular conformation after the fragment deletion is completed is used as the corresponding current ligand context information;

[0063] And identify the first model type; if the first model type is the first type, call the Pocket2Mol model to perform corresponding backbone fragment transition processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding fifth ligand molecular conformation; if the first model type is the second type, call the GraphBP model to perform corresponding backbone fragment transition processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding fifth ligand molecular conformation; if the first model type is the third type, call the FLAG model to perform corresponding backbone fragment transition processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding fifth ligand molecular conformation;

[0064] And the obtained fifth ligand molecular conformation is sent to the task execution module as the corresponding first generated molecular conformation.

[0065] The embodiment of the present invention provides a system for processing molecular generation tasks based on an autoregressive model, including: a task scheduling module, a task training module, a task execution module, a complex database, an autoregressive model library and a task interface library; wherein the task scheduling module is used to prepare model training reference data according to the task data after receiving the molecular generation task data input from the external interface, obtain the corresponding reference data and send it to the task training module, and then prepare the molecular generation task reference data according to the task data when the task training module confirms that the model training is finished, obtain the corresponding reference data and send it to the task execution module, and finally feedback the generated molecular conformation sent back by the task execution module to the external interface; the task training module is used to train the Pocket2Mol model, GraphBP model and FLAG model of the autoregressive model library according to the reference data and the complex database; the task execution module is used to perform molecular generation task execution processing according to the reference data and five types of task interfaces (ligand generation task interface, linker generation task interface, fragment growth task interface, side chain modification task interface and skeleton transition task interface) of the task interface library. The embodiment of the present invention can improve the conventional SBDD design process, shorten the design cycle and improve the design efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 A module structure diagram of a system for processing molecular generation tasks based on an autoregressive model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] The embodiment of the present invention provides a system for processing molecular generation tasks based on an autoregressive model, such as Figure 1 As shown in the module structure diagram of a system for processing molecular generation tasks based on an autoregressive model provided in an embodiment of the present invention, the system mainly includes: a task scheduling module 1, a task training module 2, a task execution module 3, a complex database 4, an autoregressive model library 5 and a task interface library 6.

[0069] The connection relationship between the modules in the system of the embodiment of the present invention is: the task scheduling module 1 is connected to the task training module 2 and the task execution module 3 respectively; the task training module 2 is connected to the composite database 4 and the autoregressive model library 5 respectively; the task execution module 3 is connected to the task interface library 6; the task interface library 6 is connected to the autoregressive model library 5.

[0070] (I) Task scheduling module 1:

[0071] The task scheduling module 1 is used to receive the molecular generation task data input by the external interface as the corresponding first task data; and prepare the model training reference data according to the first task data to obtain the corresponding first reference data and send it to the task training module 2; and when the first training state sent by the task training module 2 is received as the end of the model training, prepare the molecular generation task reference data according to the first task data to obtain the corresponding second reference data and send it to the task execution module 3; and feed back the first generated molecular conformation sent by the task execution module 3 to the external interface.

[0072] Here, the external interface of the embodiment of the present invention may be a human-computer interaction interface or a processing interface for connecting other machines or systems with the system of the embodiment of the present invention. The external interface may be a software API interface or a hardware API interface.

[0073] The first generated molecular conformation is a three-dimensional molecular conformation of a generated molecule, which in the embodiment of the present invention is a de novo generated or modified ligand molecular conformation; all conformations mentioned in the embodiment of the present invention refer to three-dimensional molecular conformations, composed of corresponding atomic object sets and chemical bond object sets; the atomic object set is composed of one or more atomic objects; each atomic object includes an atomic index, an atomic type and atomic coordinates; the chemical bond object set is composed of one or more chemical bond objects; each chemical bond object includes a chemical bond index, a chemical bond type and a chemical bond atom index group; the chemical bond atom index group is composed of the atomic indices of two atomic objects corresponding to the current chemical bond object.

[0074] The first task data of the embodiment of the present invention includes a first task type, a first model type, a first data set index and a first task information set; wherein,

[0075] 1) The first task type includes ligand generation tasks, linker generation tasks, fragment growth tasks, side chain modification tasks and backbone transition tasks;

[0076] The ligand generation task is to use the autoregressive prediction model (Pocket2Mol, GraphBP, FLAG) to predict the full conformation information of the ligand molecule based on the known protein pocket P (protein pocket) and the latest ligand context information C (context) through the internal loop iteration of the model;

[0077] The linker generation task is to use the autoregressive prediction model to add missing linkers (fragments consisting of two or more atoms) in the current ligand molecular conformation according to the known protein pocket P and the latest ligand context information C through internal loop iteration of the model;

[0078] The fragment growth task is to use the autoregressive prediction model to perform fragment growth processing on a missing molecular fragment in the current ligand molecular conformation based on the known protein pocket P and the latest ligand context information C through the internal loop iteration of the model;

[0079] The side chain (subchain) modification task is to use the autoregressive prediction model to fine-tune all side chains in the current ligand molecular conformation according to the known protein pocket P and the latest ligand context information C through the internal loop iteration of the model. It should be noted that the autoregressive prediction model of the embodiment of the present invention will regard all terminal non-cyclic structures as side chains through training;

[0080] The scaffold hopping task is to use the autoregressive prediction model to replace / reset a skeleton fragment on the skeleton structure of the current ligand molecular conformation according to the known protein pocket P and the latest ligand context information C through the internal loop iteration of the model;

[0081] 2) The first model type includes the first type, the second type and the third type; when the first model type is the first model, it indicates that the Pocket2Mol model 51 is specified as the autoregressive prediction model of this task, when the first model type is the first model, it indicates that the GraphBP model 52 is specified as the autoregressive prediction model of this task, and when the first model type is the first model, it indicates that the FLAG model 53 is specified as the autoregressive prediction model of this task;

[0082] 3) The first dataset index is the composite dataset index specified for this task;

[0083] 4) The specific content of the first task information set is related to the first task type. The corresponding relationship between the two is as follows:

[0084] When the first task type is a ligand generation task, the corresponding first task information set consists of a first protein pocket conformation; the first protein pocket conformation is a three-dimensional molecular conformation of a protein pocket;

[0085] When the first task type is a connector generation task, the corresponding first task information set includes a first protein-ligand complex conformation, a first pocket atom set, and a first ligand atom set; the first protein-ligand complex conformation is a three-dimensional molecular conformation of a protein-ligand complex A; the protein-ligand complex A consists of a protein pocket P A and a ligand molecule M A Composition, ligand molecule M A Composed of at least two or more molecular fragments F A The first pocket atom set and the first ligand atom set are protein pocket P A and ligand molecule M A The set of atomic indices of ;

[0086] When the first task type is a fragment growth task, the corresponding first task information set includes the second protein-ligand complex conformation, the second pocket atom set, the second ligand atom set and the first fragment atom set to be grown; the second protein-ligand complex conformation is a three-dimensional molecular conformation of a protein-ligand complex B; the protein-ligand complex B consists of a protein pocket P B and a ligand molecule M B Composition, ligand molecule M B The molecular fragment that needs to be grown is recorded as the fragment to be grown F BThe second pocket atom set, the second ligand atom set and the first fragment atom set to be grown are protein pocket P B , ligand molecule M B and the fragment to be grown F B The set of atomic indices of ;

[0087] When the first task type is a side chain modification task, the corresponding first task information set includes a third protein-ligand complex conformation, a third pocket atom set, and a third ligand atom set; the third protein-ligand complex conformation is a three-dimensional molecular conformation of a protein-ligand complex C; the protein-ligand complex C consists of a protein pocket P C and a ligand molecule M C The third pocket atom set and the third ligand atom set are protein pocket P C and ligand molecule M C The set of atomic indices of ;

[0088] When the first task type is a skeleton transition task, the corresponding first task information set includes the fourth protein-ligand complex conformation, the fourth pocket atom set, the fourth ligand atom set and the first to-be-transitioned fragment atom set; the fourth protein-ligand complex conformation is a three-dimensional molecular conformation of a protein-ligand complex D; the protein-ligand complex D consists of a protein pocket P D and a ligand molecule M D Composition, ligand molecule M D The skeleton fragment that needs to be transitioned on the main skeleton is recorded as the fragment to be transitioned F D The fourth pocket atom set, the fourth ligand atom set and the first transition fragment atom set are protein pocket P D , ligand molecule M D and the fragment to be transitioned F D The set of atomic indices.

[0089] In a specific implementation method of an embodiment of the present invention, the task scheduling module 1 is specifically used to send the corresponding first reference data composed of the first task type, the first model type and the first data set index of the first task data to the task training module 2 when preparing the model training reference data according to the first task data to obtain the corresponding first reference data and sending it to the task training module 2.

[0090] Here, the first reference data of the embodiment of the present invention includes a first task type, a first model type and a first data set index; the first reference data is used to guide the task training module 2 to perform corresponding model training.

[0091] In another specific implementation method of an embodiment of the present invention, the task scheduling module 1 is specifically used to send the corresponding second reference data composed of the first task type, the first model type and the first task information set of the first task data to the task training module 2 when preparing the molecular generation task reference data according to the first task data and sending it to the task execution module 3.

[0092] Here, the second reference data of the embodiment of the present invention includes the first task type, the first model type and the first task information set; the second reference data is used to guide the task execution module 3 to execute the corresponding molecule generation task.

[0093] (II) Task training module 2:

[0094] The task training module 2 is used to train a specified model in the autoregressive model library 5 according to the first reference data and the composite database 4 to obtain a corresponding first training state and send it to the task scheduling module 1 .

[0095] In another specific implementation of the embodiment of the present invention, the task training module 2 is specifically used to train the specified model in the autoregressive model library 5 according to the first reference data and the composite database 4 to obtain the corresponding first training state and send it to the task scheduling module 1:

[0096] Step A1, extracting a corresponding first task type, a first model type and a first data set index from first reference data;

[0097] Step A2, and taking the composite data set 41 whose data set index in the composite data set database 4 matches the first data set index as the corresponding first data set;

[0098] Step A3, identifying the first task type;

[0099] Step A4, when the first task type is a ligand generation task, the first model type is identified; if the first model type is the first type, the Pocket2Mol model 51 is trained according to the first data set in the training method of the ligand generation task; if the first model type is the second type, the GraphBP model 52 is trained according to the first data set in the training method of the ligand generation task; if the first model type is the third type, the FLAG model 53 is trained according to the first data set in the training method of the ligand generation task;

[0100] It should be noted that the Pocket2Mol model 51, GraphBP model 52 and FLAG model 53 trained in the training method of the ligand generation task are all used to perform corresponding ligand molecule generation processing according to the input protein pocket conformation and ligand context information and output a three-dimensional molecular conformation of a ligand molecule recorded as the corresponding ligand molecular conformation;

[0101] Step A5, when the first task type is a connector generation task, the first model type is identified; if the first model type is the first type, the Pocket2Mol model 51 is trained according to the first data set in the training method of the connector generation task; if the first model type is the second type, the GraphBP model 52 is trained according to the training method of the connector generation task in the first data set; if the first model type is the third type, the FLAG model 53 is trained according to the training method of the connector generation task in the first data set;

[0102] It should be noted that the Pocket2Mol model 51, GraphBP model 52 and FLAG model 53 trained in the training method of the linker generation task are all used to perform corresponding linker generation processing according to the input protein pocket conformation and ligand context information and output a three-dimensional molecular conformation of a new ligand molecule recorded as the corresponding ligand molecular conformation;

[0103] Step A6, when the first task type is a fragment growing task, the first model type is identified; if the first model type is the first type, the Pocket2Mol model 51 is trained according to the first data set in the training method of the fragment growing task; if the first model type is the second type, the GraphBP model 52 is trained according to the training method of the fragment growing task in the first data set; if the first model type is the third type, the FLAG model 53 is trained according to the training method of the fragment growing task in the first data set;

[0104] It should be noted that the Pocket2Mol model 51, GraphBP model 52 and FLAG model 53 trained in the training mode of the fragment growth task are all used to perform corresponding fragment growth processing according to the input protein pocket conformation and ligand context information and output a three-dimensional molecular conformation of a new ligand molecule recorded as the corresponding ligand molecular conformation;

[0105] Step A7, when the first task type is a side chain modification task, the first model type is identified; if the first model type is the first type, the Pocket2Mol model 51 is trained according to the first data set in the training method of the side chain modification task; if the first model type is the second type, the GraphBP model 52 is trained according to the training method of the side chain modification task in the first data set; if the first model type is the third type, the FLAG model 53 is trained according to the training method of the side chain modification task in the first data set;

[0106] It should be noted that the Pocket2Mol model 51, GraphBP model 52 and FLAG model 53 trained in the training method of the side chain modification task are all used to perform corresponding side chain modification processing according to the input protein pocket conformation and ligand context information and output a three-dimensional molecular conformation of a new ligand molecule recorded as the corresponding ligand molecular conformation;

[0107] Step A8, when the first task type is a skeleton transition task, the first model type is identified; if the first model type is the first type, the Pocket2Mol model 51 is trained according to the first data set in the training method of the skeleton transition task; if the first model type is the second type, the GraphBP model 52 is trained according to the first data set in the training method of the skeleton transition task; if the first model type is the third type, the FLAG model 53 is trained according to the first data set in the training method of the skeleton transition task;

[0108] It should be noted that the Pocket2Mol model 51, GraphBP model 52 and FLAG model 53 trained in the training method of the backbone transition task are all used to perform corresponding backbone fragment transition processing according to the input protein pocket conformation and ligand context information and output a three-dimensional molecular conformation of a new ligand molecule recorded as the corresponding ligand molecular conformation;

[0109] Step A9, when the current model training is finished, the first training state specifically set as the end of the model training is sent to the task scheduling module 1.

[0110] It should also be noted that the training method according to the ligand generation task mentioned in the above step A4 refers to training the Pocket2Mol model 51, GraphBP model 52 and FLAG model 53 with the goal of enabling the model to achieve the ligand generation task function; the task training module 2 adopts a supervised method during training. The task training module 2 is specifically used for training the Pocket2Mol model 51, GraphBP model 52 or FLAG model 53 according to the first data set in the training method of the ligand generation task:

[0111] Step B1, record the currently trained Pocket2Mol model 51, GraphBP model 52 or FLAG model 53 as the corresponding current model;

[0112] Step B2, based on a preset first split ratio (e.g., 8:2), the first data set is divided into two sub-data sets, which are recorded as a corresponding first training set and a first evaluation set;

[0113] Wherein, both the first training set and the first evaluation set consist of a plurality of composite records, and a ratio of the total number of records in the first training set to the total number of records in the first evaluation set satisfies a first split ratio;

[0114] Step B3, taking the first composite record of the first training set as the corresponding current training record;

[0115] Step B4, taking the protein pocket conformation and ligand molecular conformation of the current training record as the corresponding current protein pocket conformation and current label ligand molecular conformation; and setting the corresponding current ligand context information to empty;

[0116] Step B5, the current model performs corresponding ligand molecule generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding current predicted molecular conformation;

[0117] Step B6, bringing the current predicted molecular conformation and the current label ligand molecular conformation into the model loss function corresponding to the current model to calculate and obtain the corresponding first loss value;

[0118] Here, the model loss function of the Pocket2Mol model 51 is given in the technical document A "Pocket2Mol: EfficientMolecular Sampling Based on 3D Protein Pockets"; the model loss function of the GraphBP model 52 is given in the technical document B "Generating 3D Molecules for Target Protein Binding"; the model loss function of the FLAG model 53 is given in the technical document C "MOLECULE GENERATION FOR TARGET PROTEINBINDING WITH STRUCTURAL MOTIFS"; no further repetition is given here;

[0119] Step B7, identifying whether the first loss value satisfies a preset first loss value range; if the first loss value satisfies the first loss value range, identifying whether the current training record is the last composite record of the first training set, if so, turning to step B8, if not, extracting the next composite record of the first training set as a new current training record and returning to step B4; if the first loss value does not satisfy the first loss value range, based on a preset model parameter optimization algorithm or optimizer (such as SGD optimization algorithm / optimizer, ADAM optimization algorithm / optimizer, etc.), a round of modulation is performed on the model parameters of the current model in the direction of minimizing the loss function of the current model, and returning to step B5 at the end of this round of modulation;

[0120] Step B8, taking each complex record of the first evaluation set as the corresponding current evaluation record in turn; taking the protein pocket conformation and ligand molecular conformation of the current evaluation record as the corresponding current protein pocket conformation and current label ligand molecular conformation; setting the corresponding current ligand context information to be empty; and the current model performs corresponding ligand molecule generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding current predicted molecular conformation; and the current predicted molecular conformation and the corresponding current label ligand molecular conformation form a corresponding first prediction-label pair; and bringing all the obtained first prediction-label pairs into the preset first evaluation function for calculation to obtain the corresponding first evaluation value;

[0121] Here, the first evaluation function includes at least conventional MAE, RMSE and other functions;

[0122] Step B9, identifying whether the first evaluation value satisfies a preset first evaluation value range; if the first evaluation value does not satisfy the first evaluation value range, returning to step B1; if the first evaluation value satisfies the first evaluation value range, confirming that the current model training is completed.

[0123] It should also be noted that the training method according to the connector generation task mentioned in the above step A5 refers to training the Pocket2Mol model 51, GraphBP model 52 and FLAG model 53 with the goal of the model being able to realize the connector generation task function; the task training module 2 adopts a supervised method during training. The task training module 2 is specifically used for training the Pocket2Mol model 51, GraphBP model 52 or FLAG model 53 according to the first data set in the training method according to the connector generation task:

[0124] Step C1, recording the currently trained Pocket2Mol model 51, GraphBP model 52 or FLAG model 53 as the corresponding current model;

[0125] Step C2, based on a preset second split ratio (e.g., 8:2), the first data set is divided into two sub-data sets, which are recorded as corresponding second training sets and second evaluation sets;

[0126] Wherein, both the second training set and the second evaluation set are composed of a plurality of composite records, and the ratio of the total number of records in the second training set to the total number of records in the second evaluation set satisfies the second segmentation ratio;

[0127] Step C3, taking the first composite record of the second training set as the corresponding current training record;

[0128] Step C4, taking the protein pocket conformation and ligand molecular conformation of the current training record as the corresponding current protein pocket conformation and current label ligand molecular conformation; and performing connector object identification on the current label ligand molecular conformation to obtain one or more corresponding connector objects to form a corresponding connector object set; and performing a copy of the current label ligand molecular conformation to obtain the corresponding current copy molecular conformation; and randomly selecting one or more connector objects from the connector object set as mask connectors; and performing mask processing on the atomic types of the atomic objects corresponding to all mask connectors in the current copy molecular conformation, and performing Gaussian noise addition processing on the atomic coordinates; and taking the current copy molecular conformation with the atomic type mask and atomic coordinate noise added as the corresponding current ligand context information;

[0129] Step C5, the current model performs corresponding linker generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding current predicted molecular conformation;

[0130] Step C6, bringing the current predicted molecular conformation and the current label ligand molecular conformation into the model loss function corresponding to the current model to calculate and obtain the corresponding second loss value;

[0131] Step C7, identifying whether the second loss value satisfies a preset second loss value range; if the second loss value satisfies the second loss value range, identifying whether the current training record is the last composite record of the second training set, if so, turning to step C8, if not, extracting the next composite record of the second training set as a new current training record and returning to step C4; if the second loss value does not satisfy the second loss value range, based on a preset model parameter optimization algorithm or optimizer (such as SGD optimization algorithm / optimizer, ADAM optimization algorithm / optimizer, etc.), a round of modulation is performed on the model parameters of the current model in the direction of minimizing the loss function of the current model, and returning to step C5 at the end of this round of modulation;

[0132] Step C8, taking each complex record of the second evaluation set as the corresponding current evaluation record in turn; taking the protein pocket conformation and ligand molecular conformation of the current evaluation record as the corresponding current protein pocket conformation and current label ligand molecular conformation; and performing connector object identification on the current label ligand molecular conformation to obtain one or more corresponding connector objects to form a corresponding connector object set; and performing a copy of the current label ligand molecular conformation to obtain the corresponding current copied molecular conformation; and randomly selecting one or more connector objects from the connector object set as mask connectors; and performing a copy of the current copied molecular conformation with all mask connectors. The atomic type of the atomic object corresponding to the connector is masked, and the atomic coordinates are subjected to Gaussian noise addition processing; the current copied molecular conformation after the atomic type masking and atomic coordinate noise addition is used as the corresponding current ligand context information; the current model performs corresponding connector generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding current predicted molecular conformation; the current predicted molecular conformation and the corresponding current label ligand molecular conformation form a corresponding second prediction-label pair; all the obtained second prediction-label pairs are brought into the preset second evaluation function for calculation to obtain the corresponding second evaluation value;

[0133] Here, the second evaluation function includes at least conventional MAE, RMSE and other functions;

[0134] Step C9, identifying whether the second evaluation value satisfies a preset second evaluation value range; if the second evaluation value does not satisfy the second evaluation value range, returning to step C1; if the second evaluation value satisfies the second evaluation value range, confirming that the current model training is completed.

[0135] It should also be noted that the training method according to the fragment growth task mentioned in the above step A6 refers to training the Pocket2Mol model 51, GraphBP model 52 and FLAG model 53 with the goal of the model being able to realize the fragment growth task function; the task training module 2 adopts a supervised method during training. The task training module 2 is specifically used for training the Pocket2Mol model 51, GraphBP model 52 or FLAG model 53 according to the first data set in the training method according to the fragment growth task:

[0136] Step D1, record the currently trained Pocket2Mol model 51, GraphBP model 52 or FLAG model 53 as the corresponding current model;

[0137] Step D2, based on a preset third split ratio (for example, 8:2), the first data set is divided into two sub-data sets recorded as a corresponding third training set and a third evaluation set;

[0138] Wherein, the third training set and the third evaluation set are both composed of a plurality of composite records, and the ratio of the total number of records in the third training set to the total number of records in the third evaluation set satisfies a third segmentation ratio;

[0139] Step D3, taking the first composite record of the third training set as the corresponding current training record;

[0140] Step D4, taking the protein pocket conformation and ligand molecular conformation of the current training record as the corresponding current protein pocket conformation and current label ligand molecular conformation; and performing molecular fragment object recognition on the current label ligand molecular conformation to obtain one or more corresponding molecular fragment objects, and taking the molecular fragment object with the shortest length as the corresponding current molecular fragment; and copying the current label ligand molecular conformation once to obtain the corresponding current copied molecular conformation; and deleting all atomic objects corresponding to the current molecular fragment in the current copied molecular conformation; and taking the deleted current copied molecular conformation as the corresponding current ligand context information;

[0141] Step D5, the current model performs corresponding fragment growth processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding current predicted molecular conformation;

[0142] Step D6, bringing the current predicted molecular conformation and the current label ligand molecular conformation into the model loss function corresponding to the current model to calculate and obtain the corresponding third loss value;

[0143] Step D7, identifying whether the third loss value satisfies a preset third loss value range; if the third loss value satisfies the third loss value range, identifying whether the current training record is the last composite record of the third training set, if so, turning to step D8, if not, extracting the next composite record of the third training set as a new current training record and returning to step D4; if the third loss value does not satisfy the third loss value range, based on a preset model parameter optimization algorithm or optimizer (such as SGD optimization algorithm / optimizer, ADAM optimization algorithm / optimizer, etc.), a round of modulation is performed on the model parameters of the current model in the direction of minimizing the loss function of the current model, and returning to step D5 at the end of this round of modulation;

[0144] Step D8, taking each complex record of the third evaluation set as the corresponding current evaluation record in turn; and taking the protein pocket conformation and ligand molecular conformation of the current evaluation record as the corresponding current protein pocket conformation and current label ligand molecular conformation; and performing molecular fragment object identification on the current label ligand molecular conformation to obtain one or more corresponding molecular fragment objects, and taking the molecular fragment object with the shortest length as the corresponding current molecular fragment; and copying the current label ligand molecular conformation once to obtain the corresponding current copied molecular conformation; and deleting all atomic objects corresponding to the current molecular fragment in the current copied molecular conformation; and taking the deleted current copied molecular conformation as the corresponding current ligand context information; and performing corresponding fragment growth processing according to the current protein pocket conformation and the current ligand context information by the current model to obtain the corresponding current predicted molecular conformation; and forming a corresponding third prediction-label pair by the current predicted molecular conformation and the corresponding current label ligand molecular conformation; and bringing all the obtained third prediction-label pairs into the preset third evaluation function for calculation to obtain the corresponding third evaluation value;

[0145] Here, the third evaluation function includes at least conventional MAE, RMSE and other functions;

[0146] Step D9, identifying whether the third evaluation value meets the preset third evaluation value range; if the third evaluation value does not meet the third evaluation value range, returning to step D1; if the third evaluation value meets the third evaluation value range, confirming that the current model training is completed.

[0147] It should also be noted that the training method according to the side chain modification task mentioned in the above step A7 refers to training the Pocket2Mol model 51, GraphBP model 52 and FLAG model 53 with the goal of the model being able to realize the side chain modification task function; the task training module 2 adopts a supervised method during training. The task training module 2 is specifically used for training the Pocket2Mol model 51, GraphBP model 52 or FLAG model 53 according to the first data set in the training method of the side chain modification task:

[0148] Step E1, record the currently trained Pocket2Mol model 51, GraphBP model 52 or FLAG model 53 as the corresponding current model;

[0149] Step E2, based on a preset fourth split ratio (for example, 8:2), the first data set is divided into two sub-data sets recorded as a corresponding fourth training set and a fourth evaluation set;

[0150] Wherein, both the fourth training set and the fourth evaluation set are composed of a plurality of composite records, and a ratio of the total number of records in the fourth training set to the total number of records in the fourth evaluation set satisfies a fourth segmentation ratio;

[0151] Step E3, taking the first composite record of the fourth training set as the corresponding current training record;

[0152] Step E4, taking the protein pocket conformation and ligand molecular conformation of the current training record as the corresponding current protein pocket conformation and current label ligand molecular conformation; and performing side chain object identification on the current label ligand molecular conformation to obtain one or more corresponding side chain objects; and performing a copy of the current label ligand molecular conformation to obtain the corresponding current copied molecular conformation; and selecting any one of the multiple atomic objects corresponding to each side chain object in the current copied molecular conformation as the corresponding perturbation atom, and performing mask processing on the atomic type of each perturbation atom and performing Gaussian noise addition processing on the atomic coordinates; and taking the current copied molecular conformation with the atomic type mask and atomic coordinate noise added as the corresponding current ligand context information;

[0153] Step E5, the current model performs corresponding side chain modification processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding current predicted molecular conformation;

[0154] Step E6, bringing the current predicted molecular conformation and the current label ligand molecular conformation into the model loss function corresponding to the current model to calculate and obtain the corresponding fourth loss value;

[0155] Step E7, identifying whether the fourth loss value satisfies a preset fourth loss value range; if the fourth loss value satisfies the fourth loss value range, identifying whether the current training record is the last composite record of the fourth training set, if so, turning to step E8, if not, extracting the next composite record of the fourth training set as a new current training record and returning to step E4; if the fourth loss value does not satisfy the fourth loss value range, based on a preset model parameter optimization algorithm or optimizer (such as SGD optimization algorithm / optimizer, ADAM optimization algorithm / optimizer, etc.), a round of modulation is performed on the model parameters of the current model in a direction in which the current model loss function reaches a minimum value, and returning to step E5 at the end of this round of modulation;

[0156] Step E8, taking each complex record of the fourth evaluation set as the corresponding current evaluation record in turn; taking the protein pocket conformation and ligand molecular conformation of the current evaluation record as the corresponding current protein pocket conformation and current label ligand molecular conformation; taking the protein pocket conformation and ligand molecular conformation of the current training record as the corresponding current protein pocket conformation and current label ligand molecular conformation; performing side chain object identification on the current label ligand molecular conformation to obtain one or more corresponding side chain objects; performing a copy of the current label ligand molecular conformation to obtain the corresponding current copied molecular conformation; and performing a plurality of atom pairs corresponding to each side chain object in the current copied molecular conformation. Select any one of the images as the corresponding perturbation atom, and perform mask processing on the atomic type of each perturbation atom and perform Gaussian noise addition processing on the atomic coordinates; and use the current copied molecular conformation with the atomic type mask and atomic coordinate noise added as the corresponding current ligand context information; and the current model performs corresponding side chain modification processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding current predicted molecular conformation; and the current predicted molecular conformation and the corresponding current label ligand molecular conformation form a corresponding fourth prediction-label pair; and bring all the fourth prediction-label pairs obtained into the preset fourth evaluation function to calculate and obtain the corresponding fourth evaluation value;

[0157] Here, the fourth evaluation function includes at least conventional MAE, RMSE and other functions;

[0158] Step E9, identifying whether the fourth evaluation value meets the preset fourth evaluation value range; if the fourth evaluation value does not meet the fourth evaluation value range, returning to step E1; if the fourth evaluation value meets the fourth evaluation value range, confirming that the current model training is completed.

[0159] It should also be noted that the training method according to the skeleton transition task mentioned in the above step A8 refers to training the Pocket2Mol model 51, GraphBP model 52 and FLAG model 53 with the goal of the model being able to realize the skeleton transition task function; the task training module 2 adopts a supervised method during training. The task training module 2 is specifically used for training the Pocket2Mol model 51, GraphBP model 52 or FLAG model 53 according to the first data set in the training method of the skeleton transition task:

[0160] Step F1, record the currently trained Pocket2Mol model 51, GraphBP model 52 or FLAG model 53 as the corresponding current model;

[0161] Step F2, based on a preset fifth split ratio (for example, 8:2), the first data set is divided into two sub-data sets recorded as a corresponding fifth training set and a fifth evaluation set;

[0162] Wherein, both the fifth training set and the fifth evaluation set consist of a plurality of composite records, and a ratio of the total number of records in the fifth training set to the total number of records in the fifth evaluation set satisfies a fifth split ratio;

[0163] Step F3, taking the first composite record of the fifth training set as the corresponding current training record;

[0164] Step F4, taking the protein pocket conformation and ligand molecular conformation of the current training record as the corresponding current protein pocket conformation and current label ligand molecular conformation; performing skeleton fragment recognition on the current label ligand molecular conformation to obtain one or more corresponding skeleton fragments, and selecting one of them as the corresponding current skeleton fragment; performing a copy of the current label ligand molecular conformation to obtain the corresponding current copied molecular conformation; deleting all atomic objects corresponding to the current skeleton fragment in the current copied molecular conformation; and taking the deleted current copied molecular conformation as the corresponding current ligand context information;

[0165] Step F5, the current model performs corresponding backbone fragment transition processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding current predicted molecular conformation;

[0166] Step F6, bringing the current predicted molecular conformation and the current label ligand molecular conformation into the model loss function corresponding to the current model to calculate and obtain the corresponding fifth loss value;

[0167] Step F7, identifying whether the fifth loss value satisfies the preset fifth loss value range; if the fifth loss value satisfies the fifth loss value range, identifying whether the current training record is the last composite record of the fifth training set, if so, turning to step F8, if not, extracting the next composite record of the fifth training set as the new current training record and returning to step F4; if the fifth loss value does not satisfy the fifth loss value range, based on a preset model parameter optimization algorithm or optimizer (such as SGD optimization algorithm / optimizer, ADAM optimization algorithm / optimizer, etc.), a round of modulation is performed on the model parameters of the current model in the direction of minimizing the loss function of the current model, and returning to step F5 at the end of this round of modulation;

[0168] Step F8, taking each complex record of the fifth evaluation set as the corresponding current evaluation record in turn; and taking the protein pocket conformation and ligand molecular conformation of the current evaluation record as the corresponding current protein pocket conformation and current label ligand molecular conformation; and performing backbone fragment recognition on the current label ligand molecular conformation to obtain one or more corresponding backbone fragments, and selecting any one of them as the corresponding current backbone fragment; and performing a copy of the current label ligand molecular conformation to obtain the corresponding current copied molecular conformation; and deleting all atomic objects corresponding to the current backbone fragment in the current copied molecular conformation; and taking the deleted current copied molecular conformation as the corresponding current ligand context information; and performing corresponding backbone fragment transition processing according to the current protein pocket conformation and the current ligand context information by the current model to obtain the corresponding current predicted molecular conformation; and forming a corresponding fifth prediction-label pair by the current predicted molecular conformation and the corresponding current label ligand molecular conformation; and bringing all the obtained fifth prediction-label pairs into the preset fifth evaluation function for calculation to obtain the corresponding fifth evaluation value;

[0169] Here, the fifth evaluation function includes at least conventional MAE, RMSE and other functions;

[0170] Step F9, identifying whether the fifth evaluation value meets the preset fifth evaluation value range; if the fifth evaluation value does not meet the fifth evaluation value range, returning to step F1; if the fifth evaluation value meets the fifth evaluation value range, confirming that the current model training is completed.

[0171] (III) Task execution module 3:

[0172] The task execution module 3 is used to perform molecular generation task execution processing according to the second reference data and the task interface library 6 to obtain the corresponding first generated molecular conformation and send it to the task scheduling module 1.

[0173] In another specific implementation of the embodiment of the present invention, the task execution module 3 is specifically used to perform the molecular generation task execution processing according to the second reference data and the task interface library 6 to obtain the corresponding first generated molecular conformation and send it to the task scheduling module 1:

[0174] Step G1, identifying the first task type in the second reference data;

[0175] Step G2: if the first task type is a ligand generation task, the second reference data is sent to the ligand generation task interface 61; and the first generated molecular conformation sent back by the ligand generation task interface 61 is sent to the task scheduling module 1;

[0176] Step G2, if the first task type is a linker generation task, sending the second reference data to the linker generation task interface 62; and sending the first generated molecular conformation sent back by the linker generation task interface 62 to the task scheduling module 1;

[0177] Step G3, if the first task type is a fragment growth task, the second reference data is sent to the fragment growth task interface 63; and the first generated molecular conformation sent back by the fragment growth task interface 63 is sent to the task scheduling module 1;

[0178] Step G4, if the first task type is a side chain modification task, sending the second reference data to the side chain modification task interface 64; and sending the first generated molecular conformation sent back by the side chain modification task interface 64 to the task scheduling module 1;

[0179] Step G5, if the first task type is a skeleton transition task, the second reference data is sent to the skeleton transition task interface 65; and the first generated molecular conformation sent back by the skeleton transition task interface 65 is sent to the task scheduling module 1.

[0180] (IV) Compound database 4:

[0181] The complex database 4 of the embodiment of the present invention includes multiple complex data sets 41; each complex data set 41 corresponds to a data set index; each complex data set 41 is composed of multiple complex records; each complex record corresponds to a protein-ligand complex, which is composed of a protein pocket and a ligand molecule; the complex record includes a complex conformation, a protein pocket conformation and a ligand molecule conformation; the complex conformation, the protein pocket conformation and the ligand molecule conformation are respectively the three-dimensional molecular conformations of the protein-ligand complex, the protein pocket and the ligand molecule corresponding to the current record, and the protein pocket conformation and the ligand molecule conformation are each a sub-conformation of the complex conformation.

[0182] (V) Autoregressive Model Library 5:

[0183] The autoregressive model library 5 of the embodiment of the present invention includes a Pocket2Mol model 51 , a GraphBP model 52 and a FLAG model 53 .

[0184] The model implementation principle and molecular conformation prediction mechanism of the Pocket2Mol model 51 have been specifically described in the technical document A "Pocket2Mol: Efficient Molecular Sampling Based on 3D Protein Pockets", so we will not go into further detail here. In short, after receiving the protein pocket P and ligand context information C as model input, the Pocket2Mol model 51 first encodes the protein pocket P + ligand context information C based on the E(3) equivariant neural network to obtain the corresponding encoding tensor; then, based on a frontier atom predictor, the newly generated frontier atoms of the ligand molecule are predicted according to the encoding tensor to obtain the corresponding frontier atom set; then, the frontier atom set is cyclically sampled until all atoms in the frontier atom set are confirmed to have been sampled once, and the cycle is terminated, and the atom sampled at each step is used as the current focus atom, and a position predictor is used to predict the atomic coordinates of the current focus atom, and a type-bond predictor is used to predict the atom type and associated chemical bonds of the current focus atom; and after the cyclic sampling is completed, the final ligand molecular conformation is output.

[0185] The model implementation principle and molecular conformation prediction mechanism of GraphBP model 52 have been specifically described in the technical document B "Generating 3D Molecules for Target Protein Binding", so we will not go into further detail here. In short, after receiving the protein pocket P and ligand context information C as model input, GraphBP model 52 1) performs conformational merging of the protein pocket P and ligand context information C to obtain new context information C ’ ; 2) The latest context information C ’ The ligand molecular conformation in the protein pocket P is identified as empty. If it is empty, the focus atom is predicted based on an atom classifier according to the encoding features of the protein pocket P. If it is not empty, the focus atom is predicted based on another atom classifier according to the encoding features of the latest ligand molecular conformation. 4) A local coordinate system is formed based on the current focus atom and the other two existing atoms adjacent to it (which may belong to the protein pocket or the ligand molecular conformation), and the local coordinate system and the latest context information C are used to predict the focus atom. ’ The atomic type of the next atom is predicted based on the encoded features; 5) If the predicted result is a normal atom type, the atomic coordinates of the atom are predicted based on the currently predicted atom type, and a new object consisting of the current atom type + coordinates is added to the context information C ’and return to the above step 2; 6) If the prediction result is an invalid atom type, the loop processing is stopped and a new ligand molecule atom object set is obtained. It should be noted that the GraphBP model mentioned in technical document B only predicts the atom object set in the conformation, and does not predict the chemical bond object set; because all conformations of the embodiment of the present invention include chemical bond object sets, the GraphBP model 52 of the embodiment of the present invention will also call the preset chemical calculation tool (such as RDkits, OpenBabel, etc.) to perform the corresponding chemical bond object set analysis and calculation according to the ligand molecule atom object set to obtain the corresponding ligand molecule chemical bond object set after obtaining the new ligand molecule atom object set, and the obtained ligand molecule atom object set + ligand molecule chemical bond object set constitute the final ligand molecule conformation and output it.

[0186] The model implementation principle and molecular conformation prediction mechanism of FLAG model 53 have been specifically described in technical document C "MOLECULEGENERATION FOR TARGET PROTEIN BINDING WITH STRUCTURAL MOTIFS", so they will not be further described here. In short, the principle of FLAG model 53 is also to perform cyclic processing according to the logic of adding one by one, but the difference from the principle of adding one by one of GraphBP model 52 is that FLAG model 53 uses the method of adding components one by one (functional groups, linkers and other basic fragment components) to generate molecular conformations. In addition, the ligand molecular conformations obtained in each step of FLAG model 53 carry chemical bond information, so after the last component (functional groups, linkers and other basic fragment components) is added, the latest ligand molecular conformation can be directly output as the processing result without calling chemical calculation tools for chemical bond analysis and calculation.

[0187] (VI) Task interface library 6:

[0188] The task interface library 6 includes five types of molecule generation task interfaces, namely, a ligand generation task interface 61, a linker generation task interface 62, a fragment growth task interface 63, a side chain modification task interface 64 and a backbone transition task interface 65.

[0189] The ligand generation task interface 61 is used to extract the corresponding first model type and the first protein pocket conformation from the received second reference data; and use the first protein pocket conformation as the corresponding current protein pocket conformation; and set the corresponding current ligand context information to empty; and identify the first model type; if the first model type is the first type, the Pocket2Mol model 51 is called to perform the corresponding ligand molecule generation process according to the current protein pocket conformation and the current ligand context information to obtain the corresponding first ligand molecule conformation; if the first model type is the second type, the GraphBP model 52 is called to perform the corresponding ligand molecule generation process according to the current protein pocket conformation and the current ligand context information to obtain the corresponding first ligand molecule conformation; if the first model type is the third type, the FLAG model 53 is called to perform the corresponding ligand molecule generation process according to the current protein pocket conformation and the current ligand context information to obtain the corresponding first ligand molecule conformation; and the obtained first ligand molecule conformation is sent to the task execution module 3 as the corresponding first generated molecule conformation.

[0190] The connector generation task interface 62 is used to extract the corresponding first model type, the first protein-ligand complex conformation, the first pocket atom set and the first ligand atom set from the received second reference data; and based on the first pocket atom set, perform a corresponding protein pocket conformation extraction on the first protein-ligand complex conformation to obtain the corresponding current protein pocket conformation; and based on the first ligand atom set, perform a corresponding ligand molecule conformation extraction on the first protein-ligand complex conformation and use the extraction result as the corresponding current ligand context information; and identify the first model type; if the first model type is the first type, call the Pocket2Mol model 51 according to the current The protein pocket conformation and the current ligand context information are processed for corresponding linker generation to obtain the corresponding second ligand molecular conformation; if the first model type is the second type, the GraphBP model 52 is called to perform corresponding linker generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding second ligand molecular conformation; if the first model type is the third type, the FLAG model 53 is called to perform corresponding linker generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding second ligand molecular conformation; and the obtained second ligand molecular conformation is sent to the task execution module 3 as the corresponding first generated molecular conformation.

[0191] The fragment growth task interface 63 is used to extract the corresponding first model type, the second protein-ligand complex conformation, the second pocket atom set, the second ligand atom set and the first fragment atom set to be grown from the received second reference data; and based on the second pocket atom set, extract the corresponding protein pocket conformation of the second protein-ligand complex conformation to obtain the corresponding current protein pocket conformation; and based on the second ligand atom set, extract the corresponding ligand molecule conformation of the second protein-ligand complex conformation to obtain the corresponding current ligand molecule conformation; and based on the first fragment atom set to be grown, delete the fragment conformation corresponding to the fragment to be grown in the current ligand molecule conformation; and use the current ligand molecule conformation after the fragment deletion as the corresponding current ligand context information; and extract the corresponding ligand molecule conformation of the first model type. for identification; if the first model type is the first type, the Pocket2Mol model 51 is called to perform corresponding fragment growth processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding third ligand molecular conformation; if the first model type is the second type, the GraphBP model 52 is called to perform corresponding fragment growth processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding third ligand molecular conformation; if the first model type is the third type, the FLAG model 53 is called to perform corresponding fragment growth processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding third ligand molecular conformation; and the obtained third ligand molecular conformation is sent to the task execution module 3 as the corresponding first generated molecular conformation.

[0192] The side chain modification task interface 64 is used to extract the corresponding first model type, the third protein-ligand complex conformation, the third pocket atom set and the third ligand atom set from the received second reference data; and based on the third pocket atom set, perform a corresponding protein pocket conformation extraction on the third protein-ligand complex conformation to obtain the corresponding current protein pocket conformation; and based on the third ligand atom set, perform a corresponding ligand molecule conformation extraction on the third protein-ligand complex conformation and use the extraction result as the corresponding current ligand context information; and identify the first model type; if the first model type is the first type, call the Pocket2Mol model 51 according to the current The previous protein pocket conformation and the current ligand context information are subjected to corresponding side chain modification processing to obtain the corresponding fourth ligand molecular conformation; if the first model type is the second type, the GraphBP model 52 is called to perform corresponding side chain modification processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding fourth ligand molecular conformation; if the first model type is the third type, the FLAG model 53 is called to perform corresponding side chain modification processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding fourth ligand molecular conformation; and the obtained fourth ligand molecular conformation is sent to the task execution module 3 as the corresponding first generated molecular conformation.

[0193] The skeleton transition task interface 65 is used to extract the corresponding first model type, the fourth protein-ligand complex conformation, the fourth pocket atom set, the fourth ligand atom set and the first to-be-transitioned fragment atom set from the received second reference data; and based on the fourth pocket atom set, perform a corresponding protein pocket conformation extraction on the fourth protein-ligand complex conformation to obtain the corresponding current protein pocket conformation; and based on the fourth ligand atom set, perform a corresponding ligand molecule conformation extraction on the fourth protein-ligand complex conformation to obtain the corresponding current ligand molecule conformation; and based on the first to-be-transitioned fragment atom set, delete the fragment conformation corresponding to the to-be-transitioned fragment in the current ligand molecule conformation; and use the current ligand molecule conformation after the fragment deletion as the corresponding current ligand context information; and identify the first model type. If the first model type is the first type, the Pocket2Mol model 51 is called to perform the corresponding backbone fragment transition processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding fifth ligand molecular conformation; if the first model type is the second type, the GraphBP model 52 is called to perform the corresponding backbone fragment transition processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding fifth ligand molecular conformation; if the first model type is the third type, the FLAG model 53 is called to perform the corresponding backbone fragment transition processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding fifth ligand molecular conformation; and the obtained fifth ligand molecular conformation is sent to the task execution module 3 as the corresponding first generated molecular conformation.

[0194] It should be noted that it should be understood that the division of the various modules of the above system is only a division of logical functions. In actual implementation, they can be fully or partially integrated into one physical entity, or they can be physically separated. And these modules can all be implemented in the form of software called by processing elements; they can also be all implemented in the form of hardware; some modules can also be implemented in the form of software called by processing elements, and some modules can be implemented in the form of hardware. For example, the task scheduling module can be a separately established processing element, or it can be integrated in a chip of the above-mentioned device. In addition, it can also be stored in the memory of the above-mentioned device in the form of program code, and called by a processing element of the above-mentioned system and execute the functions of the above-mentioned determined modules. The implementation of other modules is similar. In addition, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element described here can be an integrated circuit with signal processing capabilities. In the implementation process, each method step of the aforementioned method or each module processing step of the aforementioned system can be completed by an integrated logic circuit of hardware in the processor element or instructions in the form of software.

[0195] For example, the modules of the above system may be one or more integrated circuits configured to implement the aforementioned method, such as one or more application specific integrated circuits (ASIC), or one or more digital signal processors (DSP), or one or more field programmable gate arrays (FPGA). For another example, when a module of the above system is implemented in the form of a processing element scheduling program code, the processing element may be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call program code. For another example, these modules may be integrated together and implemented in the form of a system-on-a-chip (SOC).

[0196] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments may be implemented 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, all or part of the processes or functions described in the above method embodiments are generated. The above computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The above-mentioned computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the above-mentioned computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, Bluetooth, microwave, etc.) methods. The above-mentioned computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media integrated. The above-mentioned available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

[0197] The embodiment of the present invention provides a system for processing molecular generation tasks based on an autoregressive model, including: a task scheduling module, a task training module, a task execution module, a complex database, an autoregressive model library and a task interface library; wherein the task scheduling module is used to prepare model training reference data according to the task data after receiving the molecular generation task data input from the external interface, obtain the corresponding reference data and send it to the task training module, and then prepare the molecular generation task reference data according to the task data when the task training module confirms that the model training is finished, obtain the corresponding reference data and send it to the task execution module, and finally feedback the generated molecular conformation sent back by the task execution module to the external interface; the task training module is used to train the Pocket2Mol model, GraphBP model and FLAG model of the autoregressive model library according to the reference data and the complex database; the task execution module is used to perform molecular generation task execution processing according to the reference data and five types of task interfaces (ligand generation task interface, linker generation task interface, fragment growth task interface, side chain modification task interface and skeleton transition task interface) of the task interface library. The embodiment of the present invention can improve the conventional SBDD design process, shorten the design cycle and improve the design efficiency.

[0198] The professionals should further realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to the function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0199] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0200] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A system for handling molecular generation tasks based on autoregressive models, It is characterized in that The system includes: a task scheduling module, a task training module, a task execution module, a compound database, an autoregressive model library and a task interface library; The task scheduling module is connected to the task training module and the task execution module respectively; the task training module is connected to the compound database and the autoregressive model library respectively; the task execution module is connected to the task interface library; the task interface library is connected to the autoregressive model library; The task scheduling module is used to receive the molecular generation task data input by the external interface as the corresponding first task data; and prepare the model training reference data according to the first task data to obtain the corresponding first reference data and send it to the task training module; and when the first training state sent by the task training module is received as the end of model training, prepare the molecular generation task reference data according to the first task data to obtain the corresponding second reference data and send it to the task execution module; and feed back the first generated molecular conformation sent by the task execution module to the external interface; the first generated molecular conformation is a three-dimensional molecular conformation of a generated molecule; the three-dimensional molecular conformation is composed of a corresponding atomic object set and a chemical bond object set; the atomic object set is composed of one or more atomic objects; each of the atomic objects includes an atomic index, an atomic type and an atomic coordinate; the chemical bond object set is composed of one or more chemical bond objects; each of the chemical bond objects includes a chemical bond index, a chemical bond type and a chemical bond atom index group; the chemical bond atom index group is composed of the atomic indexes of the two atomic objects corresponding to the current chemical bond object; The task training module is used to train the specified model in the autoregressive model library according to the first reference data and the composite database to obtain the corresponding first training state and send it to the task scheduling module; The task execution module is used for performing molecular generation task execution processing according to the second reference data and the task interface library to obtain the corresponding first generated molecular conformation and send it to the task scheduling module; The complex database includes a plurality of complex data sets; each of the complex data sets corresponds to a data set index; each of the complex data sets consists of a plurality of complex records; each of the complex records corresponds to a protein-ligand complex, which is formed by the combination of a protein pocket and a ligand molecule; the complex record includes a complex conformation, a protein pocket conformation and a ligand molecule conformation; the complex conformation, the protein pocket conformation and the ligand molecule conformation are respectively the three-dimensional molecular conformations of the protein-ligand complex, the protein pocket and the ligand molecule corresponding to the current record, and the protein pocket conformation and the ligand molecule conformation are each a sub-conformation of the complex conformation; The autoregressive model library includes Pocket2Mol model, GraphBP model and FLAG model; The task interface library includes five types of molecular generation task interfaces, namely, ligand generation task interface, linker generation task interface, fragment growth task interface, side chain modification task interface and backbone transition task interface.

2. The system for processing molecular generation tasks based on an autoregressive model according to claim 1, It is characterized in that The first task data includes a first task type, a first model type, a first data set index and a first task information set; The first task type includes ligand generation tasks, linker generation tasks, fragment growth tasks, side chain modification tasks and backbone transition tasks; The first model type includes a first type, a second type and a third type; When the first task type is a ligand generation task, the corresponding first task information set consists of a first protein pocket conformation; the first protein pocket conformation is the three-dimensional molecular conformation of a protein pocket; When the first task type is a connector generation task, the corresponding first task information set includes a first protein-ligand complex conformation, a first pocket atom set and a first ligand atom set; the first protein-ligand complex conformation is the three-dimensional molecular conformation of a protein-ligand complex A; the protein-ligand complex A consists of a protein pocket P A and a ligand molecule M A Composition, the ligand molecule M A Composed of at least two or more molecular fragments F A The first pocket atom set and the first ligand atom set are respectively the protein pocket P A and the ligand molecule M A The set of atomic indices of ; When the first task type is a fragment growth task, the corresponding first task information set includes a second protein-ligand complex conformation, a second pocket atom set, a second ligand atom set and a first fragment atom set to be grown; the second protein-ligand complex conformation is the three-dimensional molecular conformation of a protein-ligand complex B; the protein-ligand complex B consists of a protein pocket P B and a ligand molecule M B Composition, the ligand molecule M B The molecular fragment that needs to be grown is recorded as the fragment to be grown F B The second pocket atom set, the second ligand atom set and the first to-be-grown fragment atom set are respectively the protein pocket P B , the ligand molecule M B and the fragment to be grown F B The set of atomic indices of ; When the first task type is a side chain modification task, the corresponding first task information set includes a third protein-ligand complex conformation, a third pocket atom set and a third ligand atom set; the third protein-ligand complex conformation is the three-dimensional molecular conformation of a protein-ligand complex C; the protein-ligand complex C consists of a protein pocket P C and a ligand molecule M C The third pocket atom set and the third ligand atom set are respectively the protein pocket P C and the ligand molecule M C The set of atomic indices of ; When the first task type is a skeleton transition task, the corresponding first task information set includes a fourth protein-ligand complex conformation, a fourth pocket atom set, a fourth ligand atom set and a first to-be-transitioned fragment atom set; the fourth protein-ligand complex conformation is the three-dimensional molecular conformation of a protein-ligand complex D; the protein-ligand complex D consists of a protein pocket P D and a ligand molecule M D Composition, the ligand molecule M D The skeleton fragment that needs to be transitioned on the main skeleton is recorded as the fragment to be transitioned F D The fourth pocket atom set, the fourth ligand atom set and the first to-be-transitioned fragment atom set are respectively the protein pocket P D , the ligand molecule M D and the fragment to be transitioned F D The set of atomic indices.

3. The system for processing molecular generation tasks based on an autoregressive model according to claim 2, It is characterized in that The task scheduling module is specifically used for sending the first reference data corresponding to the first task type, the first model type and the first data set index of the first task data to the task training module when the first reference data corresponding to the first reference data is prepared according to the first task data and sent to the task training module; The task scheduling module is specifically used to send the corresponding second reference data composed of the first task type, the first model type and the first task information set of the first task data to the task training module when the corresponding second reference data is prepared by performing molecular generation task reference data according to the first task data and sent to the task execution module.

4. The system for processing molecular generation tasks based on an autoregressive model according to claim 3, It is characterized in that The task training module is specifically used for, when the first training state is obtained by training the designated model in the autoregressive model library according to the first reference data and the composite database and then sending the first training state to the task scheduling module, Extracting the corresponding first task type, first model type and first data set index from the first reference data; and taking the composite data set whose data set index in the composite database matches the first data set index as the corresponding first data set; and identifying the first task type; When the first task type is a ligand generation task, the first model type is identified; if the first model type is the first type, the Pocket2Mol model is trained according to the first data set in a training manner for the ligand generation task; if the first model type is the second type, the GraphBP model is trained according to the first data set in a training manner for the ligand generation task; if the first model type is the third type, the FLAG model is trained according to the first data set in a training manner for the ligand generation task; The Pocket2Mol model, the GraphBP model and the FLAG model trained in the training method of the ligand generation task are all used to perform corresponding ligand molecule generation processing according to the input protein pocket conformation and ligand context information and output the three-dimensional molecular conformation of a ligand molecule as the corresponding ligand molecular conformation; When the first task type is a connector generation task, the first model type is identified; if the first model type is the first type, the Pocket2Mol model is trained according to the first data set in a training manner of the connector generation task; if the first model type is the second type, the GraphBP model is trained according to the first data set in a training manner of the connector generation task; if the first model type is the third type, the FLAG model is trained according to the first data set in a training manner of the connector generation task; the Pocket2Mol model, the GraphBP model and the FLAG model trained according to the training manner of the connector generation task are all used to perform corresponding connector generation processing according to the input protein pocket conformation and ligand context information and output the three-dimensional molecular conformation of a new ligand molecule as the corresponding ligand molecular conformation; When the first task type is a fragment growth task, the first model type is identified; if the first model type is the first type, the Pocket2Mol model is trained according to the first data set in a training manner of the fragment growth task; if the first model type is the second type, the GraphBP model is trained according to the first data set in a training manner of the fragment growth task; if the first model type is the third type, the FLAG model is trained according to the first data set in a training manner of the fragment growth task; the Pocket2Mol model, the GraphBP model and the FLAG model trained in the training manner of the fragment growth task are all used to perform corresponding fragment growth processing according to the input protein pocket conformation and ligand context information and output the three-dimensional molecular conformation of a new ligand molecule as the corresponding ligand molecular conformation; When the first task type is a side chain modification task, the first model type is identified; if the first model type is the first type, the Pocket2Mol model is trained according to the first data set in a training manner for the side chain modification task; if the first model type is the second type, the GraphBP model is trained according to the first data set in a training manner for the side chain modification task; if the first model type is the third type, the FLAG model is trained according to the first data set in a training manner for the side chain modification task; the Pocket2Mol model, the GraphBP model and the FLAG model trained according to the training manner for the side chain modification task are all used to perform corresponding side chain modification processing according to the input protein pocket conformation and ligand context information and output the three-dimensional molecular conformation of a new ligand molecule as the corresponding ligand molecular conformation; When the first task type is a skeleton transition task, the first model type is identified; if the first model type is the first type, the Pocket2Mol model is trained according to the first data set in a training manner of the skeleton transition task; if the first model type is the second type, the GraphBP model is trained according to the first data set in a training manner of the skeleton transition task; if the first model type is the third type, the FLAG model is trained according to the first data set in a training manner of the skeleton transition task; the Pocket2Mol model, the GraphBP model and the FLAG model trained in the training manner of the skeleton transition task are all used to perform corresponding skeleton fragment transition processing according to the input protein pocket conformation and ligand context information and output the three-dimensional molecular conformation of a new ligand molecule as the corresponding ligand molecular conformation; And when the current model training is finished, the first training state specifically set as the end of the model training is sent to the task scheduling module.

5. The system for processing molecular generation tasks based on an autoregressive model according to claim 3, It is characterized in that The task execution module is specifically used for, when the molecular generation task execution processing is performed according to the second reference data and the task interface library to obtain the corresponding first generated molecular conformation and send it to the task scheduling module, identifying the first task type in the second reference data; If the first task type is a ligand generation task, sending the second reference data to the ligand generation task interface; and sending the first generated molecular conformation sent back by the ligand generation task interface to the task scheduling module; If the first task type is a linker generation task, sending the second reference data to the linker generation task interface; and sending the first generated molecular conformation sent back by the linker generation task interface to the task scheduling module; If the first task type is a fragment growth task, sending the second reference data to the fragment growth task interface; and sending the first generated molecular conformation sent back by the fragment growth task interface to the task scheduling module; If the first task type is a side chain modification task, sending the second reference data to the side chain modification task interface; and sending the first generated molecular conformation sent back by the side chain modification task interface to the task scheduling module; If the first task type is a skeleton transition task, the second reference data is sent to the skeleton transition task interface; and the first generated molecular conformation sent back by the skeleton transition task interface is sent to the task scheduling module.

6. The system for processing molecular generation tasks based on an autoregressive model according to claim 5, It is characterized in that The ligand generation task interface is used to extract the corresponding first model type and the first protein pocket conformation from the received second reference data; and taking the first protein pocket conformation as the corresponding current protein pocket conformation; and setting the corresponding current ligand context information to be empty; and identifying the first model type; if the first model type is the first type, calling the Pocket2Mol model to perform corresponding ligand molecule generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding first ligand molecule conformation; If the first model type is the second type, calling the GraphBP model to perform corresponding ligand molecule generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding first ligand molecule conformation; If the first model type is the third type, calling the FLAG model to perform corresponding ligand molecule generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding first ligand molecule conformation; And the obtained first ligand molecular conformation is sent to the task execution module as the corresponding first generated molecular conformation.

7. The system for processing molecular generation tasks based on an autoregressive model according to claim 5, It is characterized in that The connector generation task interface is used to extract the corresponding first model type, the first protein-ligand complex conformation, the first pocket atom set and the first ligand atom set from the received second reference data; and extracting the corresponding protein pocket conformation of the first protein-ligand complex conformation based on the first pocket atom set to obtain the corresponding current protein pocket conformation; and extracting the corresponding ligand molecular conformation of the first protein-ligand complex conformation based on the first ligand atom set and using the extraction result as the corresponding current ligand context information; and identifying the first model type; if the first model type is the first type, calling the Pocket2Mol model to perform corresponding linker generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding second ligand molecular conformation; If the first model type is the second type, calling the GraphBP model to perform corresponding linker generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding second ligand molecular conformation; If the first model type is the third type, calling the FLAG model to perform corresponding linker generation processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding second ligand molecular conformation; And the obtained second ligand molecular conformation is sent to the task execution module as the corresponding first generated molecular conformation.

8. The system for processing molecular generation tasks based on an autoregressive model according to claim 5, It is characterized in that The fragment growth task interface is used to extract the corresponding first model type, the second protein-ligand complex conformation, the second pocket atom set, the second ligand atom set and the first to-be-grown fragment atom set from the received second reference data; and extracting the corresponding protein pocket conformation of the second protein-ligand complex conformation based on the second pocket atom set to obtain the corresponding current protein pocket conformation; and extracting the corresponding ligand molecular conformation of the second protein-ligand complex conformation based on the second ligand atom set to obtain the corresponding current ligand molecular conformation; Based on the first set of atoms of the fragment to be grown, the fragment conformation corresponding to the fragment to be grown in the current ligand molecular conformation is deleted; and the current ligand molecular conformation after the fragment deletion is completed is used as the corresponding current ligand context information; and identifying the first model type; if the first model type is the first type, calling the Pocket2Mol model to perform corresponding fragment growth processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding third ligand molecular conformation; if the first model type is the second type, calling the GraphBP model to perform corresponding fragment growth processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding third ligand molecular conformation; if the first model type is the third type, calling the FLAG model to perform corresponding fragment growth processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding third ligand molecular conformation; The obtained third ligand molecular conformation is sent to the task execution module as the corresponding first generated molecular conformation.

9. The system for processing molecular generation tasks based on an autoregressive model according to claim 5, It is characterized in that The side chain modification task interface is used to extract the corresponding first model type, the third protein-ligand complex conformation, the third pocket atom set and the third ligand atom set from the received second reference data; and extracting the corresponding protein pocket conformation of the third protein-ligand complex conformation based on the third pocket atom set to obtain the corresponding current protein pocket conformation; and extracting the corresponding ligand molecular conformation of the third protein-ligand complex conformation based on the third ligand atom set and using the extraction result as the corresponding current ligand context information; and identifying the first model type; if the first model type is the first type, calling the Pocket2Mol model to perform corresponding side chain modification processing according to the current protein pocket conformation and the current ligand context information to obtain a corresponding fourth ligand molecular conformation; If the first model type is the second type, calling the GraphBP model to perform corresponding side chain modification processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding fourth ligand molecular conformation; If the first model type is the third type, calling the FLAG model to perform corresponding side chain modification processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding fourth ligand molecular conformation; And the obtained fourth ligand molecular conformation is sent to the task execution module as the corresponding first generated molecular conformation.

10. The system for processing molecular generation tasks based on autoregressive model according to claim 5, It is characterized in that The backbone transition task interface is used to extract the corresponding first model type, the fourth protein-ligand complex conformation, the fourth pocket atom set, the fourth ligand atom set and the first to-be-transitioned fragment atom set from the received second reference data; and extracting the corresponding protein pocket conformation of the fourth protein-ligand complex conformation based on the fourth pocket atom set to obtain the corresponding current protein pocket conformation; and extracting the corresponding ligand molecular conformation of the fourth protein-ligand complex conformation based on the fourth ligand atom set to obtain the corresponding current ligand molecular conformation; Based on the first set of atoms of the fragments to be transitioned, the fragment conformation corresponding to the fragment to be transitioned in the current ligand molecular conformation is deleted; and the current ligand molecular conformation after the fragment deletion is completed is used as the corresponding current ligand context information; and identifying the first model type; if the first model type is the first type, calling the Pocket2Mol model to perform corresponding backbone fragment transition processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding fifth ligand molecular conformation; If the first model type is the second type, calling the GraphBP model to perform corresponding backbone fragment transition processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding fifth ligand molecular conformation; If the first model type is the third type, calling the FLAG model to perform corresponding backbone fragment transition processing according to the current protein pocket conformation and the current ligand context information to obtain the corresponding fifth ligand molecular conformation; And the obtained fifth ligand molecular conformation is sent to the task execution module as the corresponding first generated molecular conformation.