CN116052762A - Method and server for matching drug molecules and target proteins - Google Patents

Abstract

The application provides a method and a server for matching a drug molecule with target proteins. According to the method, structural information of drug molecules to be matched, structural information of target proteins, protein family group information and protein function description information are obtained, and the multi-mode characteristic information is characterized; predicting the correlation of target proteins and drug molecules according to the characterization information of the multi-modal characteristic information of the drug molecules and the target proteins, sequencing the objects to be screened (such as known target proteins), screening a plurality of candidate objects related to a given object (such as drug molecules), and realizing the rough screening of the objects to be screened; through the characterization information of multi-mode characteristic information such as molecular structure, protein family, protein function description and the like, a plurality of candidate objects obtained by coarse screening are finely discharged, possible target proteins of given drug molecules and drug molecules matched with the given target proteins are accurately and rapidly screened, and the efficiency of matching the drug molecules with the target proteins is improved.

Description

Method and server for matching drug molecules and target proteins

technical field

This application relates to computer technology, in particular to a method and server for matching drug molecules and target proteins.

Background technique

A drug target is a special site formed by a biomolecule formed by the interaction between a drug and a biomacromolecule in the body to produce pharmacological effects and achieve the purpose of preventing and treating diseases. It is the basis for the drug to function and is of great significance in the screening of new drugs. The matching method between drug molecules and target proteins can be applied to predict drug targets, which not only plays an irreplaceable role in the evaluation of the initial druggability of drug molecules, but also has great significance in the field of new use of old drugs after the drug matures. It can be applied to predict the drugs applicable to the target, which is of great significance in scenarios such as combined drug therapy.

At present, most of the matching methods between drug molecules and target proteins focus on matching drug molecules and target proteins one by one to predict whether there is an interaction relationship between drug molecules and target proteins. However, in practical applications, it is often necessary to screen a target (drug) that may interact with a given drug (or target) from a large number of targets (drugs), and it takes a long time to match the drug molecule with the target protein. ,low efficiency.

Contents of the invention

The present application provides a method and a server for matching drug molecules and target proteins to solve the problem of low efficiency of existing methods for matching drug molecules and target proteins.

In the first aspect, the present application provides a method for matching a drug molecule with a target protein, including: obtaining the structural information of the drug molecule to be matched, the structural information of the target protein, the protein family group information, and the protein function description information; The characterization information of the drug molecule and the characterization information of the target protein, the characterization information of the drug molecule includes the structural characterization of the drug molecule, the characterization information of the target protein includes the structural characterization of the target protein, protein family group Characterization and protein function description and characterization; according to the structural characterization of the drug molecule, the structural characterization of the target protein and the characterization of protein family groups, predict the correlation information between the target protein and the drug molecule, and according to the According to the above correlation information, select candidate objects from the objects to be screened, and the objects to be screened are the target protein or the drug molecule; according to the characterization information of the drug molecule and the characterization information of the target protein , sort the candidate objects, and output a matching result between the drug molecule and the target protein according to the sorting result.

In the second aspect, the present application provides a method for matching drug molecules and target proteins, including: responding to drug multi-target protein prediction instructions, obtaining structural information of a given drug molecule and structural information of target proteins to be screened , protein family group information and protein function description information; determine the characterization information of the drug molecule and the characterization information of the target protein, the characterization information of the drug molecule includes the structural characterization of the drug molecule, and the characterization information of the target protein The characterization information includes the structural characterization of the target protein, the characterization of the protein family group and the description and characterization of the protein function; according to the structural characterization of the drug molecule, the structural characterization of the target protein and the characterization of the protein family group, predict the target The correlation information between the protein and the drug molecule, and according to the correlation information, select the candidate target protein from the target protein to be screened; according to the characterization information of the drug molecule and the target protein To characterize the information, sort the candidate target proteins, and output the multi-target protein sequence matching the drug molecule according to the sorting results.

In a third aspect, the present application provides a method for matching a drug molecule with a target protein, including: responding to a matching instruction for a drug molecule applicable to the target protein, obtaining structural information, protein family group information, and protein Functional description information, and structural information of the drug molecule to be screened; determine the characterization information of the target protein and the characterization information of the drug molecule, the characterization information of the target protein includes the structural characterization of the target protein, protein family Group characterization and protein function description and characterization, the characterization information of the drug molecule includes the structural characterization of the drug molecule; according to the structural characterization of the drug molecule, the structural characterization of the target protein and the characterization of the protein family group, predict the Correlation information between the target protein and the drug molecule, and according to the correlation information, select candidate drug molecules from the drug molecules to be screened; according to the characterization information of the drug molecule and the target protein To characterize the information, sort the candidate drug molecules, and output the sequence of multi-drug molecules matching the target protein according to the sorting results.

In the fourth aspect, the present application provides a method for matching drug molecules and target proteins, which is applied to end-to-end devices, including:

Responding to the matching request between the drug molecule and the target protein, obtaining the information of the given object to be matched and the object to be screened, one of the given object and the object to be screened is the drug molecule and the other is the target protein; The information of the given object to be matched and the object to be screened is sent to the server, wherein the matching result is that the server, according to the information of the given object to be matched and the object to be screened, through any of the above aspects Determined by the method; receiving the matching result of the drug molecule and the target protein sent by the server; outputting the matching result of the drug molecule and the target protein.

In a fifth aspect, the present application provides a server, including: a processor, and a memory communicatively connected to the processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory, to Implement the methods described above.

The method and server for matching a drug molecule with a target protein provided in this application obtain the structural information of the drug molecule to be matched, the structural information of the target protein, the protein family group information, and the protein function description information; determine the characterization of the drug molecule Information and characterization information of the target protein, the characterization information of the drug molecule includes the structural characterization of the drug molecule, and the characterization information of the target protein includes the structural characterization of the target protein, the characterization of the protein family group and the description and characterization of the protein function; according to the characterization of the drug molecule Structural characterization, structural characterization of target protein and protein family group characterization, prediction of correlation information between target protein and drug molecules, and screening of candidate objects from objects to be screened based on correlation information Among the many objects to be screened, multiple candidates with high correlation with the given object are roughly screened out; furthermore, according to the characterization information of the drug molecule and the characterization information of the target protein, the candidate objects are sorted, and according to the ranking The results output the matching results of drug molecules and target proteins. Through the characterization information of multi-modal feature information such as molecular structure, protein family group, and protein function description, the multiple candidate objects obtained by rough screening can be fine-sorted, which can automatically And quickly screen out the to-be-screened object that matches the given object to realize the matching of the drug molecule and the target protein, and improve the matching efficiency of the drug molecule and the target protein.

Description of drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the application and together with the description serve to explain the principles of the application.

Fig. 1 is an exemplary system architecture diagram applicable to the present application;

Figure 2 is a flowchart of a method for matching a drug molecule with a target protein provided in an exemplary embodiment of the present application;

Fig. 3 is a framework diagram of the matching process between a drug molecule and a target protein provided by an exemplary embodiment of the present application;

Fig. 4 is a flowchart of a method for matching a drug molecule with a target protein according to another exemplary embodiment of the present application;

Fig. 5 is a flow chart of the method for training the molecular target relationship prediction model provided by an exemplary embodiment of the present application;

FIG. 6 is a flowchart of a method for sequence model training provided by an exemplary embodiment of the present application;

Fig. 7 is a flowchart of a method for matching a drug molecule with a target protein according to another exemplary embodiment of the present application;

Fig. 8 is a flowchart of a method for matching a drug molecule with a target protein according to another exemplary embodiment of the present application;

Fig. 9 is a flowchart of a method for matching a drug molecule with a target protein according to another exemplary embodiment of the present application;

Fig. 10 is a schematic structural diagram of a drug multi-target protein prediction device provided by an exemplary embodiment of the present application;

Fig. 11 is a schematic structural diagram of a server provided by an example embodiment of the present application.

By means of the above drawings, specific embodiments of the present application have been shown, which will be described in more detail hereinafter. These drawings and text descriptions are not intended to limit the scope of the concept of the application in any way, but to illustrate the concept of the application for those skilled in the art by referring to specific embodiments.

Detailed ways

Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, the same numerals in different drawings refer to the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with aspects of the present application as recited in the appended claims.

First, the nouns involved in this application are explained:

Target protein: refers to the drug target, also known as target protein, or protein target, generally refers to those proteins directly bound by drug molecules, such as enzymes, ion channels and receptors or other biomolecules (such as deoxyribonucleic acid DNA , ribonucleic acid RNA, heparin and peptides, etc.). Most drug targets are proteins, and chemicals used to treat or diagnose disease selectively interact with target proteins, resulting in changes in their biological pathways or functions.

Drug molecule: usually refers to a small molecule drug.

Multimodality: refers to data collaborative reasoning of multiple heterogeneous modalities. In the field of artificial intelligence, multimodal information often refers to perceptual information, such as heterogeneous modal information such as images, texts, and voices. In the embodiment of this application, the multimodal information includes the structural information (graph data) of the drug molecule and the target protein, the protein function description information (text) of the target protein, and the protein family group information of the target protein (such as coding sequence ), multiple heterogeneous modal information such as triplets composed of drug molecules, target proteins, and protein family group information.

Aiming at the time-consuming and low-efficiency problems of matching a drug molecule with a target protein, this application provides a method for matching a drug molecule with a target protein. By obtaining the structural information of the drug molecule to be matched, the structural information of the target protein , protein family group information and protein function description information; determine the characterization information of the drug molecule and the characterization information of the target protein, the characterization information of the drug molecule includes the structural characterization of the drug molecule, and the characterization information of the target protein includes the structure of the target protein Characterization, protein family group characterization and protein function description and characterization; according to the structural characterization of drug molecules, target protein structural characterization and protein family group characterization, predict the correlation information between target proteins and drug molecules, and based on the correlation information , to screen out candidate objects from the objects to be screened, which can quickly and roughly screen out multiple candidate objects with high correlation with a given object from a large number of objects to be screened; further, according to the characterization information of drug molecules According to the characterization information of the target protein, the candidate objects are sorted, and the matching result of the drug molecule and the target protein is output according to the sorting result, and the characterization of the multimodal characteristic information such as molecular structure, protein family group, protein function description, etc. Information, fine sorting of multiple candidate objects obtained by coarse screening, can automatically and quickly screen out the objects to be screened that match the given object, so as to realize the matching of drug molecules and target proteins, which greatly improves the ability of drug targets. Efficiency of point protein screening.

In a possible application scenario, it is applied to predicting multi-target protein sequences for a given drug. In this case, the given object is a drug molecule, and the object to be screened is a known target protein, or multiple user-specified A collection of known target proteins. Specifically, in response to the drug multi-target protein prediction instruction, obtain the structural information of the given drug molecule, as well as the structural information of the target protein to be screened, the protein family group information and the protein function description information; determine the characterization of the drug molecule Information and characterization information of the target protein, the characterization information of the drug molecule includes the structural characterization of the drug molecule, and the characterization information of the target protein includes the structural characterization of the target protein, the characterization of the protein family group and the description and characterization of the protein function; according to the characterization of the drug molecule Structural characterization, structural characterization of target proteins and protein family group characterization, predict the correlation information between target proteins and drug molecules, and screen candidate target proteins from target proteins to be screened based on the correlation information, Can roughly screen out multiple candidate target proteins that are highly related to a given drug molecule from a large number of known target proteins; further, according to the characterization information of the drug molecule and the characterization information of the candidate target protein , sort the candidate target proteins, and output the multi-target protein sequences that match the drug molecules according to the sorting results. Through the characterization information of multimodal characteristic information such as molecular structure, protein family group, protein function description, etc., the The fine sorting of multiple candidate target proteins obtained by coarse screening can automatically and quickly screen out the possible target protein sequences of a given drug molecule, which greatly improves the efficiency of drug target protein screening and saves A large number of manpower and experimental resources can shorten the cycle of multi-target drug development and promote the process of new drug development. Multiple target proteins in the output multi-target protein sequence can be referenced by users as target proteins that a given drug can act on.

In another possible application scenario, it is applied to predict suitable drug molecules for a given target protein, so as to provide relevant medical personnel with information on target drugs that can be selected. In this case, the given object is the target protein, and the object to be screened is a known drug molecule, or a set of multiple drug molecules specified by the user. Specifically, in response to the matching instruction of the drug molecule applicable to the target protein, the structural information, protein family group information, and protein function description information of a given target protein, as well as the structural information of the drug molecule to be screened; determine the target Protein characterization information and drug molecule characterization information. Target protein characterization information includes target protein structure characterization, protein family group characterization, and protein function description characterization. Drug molecule characterization information includes drug molecule structure characterization; Molecular structural characterization, target protein structural characterization and protein family group characterization, predict the correlation information between the target protein and drug molecules, and screen candidate drug molecules from the drug molecules to be screened based on the correlation information, It can roughly screen out multiple candidate drug molecules that are highly related to a given target protein from a large number of drug molecules. Further, according to the characterization information of the candidate drug molecules and the characterization information of the target protein, the candidate drug molecules are sorted, and the sequence of multi-drug molecules matching the target protein is output according to the sorting results. Characterization information of multi-modal feature information such as gender and protein function description, and fine sorting of multiple candidate drug molecules obtained by coarse screening, can automatically and quickly screen out effective drug molecule sequences for a given target protein, which is extremely It greatly improves the efficiency of molecular screening of target proteins for drugs. The multiple drug molecules included in the output multi-drug molecular sequence can be used as an effective drug for a given target protein for user reference.

FIG. 1 is an exemplary system architecture diagram applicable to this application. As shown in FIG. 1 , the system architecture may specifically include a server and end-side devices. Wherein, the server may specifically be a server cluster deployed in the cloud, or a computing device deployed locally. End-side devices can specifically be hardware devices with network communication functions, computing functions, and information display functions, including but not limited to smart phones, tablet computers, desktop computers, Internet of Things devices and other physical devices, and can also be application programs (APP) , software products or databases, etc. can be used as hardware/software devices that interact with the server as the client device on the user side.

The server stores the trained molecular target relationship prediction model and sequence model. The server can obtain the structural information of the drug molecule to be matched, the structural information of the target protein, the protein family group information and the protein function description information, and the structural information of the drug molecule, as well as the structural information of the target protein and the protein family Multimodal feature information such as group information and protein function description information are respectively characterized to determine the structural characterization of drug molecules, the structural characterization of target proteins, the characterization of protein family groups, and the description and characterization of protein functions. The server can also use the molecular-target relationship prediction model to predict the correlation information between the target protein and the drug molecule based on the structural characterization of the drug molecule, the structural characterization of the target protein, and the characterization of the protein family group. A plurality of candidate objects related to a given object are coarsely screened out from the objects to be filtered, thereby greatly reducing the number of candidate objects to be filtered. Furthermore, the server can also use the structural characterization of the drug molecule, the structural characterization of the target protein, the characterization of the protein family group and the description of the protein function to accurately sort the candidate objects to be screened, and output the drug molecule and Matching results for target proteins.

The server can also pre-store the information of the object to be screened, including at least one of the following: the structural information of the known target protein, the information of the protein family group and the description information of the protein function, and the structural information of the known drug molecule. Exemplarily, the server may store a pre-established molecular-target relationship knowledge graph, which includes the interaction relationship between known drug molecules and known target proteins, as well as the interaction relationship between known target proteins. Protein family group information. Based on the molecular-target relationship knowledge map, the triplet information of drug molecules, target proteins, and protein family groups can also be characterized, the triplet representation can be determined, and the representation information of multimodal feature information can be obtained.

In an exemplary usage scenario, the user can submit the information of the given object and the object to be screened to the server through the end-side device, for example, the structural information of a given drug molecule, or the structural information of a given target protein, protein Family group information and protein function description information. The server receives information for a given object. The server can also obtain the information of the object to be screened, such as the structural information of the given target protein, the information of the protein family group and the description information of the protein function, or, the structural information of the given drug molecule, the structural information of the drug molecule, the target Characterize the multi-modal feature information such as protein structure information, protein family group information, and protein function description information to determine the structural characterization of drug molecules, target protein structure characterization, protein family group characterization, and protein function description characterization . The server predicts the correlation information between the target protein and the drug molecule based on the structural characterization of the drug molecule, the structural characterization of the target protein, and the characterization of the protein family group, and roughly screens the given object from the objects to be screened based on the correlation information. Related multiple candidate objects to be screened, thereby greatly reducing the number of objects to be screened. According to the structural characterization of drug molecules, the structural characterization of target proteins, the characterization of protein family groups, and the characterization of protein functions, the server accurately sorts the candidate objects, and outputs the candidates that match the given object orientation according to the sorting results. object, to obtain the matching results between the drug molecule and the target protein. The server can also output the matching results of the drug molecule and the target protein to the end-side device.

Exemplarily, in Fig. 1, the system architecture is exemplified for predicting a multi-target protein sequence for a given drug. As shown in Fig. 1, the user can submit a drug multi-target protein prediction instruction to the server through the end-side device , and submit the given structural information of the drug molecule to the server, such as the structure diagram of the drug molecule. The server receives the drug multi-target protein prediction instruction, and receives the structural information of the given drug molecule. The server can also obtain structural information of known target proteins (objects to be screened), protein family group information, and protein function description information, as well as structural information of drug molecules, structural information of known target proteins, and protein family information. Multimodal feature information such as group information and protein function description information are respectively characterized to determine the structural characterization of drug molecules, the structural characterization of target proteins, the characterization of protein family groups, and the description and characterization of protein functions. The server predicts the correlation information between the target protein and the drug molecule based on the structural characterization of the drug molecule, the structural characterization of the target protein, and the characterization of the protein family group, and roughly screens out the known target proteins based on the correlation information. Multiple candidate target proteins associated with drug molecules, thereby greatly reducing the number of candidate target proteins. According to the structural characterization of the drug molecule, the structural characterization of the candidate target protein, the characterization of the protein family group and the description of the protein function, the server can accurately sort the candidate target proteins, and output the multi-target matching drug molecule according to the sorting results. Based on the characterization information of multimodal characteristic information such as molecular structure, protein family group information, and protein function description, the multiple candidate target proteins obtained by rough screening are fine-sorted, and according to the fine-sorting results Determine the multi-target protein sequence matching the drug molecule, improve the efficiency of matching the drug molecule and the target protein, and at the same time improve the accuracy of the matching between the drug molecule and the target protein. The server can also output the multi-target protein sequence matched by the drug molecule to the end-side device. Optionally, the information of the object to be filtered can also be specified by the user as a set containing multiple objects to be filtered. In the set of objects to be filtered specified by the user, the object to be filtered that matches the given object can be filtered, and the scope of filtering can be reduced. , effectively improving the matching efficiency of drug molecules and target proteins. For example, the user can specify a drug molecule and a set of target proteins to be screened; or, a user can specify a target protein and a set of drug molecules to be screened.

Based on the method provided by this application, the multi-target protein sequence that matches the determined drug molecule is predicted, and the server automatically analyzes and predicts the drug molecule and the known target based on the structural information of the drug molecule to be predicted and the information of the known target protein. The possibility information of protein interaction and the multi-target protein sequence of the drug molecule generated after the possibility sorting can predict the multi-target protein sequence for the drug molecule with drug potential. The multi-target protein sequence contains multiple alternatives The target protein, and the higher the candidate target protein in the sequence, the higher the possibility of interaction with the drug molecule. The target protein in the generated multi-target protein sequence can be selected to the kinase map, and the drug designer will screen out the target protein that meets the drug concept, and then enter the next biological activity test section, without the need for drug developers to manually screen the target This point can greatly shorten the cycle of multi-target drug development and promote the process of new drug development.

The technical solution of the present application and how the technical solution of the present application solves the above technical problems will be described in detail below with specific embodiments. The following specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. Embodiments of the present application will be described below in conjunction with the accompanying drawings.

Fig. 2 is a flowchart of a method for matching a drug molecule with a target protein according to an exemplary embodiment of the present application. The execution subject of this embodiment is the server mentioned above. As shown in Figure 2, the specific steps of the method of this embodiment are as follows:

Step S201, obtaining the structural information of the drug molecule to be matched, the structural information of the target protein, the protein family group information and the protein function description information.

Wherein, the structural information of the drug molecule may be a structure diagram of the drug molecule, which is characteristic information of the drug molecule. The structural information of the target protein may be a 3D structure diagram of the target protein. The protein family group information of the target protein is used to indicate the protein family (ie protein family group) to which the target protein belongs, and the protein family group of the target protein can be represented by a discrete value. The protein function description information of the target protein includes text information describing the function of the target protein, corresponding diseases, supporting genes, etc. The structural information, protein family group information, and protein function description information of the target protein are characteristic information of the target protein.

In this embodiment, the first of the drug molecule and the target protein is the given object, and the other is the object to be screened. The feature information of the given object may be received by the server from the feature information of the given object sent by the user through the end-side device used. The number of objects to be screened is usually large, such as the number of known target proteins, known drug molecules, and a certain type of drug molecules, etc., and the server can pre-store the feature information of the objects to be screened in the local storage space. It needs to be read from the local storage space.

Optionally, both the feature information of the given object and the feature information of the object to be screened may be sent to the server by the user through the end-side device used. Optionally, the feature information of the possible given object and the object to be screened can be stored on the server, and the user submits the identification of the given object and the identification of the object to be screened to the server through the end-side device used during specific use, The server reads the characteristic information of the given object from the local storage space based on the identification information of the given object, and reads the characteristic information of the object to be screened from the local storage space based on the identification of the object to be screened.

Step S202, determine the characterization information of the drug molecule and the characterization information of the target protein, the characterization information of the drug molecule includes the structural characterization of the drug molecule, and the characterization information of the target protein includes the structural characterization of the target protein, the characterization of the protein family group and the protein Functional description characterization.

In this example, the structural information of the drug molecule and the target protein are usually stored in the form of graph data, the protein family group information of the target protein is usually represented by the coding sequence, and the protein function description information of the target protein is text , which contains feature information of a variety of different modalities. After obtaining the structural information of the drug molecule, as well as multi-modal feature information such as the structural information of the target protein, protein family group information, and protein function description information, different information can be used to characterize the feature information of different modalities. Methods Characterization was carried out to obtain the structural characterization of the drug molecule, the structural characterization of the target protein, the characterization of the protein family group and the characterization of the protein function.

Exemplarily, for the structural characterization of a drug molecule, the structural information of the drug molecule is usually a graph, and the small molecule pre-training model, the drug molecule pre-training model, etc. can be used to characterize (embedding) the structure information of the input drug molecule to obtain the drug Structural characterization of molecules. Among them, the small molecule pre-training model and drug molecule pre-training can be any pre-trained model used to represent the structure diagram of a small molecule as a representation vector.

Exemplarily, for the structural characterization of the target protein, a pre-trained Transformer-based protein structure characterization model can be used to characterize the structural diagram of the target protein. For example, a pre-trained Transformer-based GPT model, such as GPT, GPT2.0, etc., can be used. Transformer is a neural network model that includes an encoding module and a decoding module, using a self-attention mechanism. GPT (Generative Pre-Training) model: It is a generative pre-training model. GPT and GPT2.0 are different versions of the GPT model.

Exemplarily, for the protein function description and characterization of the target protein, the protein function description information of the target protein is text information, and the protein function description information of the target protein can be represented as a protein function description characterization by using a pre-trained text representation model . For example, a Bert model or other pre-trained text representation models, text representation models, text encoding models, etc. may be used, which will not be described in detail here in this embodiment. In addition, the protein family group information of the target protein can be text information. At this time, a pre-trained text representation model can be used to represent the protein family group information of the target protein as a protein family group representation, which is not specifically limited here. . Among them, Bert (Bidirectional Encoder Representations from Transformers) is a Transformer-based bidirectional encoder, which aims to pre-train deep bidirectional representations from unlabeled text through conditional calculations shared in the left and right contexts. It is a neural network model for representing text. .

Exemplarily, the protein family group information of the target protein indicates the protein family to which the target protein belongs, which may be a discrete value or a coding sequence. For the protein family group characterization of the target protein, the protein family group information of the target protein can be directly mapped to the corresponding vector representation according to the preset mapping rules, or any of the existing technologies can be used to convert the discrete Numerical mapping is implemented in the form of representation vectors, which will not be repeated here.

In this embodiment, the corresponding characterization information is obtained by characterizing the structural information of the drug molecule, the structural information of the target protein, the protein family group information, and the protein function description information and other characteristic information of various modalities. Predicting the correlation between drug molecules and target proteins based on the characterization information of multimodal feature information can improve the accuracy of correlation information prediction.

Step S203, according to the structural characterization of the drug molecule, the structural characterization of the target protein, and the characterization of the protein family group, predict the correlation information between the target protein and the drug molecule, and select candidate objects from the objects to be screened according to the correlation information, The object to be screened is a target protein or a drug molecule.

In this step, according to the structural characterization of the drug molecule, the structural characterization of the target protein, and the characterization of the protein family group, the correlation information between the target protein and the drug molecule is predicted, so that the characterization of the multimodal feature information of the target protein can be integrated Information, automatically predicts the correlation information between the target protein and the drug molecule, and provides a data basis for the matching of the drug molecule and the target protein.

Wherein, the correlation information between the target protein and the drug molecule indicates the possibility of an interaction relationship between the drug molecule and the target protein. The higher the correlation information between the target protein and the drug molecule, the greater the possibility that the target protein has an interaction relationship with the drug molecule. The lower the correlation information between the target protein and the drug molecule, the less likely it is that the target protein has an interaction relationship with the drug molecule.

Further, according to the correlation information between the target protein and the drug molecule, the objects to be screened can be roughly sorted, and according to the results of the rough sorting, multiple objects to be screened with higher correlation information with the given object can be selected as candidates Objects, so as to achieve coarse screening of objects to be screened, greatly reducing the number of objects to be screened.

When this embodiment is applied to different application scenarios, the reference objects of the objects to be screened may be different, and the specified objects of the given objects may also be the same.

For example, in the scenario of predicting a multi-target protein sequence for a given drug, the given object refers to the given drug molecule, the object to be screened refers to the target protein to be screened, and the candidate object refers to the candidate target protein . In this step, according to the correlation information between the target protein to be screened and the given drug molecule, the target protein to be screened can be roughly sorted, and the correlation with the given drug molecule can be screened out according to the results of the rough sorting Multiple target proteins with higher information are used as alternative target proteins, so as to achieve coarse screening of target proteins and greatly reduce the number of target proteins to be screened. Usually, there are tens of thousands of known target proteins, and the number of candidate target proteins determined by coarse screening can be reduced to a few hundred.

For example, in the scenario of predicting applicable drug molecules for a given target protein, the given object refers to the given target protein, the object to be screened refers to the drug molecule to be screened, and the candidate object refers to the candidate drug molecule . In this step, according to the correlation information between the drug molecules to be screened and the given target protein, the drug molecules to be screened can be roughly sorted, and the correlation with the given target protein can be screened out according to the results of the rough sorting Multiple drug molecules with higher information are used as candidate drug molecules, so as to realize the coarse screening of drug molecules and greatly reduce the number of drug molecules to be screened.

Step S204, sort the candidate objects according to the characterization information of the drug molecule and the characterization information of the target protein, and output the matching result of the drug molecule and the target protein according to the sorting result.

In this step, for the multiple candidates determined by the coarse screening, according to the structural characterization of the target protein, the characterization of the protein family group and the description and characterization of the protein function, as well as the structural characterization of the drug molecule, analyze and predict the candidate and the given object Interaction possibility information, and generate fine sorting results after possibility sorting, to achieve precise sorting of candidate objects. According to the fine sorting results, several candidate objects with the highest ranking are selected to obtain the matching results between the drug molecule and the target protein. Wherein, the matching result of the drug molecule and the target protein may be a sequence of multiple candidate objects, and the higher the candidate object in the sequence, the higher the possibility of interacting with the given object. For example, in the scenario of predicting a multi-target protein sequence for a given drug, for multiple candidate target proteins of a given drug molecule determined by coarse screening, according to the structural characterization of the candidate target protein and the characterization of protein family groups and protein function description and characterization, as well as structural characterization of drug molecules, analyze and predict the possibility information of drug molecules interacting with known target proteins, and generate fine sorting results after sorting the possibilities, so as to realize the identification of alternative target proteins precise sorting. According to the fine sorting results, several candidate target proteins that are ranked first are selected to form the multi-target protein sequence of the drug molecule. Wherein, the predicted multi-target protein sequence of the drug molecule contains multiple candidate target proteins, and the higher the candidate target protein in the sequence, the higher the possibility of interacting with the drug molecule. The target protein in the generated multi-target protein sequence can be selected, and the target protein that meets the drug concept is screened out by the drug designer, and then enters the next biological activity test section, without the need for drug developers to manually screen the target protein , can greatly shorten the cycle of multi-target drug development and promote the process of new drug development.

For example, in the scenario of predicting applicable drug molecules for a given target protein, for multiple candidate drug molecules of a given target protein determined by coarse screening, according to the structural characterization of the given target protein, protein family group Characterization and protein function description and characterization, as well as the structural characterization of candidate drug molecules, analyze and predict the possibility information of the interaction between candidate drug molecules and a given target protein, and generate fine sorting results after sorting the possibilities, realizing the identification of candidate drugs Precise sequencing of drug molecules. According to the fine sorting results, a multi-drug molecule sequence composed of several candidate drug molecules that are ranked top is selected. Wherein, the predicted multi-drug molecule sequence contains multiple candidate drug molecules, and the higher the candidate drug molecule in the sequence, the higher the possibility of interacting with a given target protein. The drug molecules in the generated multi-drug molecular sequence can be selected as the recommended drugs for a given target protein, which can provide relevant personnel with accurate recommendations of available drugs, thereby improving the efficiency of relevant personnel.

In this embodiment, by obtaining the structural information of the drug molecule to be matched, the structural information of the target protein, the protein family group information, and the protein function description information; determining the characterization information of the drug molecule and the characterization information of the target protein, the drug molecule The characterization information of the drug molecule includes the structural characterization of the drug molecule, and the characterization information of the target protein includes the structural characterization of the target protein, the characterization of the protein family group, and the description and characterization of the protein function; Family group characterization, predicting the correlation information between target proteins and drug molecules, and screening candidate objects from the objects to be screened based on the correlation information, which can quickly and roughly screen out and give drugs from a large number of objects to be screened Identify multiple candidate objects with high correlation between objects; further, sort the candidate objects according to the characterization information of the drug molecule and the characterization information of the target protein, and output the matching result of the drug molecule and the target protein according to the sorting results , through the characterization information of multi-modal feature information such as molecular structure, protein family group, protein function description, etc., the multiple candidate objects obtained by rough screening are fine-sorted, and the matching of a given object can be automatically and quickly screened out. The objects to be screened are used to achieve the matching of drug molecules and target proteins, which greatly improves the efficiency of matching drug molecules and target proteins.

Fig. 3 is a framework diagram of a matching process between a drug molecule and a target protein provided in an exemplary embodiment of the present application, and Fig. 4 is a flow chart of a method for matching a drug molecule with a target protein provided in another exemplary embodiment of the present application. On the basis of the above-mentioned method embodiments, in this embodiment, the method for matching drug molecules and target proteins is described in detail in combination with the process framework shown in Figure 3, and the content described in the above-mentioned embodiments is included in this embodiment .

As shown in Figure 4, the specific steps of the method of this embodiment are as follows:

Step S401, obtaining the structural information of the drug molecule to be matched, the structural information of the target protein, the information of the protein family group and the description information of the protein function.

This step is implemented in the same way as the above step S201, for details, please refer to the relevant content of the above step S201, which will not be repeated here.

Step S402, determine the characterization information of the drug molecule and the characterization information of the target protein, the characterization information of the drug molecule includes the structural characterization of the drug molecule, and the characterization information of the target protein includes the structural characterization of the target protein, the characterization of the protein family group and the protein Functional description characterization.

Among them, the structural information of the drug molecule and the target protein are usually stored in the form of graph data, the protein family group information of the target protein is usually represented by the coding sequence, and the protein function description information of the target protein is text, including multiple feature information of different modes. After obtaining the structural information of the drug molecule, as well as multi-modal feature information such as the structural information of the target protein, protein family group information, and protein function description information, different information can be used to characterize the feature information of different modalities. Methods Characterization was carried out to obtain the structural characterization of the drug molecule, the structural characterization of the target protein, the characterization of the protein family group and the characterization of the protein function.

Exemplarily, for the structural characterization of a drug molecule, the structural information of the drug molecule is usually a graph, and the small molecule pre-training model, the drug molecule pre-training model, etc. can be used to characterize (embedding) the structure information of the input drug molecule to obtain the drug Structural characterization of molecules. Among them, the small molecule pre-training model and drug molecule pre-training can be any pre-trained model used to represent the structure diagram of a small molecule as a representation vector. In FIG. 3 , an example is illustrated by using a small molecule pre-training model to represent a structure diagram of a drug molecule as a structure representation of a drug molecule.

Exemplarily, as shown in FIG. 3 , for the structural characterization of the target protein, a pre-trained Transformer-based protein structure characterization model can be used to characterize the structural diagram of the target protein. For example, the protein structure characterization model can use a pre-trained Transformer-based GPT model, such as GPT, GPT2.0, etc.

Exemplarily, as shown in Figure 3, for the protein function description and characterization of the target protein, the protein function description information of the target protein is text information, and a pre-trained text representation model can be used to convert the protein function description information of the target protein Characterization is a description of protein function. For example, a Bert model or other pre-trained text representation models, text representation models, text encoding models, etc. may be used, which will not be described in detail here in this embodiment.

Exemplarily, the protein family group information of the target protein indicates the protein family to which the target protein belongs, which may be a discrete value or a coding sequence. As shown in Figure 3, for the protein family group characterization of the target protein, the protein family group information of the target protein can be directly mapped to the corresponding vector representation according to the preset mapping rules, or the existing technology can be used Any one of the ways to map discrete values into representation vectors is implemented, so I won’t go into details here. In addition, the protein family group information of the target protein can be text information. In this case, a pre-trained text representation model can be used to represent the protein family group information of the target point as a protein family group representation, which is not specifically limited here.

By using different pre-trained models to characterize the structural information of the drug molecule, as well as the structural information of the target protein, the protein family group information, and the protein function description information, the structure of the drug molecule is obtained. Characterization, structural characterization of target proteins, protein family group characterization, and protein function description characterization and other characterization information. Predicting the correlation between drug molecules and target proteins based on the characterization information of multimodal feature information can improve the accuracy of correlation information prediction.

Step S403, input the structural characterization of the drug molecule, the structural characterization of the target protein, and the characterization of the protein family group into the molecular-target relationship prediction model, and predict the correlation information between the target protein and the drug molecule through the molecular-target relationship prediction model.

In this example, the trained molecular-target relationship prediction model is obtained in advance. After obtaining the characterization information such as the structural characterization of the drug molecule, the structural characterization of the target protein, the characterization of the protein family group, and the characterization of the protein function, the trained Molecular-target relationship prediction model, which mines and predicts the correlation information between target proteins and drug molecules.

Specifically, the structural characterization of the drug molecule, the structural characterization of the target protein, and the characterization of the protein family group are input into the trained molecular-target relationship prediction model, and the relationship between the target protein and the drug molecule is predicted by the molecular-target relationship prediction model. sexual information.

Wherein, the correlation information between the target protein and the drug molecule indicates the possibility of an interaction relationship between the drug molecule and the target protein. The higher the correlation information between the target protein and the drug molecule, the greater the possibility that the target protein has an interaction relationship with the drug molecule. The lower the correlation information between the target protein and the drug molecule, the less likely it is that the target protein has an interaction relationship with the drug molecule.

Step S404 , selecting candidate objects from the objects to be screened according to the correlation information, where the objects to be screened are target proteins or drug molecules.

After the correlation information between the target protein and the drug molecule is predicted, the objects to be screened are roughly sorted according to the correlation information between the target protein and the drug molecule, and the correlation information with the given object is screened out according to the rough sorting results Higher multiple objects to be screened are used as candidate objects, so as to realize coarse screening of objects to be screened and greatly reduce the number of objects to be screened.

Exemplarily, in this step, according to the correlation information with the given object, sort the objects to be screened to obtain a second sorting result; determine a second preset number of objects to be screened according to the second sorting result, as the Object related multiple alternatives.

Wherein, the second preset number is a pre-set threshold of the number of candidate objects roughly screened out, which can be set and adjusted according to the needs of actual application scenarios, for example, the second preset number can be 300, 400, 500, etc. In addition, when applied to different application scenarios, different values may be used for the second preset quantity. Usually the second preset number is a larger value than the first preset number of target proteins in the finally determined multi-target protein sequence of the drug molecule. Therefore, the steps S403-S404 are a process of rough sorting and rough screening of objects to be screened, and a second preset number of candidate objects related to a given object are determined through rough screening.

Further, through the following steps S405-S407, according to the structural characterization of the drug molecule, the structural characterization of the target protein, the characterization of the protein family group and the description of the protein function, the candidate objects are accurately sorted to obtain the first sorting result. Further, through step S408, according to the first sorting result of precise sorting, a sequence including multiple candidate objects matching the given object can be determined, and the sequence includes a first preset number of candidate objects.

Step S405, according to the pre-established molecular target relationship knowledge map, characterize the triplets containing the given object and the candidate object, and obtain the triplet representation of the given object and the candidate object, and the molecular target relationship knowledge map Contains the interaction relationship between the known drug molecule and the known target protein, as well as the protein family group information of the known target protein. The triplet includes the drug molecule, the target protein and the protein family group of the target protein.

In an optional embodiment, according to the structural information of the known drug molecule, the structural information of the known target protein, the information of the protein family group and the description information of the protein function, the characterization is performed separately to obtain the structure of the known drug molecule Representation information of multimodal feature information such as characterization, structural characterization of known target proteins, protein family group characterization, and protein function description and characterization. Based on the target protein sequences of known drug molecules, a large-scale data set containing triplets of drug molecule-target protein-target protein family groups is constructed. Further, based on the dataset of triples, a molecular target relationship knowledge graph is established. The molecular-target relationship knowledge graph contains triplet information in the data set, that is, it includes the interaction relationship between known drug molecules and known target proteins, and the protein family group information of known target proteins.

Optionally, the molecular-target relationship knowledge map may also include structural information of drug molecules and target proteins, protein function description information of target proteins, and the like. When matching drug molecules and target proteins, the stored characterization information can be directly obtained, which can improve the matching efficiency of drug molecules and target proteins to a certain extent.

Optionally, the molecular target relationship knowledge map can also pre-store the structural representation of known drug molecules, the structural representation of known target proteins, the representation of protein family groups, and the description and representation of protein functions. When matching drug molecules and target proteins, the stored characterization data can be directly obtained, which can improve the matching efficiency of drug molecules and target proteins to a certain extent. For example, when performing multi-target prediction for any drug molecule, the stored characterization information of the target protein can be directly obtained to improve the efficiency of multi-target prediction.

Optionally, the structural information of the drug molecule and the target protein, and the protein function description information of the target protein may not be included in the molecular-target relationship knowledge graph, and be stored independently of the molecular-target relationship knowledge graph. In addition, the characterization data such as the structural representation of the known target protein, the representation of the protein family group, and the description and representation of the protein function may not be included in the knowledge map of the molecular target relationship, and be stored independently of the knowledge map of the molecular target relationship.

In this embodiment, for each given object, information about the given object and each of its candidate objects and protein family groups are formed into a triplet. The triplet is characterized according to the molecular target relationship knowledge map, and the triplet representation of each given object sample and each candidate object sample is obtained.

Exemplarily, in the scenario of predicting a multi-target protein sequence for a given drug, the TransH model can be used according to the molecular-target relationship knowledge map to compare the given drug molecule with each candidate target protein and candidate target Characterize the triplet composed of the protein family group of the target protein, and generate the triplet representation of each candidate target protein. The triplet representation is also another kind of modal feature information representation information, which is used to identify the candidate target protein. The characterization data based on the precise sequencing of the selected target proteins can improve the accuracy of the sequencing of the candidate target proteins, thereby improving the accuracy of the multi-target protein sequences matched by drug molecules. Among them, the TransH model is a knowledge base automatic representation learning model, which can be used to represent the triples in the knowledge graph as feature vectors.

Exemplarily, in the scenario of predicting applicable drug molecules for a given target protein, for each candidate drug molecule of a given target protein, and the given target The triplets composed of proteins and protein family groups of a given target protein are characterized to generate a triplet representation of each candidate drug molecule and a given target protein, and the triplet representation is also used as another model The characterization information of the state characteristic information is used for the characterization data based on the fine sorting of candidate drug molecules, which can improve the accuracy of the sorting of candidate drug molecules, thereby improving the accuracy of the matching results between drug molecules and target proteins.

Step S406, input the characterization information of the given object, the characterization information of the candidate object, and the triplet characterization of the given object and the candidate object into the sequence model, and use the coding module of the sequence model to input the characterization information of the given object, the candidate The representation information of the selected object is fused with the triplet representation of the given object and the candidate object to obtain the fusion representation of the candidate object.

In this step, according to the characterization information of the given object, the characterization information of the candidate object, and the triplet characterization of the given object and the candidate object and other characterization information, these characterization information are input into the trained sequence model, through the sequence model The encoding module of the given object is fused with the representation information of the given object, the representation information of the candidate object, and the triplet representation of the given object and the candidate object to obtain the fusion representation of each candidate object.

In another optional embodiment, it is also possible to input the sequence model according to the characterization information of the given object and the characterization information of the candidate objects, and use the coding module of the sequence model to convert the characterization information of each candidate object, It is fused with the representation information of the given object to obtain the fusion representation of each candidate object. When training the sequence model, the training samples in the training set only need to contain the representation information of the given object and the representation information of the object to be screened, and the triplet representation may not be included.

In this embodiment, the characterization information of the multimodal feature information is fused through the encoding module of the sequence model, so that the fused characterization of each candidate object contains rich features, and based on the fused characterization, it is possible to accurately predict the relationship between each candidate object and The likelihood of an interaction relationship for a given object.

Exemplarily, the encoding module of the sequence model can be implemented using an encoding layer of a transformer, a long-short-term memory network (LongShort-Term Memory, LSTM for short), and the like. For example, as shown in Figure 3, the encoding module of the sequence model can be stacked using the 8-layer encoder (encoder) layer of the transformer model (indicated by "8×"), and each encoding layer contains a cross-attention layer and a feed-forward neural network. Network (Feed-Forward Network, FFN for short) layer.

Step S407, sort the candidate objects according to the fusion representation of the candidate objects through the sorting module of the sequence model, and obtain the first sorting result.

After obtaining the fused representation of each candidate object, the candidate objects are sorted according to the fused representation of each candidate object through the sorting module of the sequence model to obtain a first sorting result. The sorting order of the candidate objects in the first sorting result reflects the possibility of interaction between the candidate objects and the given object, the higher the possibility of interaction between the candidate objects and the given object, The further back the candidate objects are, the less likely they are to have interactions with the given object.

Wherein, the sorting module of the sequence model can be realized by using a list method (ListWise), for example, specifically, it can be realized by using ListNet, LambdaRank, or other similar sorting methods. ListNet and LambdaRank are commonly used list sorting algorithms.

Step S408, according to the first sorting result, determine a first preset number of candidate objects as target objects matching the given object, and obtain a matching result of the drug molecule and the target protein.

In this step, according to the first sorting result, intercept the first preset number of candidate objects that are ranked higher as the target objects that match the given object, and form a target object sequence, and the arrangement of each target object in the target object sequence The order is the same as the sort order in the first sorted result.

Wherein, the first preset number is a preset threshold of the number of target objects included in the matching result, which can be set and adjusted by the user or relevant personnel according to the needs of the actual application scenario. For example, the first preset number may be 20, 30, 40, 50 and so on. Usually the first preset number is a smaller value than the number of candidate objects (second preset number) in the rough screening result. Therefore, the steps S405-S408 are the process of finely sorting and finely screening the candidate objects. After the fine screening, the first preset number of candidate objects matching the given object are determined, and the number of candidate objects matching the given object is obtained. The sequence of the target object, that is, the matching result of the drug molecule and the target protein.

Step S409, outputting the matching result of the drug molecule and the target protein.

After determining the matching result of the drug molecule and the target protein, the matching result of the drug molecule and the target protein can be visualized, so that the information of the multi-target object sequence matching the given object and the relevant information of the target object can be visualized. (such as structural information, identification information) and other information are displayed visually to provide users with reference data.

For example, in the scenario of predicting a multi-target protein sequence for a given drug, the matching result is the multi-target protein sequence that matches the given drug molecule, and the relevant information of each target protein in the multi-target protein sequence is output through visualization , to provide reference data to relevant drug developers to select druggable drug molecules from the perspective of drug design through manual interactive screening, and to derivate drug structures for target proteins of specific diseases, so as to put them into subsequent drug development links.

In this example, by obtaining the structural information of the drug molecule, as well as the structural information of the target protein, the information of the protein family group and the description information of the protein function, and using different pre-trained models for the characteristic information of different modalities to generate the drug molecule Representation information of multimodal feature information such as structural characterization, structural characterization of target protein, protein family group characterization, and protein function description; according to the characterization information of multimodal feature information, use the trained molecular target relationship prediction model Predict the correlation information between the target protein and the drug molecule, and select multiple objects to be screened with high correlation information with the given object as candidate objects, so as to realize the coarse screening of the objects to be screened and greatly reduce the number of objects to be screened quantity. Further, through the established molecular-target relationship knowledge graph, based on the molecular-target relationship knowledge graph, the triplet composed of the drug molecule target protein and the protein family group of the target protein is characterized, and prepared Select the triplet representation of the selected object to obtain the representation information of the feature information of one mode. According to the representation information of the multi-modal feature information, use the trained sequence model to combine the representation information of each candidate object with the triplet Representation, and the representation information of the given object, to obtain the fusion representation of each candidate object, and according to the fusion representation of each candidate object, accurately sort the candidate objects, and determine the match with the given object according to the precise sorting result The sequence of the target to be screened can quickly and accurately screen out the possible sequence of the target to be screened for a given object, which greatly improves the matching efficiency of drug molecules and target proteins.

Fig. 5 is a flowchart of a method for training a molecular target relationship prediction model provided by an exemplary embodiment of the present application. On the basis of any of the aforementioned method embodiments, as shown in Figure 5, the molecular target relationship prediction model can be trained through the following steps:

Step S501, obtaining a pre-established molecular-target relationship knowledge graph, which includes known interaction relationships between drug molecules and target proteins, and protein family group information of target proteins.

In this example, based on the target protein sequences of known drug molecules, a large-scale data set including triplets of drug molecule-target protein-protein family groups was constructed. Further, based on the dataset of triples, a molecular target relationship knowledge graph is established. The molecular-target relationship knowledge graph contains triplet information in the data set, that is, it includes the interaction relationship between known drug molecules and known target proteins, and the protein family group information of known target proteins.

Exemplarily, in the knowledge graph of molecular-target relationship, drug molecules and target proteins are respectively used as nodes in the graph, and the associated edges between drug molecules and target proteins indicate that there is an interaction relationship between drug molecules and target proteins. The protein family group can be used as a node in the graph, or the attribute information of the node corresponding to the target protein in the graph.

Optionally, the molecular-target relationship knowledge graph may also include structural information of drug molecules and target proteins, protein function description information of target proteins, and the like.

Optionally, the molecular target relationship knowledge map can also pre-store the structural representation of known drug molecules, the structural representation of known target proteins, the representation of protein family groups, and the description and representation of protein functions. When performing multi-target prediction for any drug molecule, the stored characterization information of the target protein can be directly obtained to improve the efficiency of multi-target prediction.

For example, the structural characterization of a drug molecule can be used as an attribute information of a drug molecule node, and the structural characterization of a known target protein, the characterization of a protein family group, and the description of a protein function can be used as an attribute of a node corresponding to a target protein The information is stored in the molecular target relationship knowledge graph.

Optionally, the structural information of the drug molecule and the target protein, and the protein function description information of the target may not be included in the molecular-target relationship knowledge graph, and be stored independently of the molecular-target relationship knowledge graph. In addition, the characterization data such as the structural representation of the known target protein, the representation of the protein family group, and the description and representation of the protein function may not be included in the knowledge map of the molecular target relationship, and be stored independently of the knowledge map of the molecular target relationship.

Step S502, according to the molecular-target relationship knowledge map, sample and generate a comparative learning training set, each sample in the comparative learning training set is a triplet containing drug molecules, target proteins corresponding to drug molecules, and protein family group information of target proteins , the contrastive learning training set contains positive and negative samples.

In this embodiment, the molecular target relationship prediction model can be obtained by performing comparative learning training on the deep learning model. The comparative learning training set used in the comparative learning training can be generated based on the molecular target relationship knowledge map.

In this step, multiple triplets are sampled from the molecular-target relationship knowledge map, and the triplets include drug molecules, target proteins, and protein family group information of target proteins. Each triplet is used as a sample to form a contrastive learning training set. The contrastive learning training set contains positive and negative samples. Among them, the positive sample is the correct triplet, and there is an interaction relationship between the drug molecule and the target protein in the positive sample. Negative samples are nonexistent triplets or wrong triplets, and there is no interaction relationship between drug molecules and target proteins in negative samples.

Exemplarily, a positive sample can be composed of drug molecules, target proteins that have an interaction relationship with the drug molecules, and protein family group information of the target proteins from the molecular target relationship knowledge map; from the molecular target In the relational knowledge graph, negative samples are composed of drug molecules, target proteins that do not have an interaction relationship with the drug molecules, and protein family group information of the target proteins; positive samples and negative samples constitute a comparative learning training set.

Optionally, from the molecular-target relationship knowledge map, sample drug molecules, target proteins that do not have an interaction relationship with the drug molecule, and protein family group information of the target protein to form a negative sample, and the following can also be used Enhanced sampling is achieved by:

From the molecular target relationship knowledge map, sample drug molecules, the first target protein that has an interaction relationship with the drug molecule, the protein family group information of the first target protein, the drug molecule, the first target protein and The protein family group information of the first target protein constitutes the positive sample. Further, according to the protein family group information of the first target protein, a second target protein with the same group as the first target protein is sampled from the molecular target relationship knowledge map; the drug molecule, the second target protein The protein family group information of the second target protein and the second target protein constitutes a negative sample.

For example, a drug molecule A can be randomly sampled from the molecular target relationship knowledge graph, and the target protein B1 that has an interaction relationship with the drug molecule A (with associated edges), and the protein family group of the target protein B1 C, get positive samples (drug molecule A, target protein B1, protein family group C). Further, according to the protein family group C, another target protein B2 with the same protein family group C is obtained, and there is no interaction relationship between the drug molecule A and the target protein B2 in the molecular target relationship knowledge map ( does not have associated edges), get negative samples (drug molecule A, target protein B2, protein family group C).

Through this enhanced sampling strategy, the quality of negative samples can be improved, thereby improving the training effect of the comparative learning of the molecular target relationship prediction model, and improving the prediction accuracy of the molecular target relationship prediction model.

Exemplarily, a drug molecule, a target protein that has an interaction relationship with the drug molecule, and protein family group information of the target protein can be sampled from the molecular-target relationship knowledge map to form a positive sample. Positive samples are augmented using a sample augmentation strategy to generate multiple negative samples.

Step S503 , according to the comparative learning training set, perform comparative learning training on the molecular target relationship prediction model to obtain a trained molecular target relationship prediction model.

Based on the generated comparative learning training set, the molecular target relationship prediction model is trained for comparative learning, so that the correlation information between the drug molecule and the target protein in the negative sample predicted by the molecular target relationship prediction model tends to be minimal by default value (such as 0), while the correlation information between the drug molecule and the target protein in the positive sample predicted by the molecular target relationship prediction model tends to the preset maximum value (such as 1).

Exemplarily, based on the samples in the non-learning training set, the representation information of the multimodal feature information of the samples is obtained, specifically including the structural representation of the drug molecule in the sample, the structural representation of the target protein, and the representation of the protein family group. Input the characterization information of the multimodal feature information of the sample into the molecular target relationship prediction model, predict the correlation information between the drug molecule and the target protein in the sample through the molecular target relationship prediction model, and calculate the contrast loss function value, by minimizing The comparison loss function value makes the correlation information between the drug molecule and the target protein in the negative sample in the prediction result tend to the preset minimum value (such as 0), while the molecular target relationship prediction model predicts that the drug molecule in the positive sample The correlation information between the molecule and the target protein tends to the preset maximum value (such as 1).

In this embodiment, by pre-constructing the molecular target relationship knowledge map, a comparative learning training set is generated based on the knowledge map sampling, and according to the comparative learning training set, the modules used to predict correlation information in the molecular target relationship prediction model are compared Learn and train to improve the recall rate of training results and obtain a trained molecular target relationship prediction model. Through the trained molecular target relationship prediction model, the correlation information between the drug molecule and the target protein can be predicted more accurately according to the structure representation of the input drug molecule, the structure representation of the target protein and the representation of the protein family group. And according to the correlation information, the objects to be screened are roughly sorted, and multiple candidate objects with high correlation with the given object can be roughly screened out according to the rough sorting results, thereby greatly reducing the number of objects to be screened and improving the relationship between drug molecules and targets. Point protein matching efficiency.

Fig. 6 is a flowchart of a method for training a sequence model provided by an exemplary embodiment of the present application. On the basis of any of the aforementioned method embodiments, as shown in Figure 6, the sequence model can be obtained by training through the following steps:

Step S600 , sampling the given object sample and the object samples to be screened corresponding to the given object sample from the knowledge graph of molecular-target relationship.

In this embodiment, based on the pre-built molecular target relationship knowledge graph, a large number of given objects can be randomly sampled from the molecular target relationship knowledge graph as samples, which is called a given object sample in this embodiment. For each given object sample, a larger number of objects to be screened is sampled from the molecular target relationship knowledge map, and used as the object sample to be screened corresponding to the given object sample. The objects to be filtered corresponding to different given objects may be different or the same. For example, all the objects to be screened in the molecular target relationship knowledge graph can be used as the object samples to be screened corresponding to each given object sample.

Exemplarily, in the scenario of predicting a multi-target protein sequence for a given drug, a drug molecule sample is sampled as a given drug molecule, and the target protein in the molecular-target relationship knowledge map is used as the target sample to be screened. In the scenario of predicting applicable drug molecules for a given target protein, the target protein sample is sampled as a given target protein, and the drug molecule in the molecular-target relationship knowledge map is used as the sample to be screened.

Step S601, using the trained molecular target relationship prediction model, according to the characterization information of the given object sample and the corresponding characterization information of the object sample to be screened, to predict the correlation information between the given object sample and the object sample to be screened, and according to The correlation information screens out a plurality of candidate object samples related to a given object sample from the object samples to be screened.

In this step, for each given object sample, input the characterization information of the given object sample and the characterization information of each object sample to be screened corresponding to the given object sample into the trained molecular target relationship prediction model, Through the trained molecular target relationship prediction model, predict the correlation information between the given object sample and each object sample to be screened, and determine multiple to-be-screened objects with high correlation with the given object sample according to the relevant information An object sample, as a number of candidate object samples for that given object sample.

For example, in the scenario of predicting a multi-target protein sequence for a given drug, for each drug molecule sample, the substructure characterization of the drug molecule sample, as well as the structural characterization and protein Family group characterization, input the trained molecular target relationship prediction model, through the trained molecular target relationship prediction model, predict the correlation information between the drug molecule sample and the known target protein, and determine the multiplicity of the drug molecule sample A candidate target protein sample.

For example, in the scenario of predicting applicable drug molecules for a given target protein, for each target protein sample, the structural characterization and protein family group characterization of the target protein sample, as well as the drug molecule to be screened Input the trained molecular target relationship prediction model, through the trained molecular target relationship prediction model, predict the correlation information between the drug molecule sample to be screened and the target protein, and determine the target unit sample Multiple candidate drug molecule samples.

The specific implementation of determining multiple candidate object samples for each given object sample in this step is the same as using the trained molecular target relationship prediction model in the above steps S403-S404 to predict multiple candidate objects for a given object The implementation manners are similar, and for details, refer to the related descriptions in the foregoing embodiments, and details are not repeated here in this embodiment.

Step S602, according to the molecular target relationship knowledge map, characterize the triplets including the given object sample and the candidate object sample, and obtain the triplet representation of the given object sample and the candidate object sample, the triplet includes the drug Molecules, target proteins, and protein family groups of target proteins.

Specifically, for each given object sample, information about the given object sample and each of its candidate object samples and protein family groups are formed into a triplet. The triplet is characterized according to the molecular target relationship knowledge map, and the triplet representation of each given object sample and each candidate object sample is obtained.

For example, in the scenario of predicting a multi-target protein sequence for a given drug, for each drug molecule sample, the drug molecule sample and each candidate target protein sample and candidate target protein sample of the drug molecule sample The information of the protein family group is composed of a triplet. Based on the molecular target relationship knowledge map, the triplet is characterized, and the triplet representation of the drug molecule sample and each candidate target protein sample is obtained.

For example, in the scenario of predicting applicable drug molecules for a given target protein, for each target protein sample, each candidate drug molecule sample of the target protein sample is associated with the target protein sample and the target protein sample The information of the protein family group of the protein sample forms a triplet. Based on the molecular target relationship knowledge map, each triplet is characterized, and the triplet representation of the target protein sample and each candidate drug molecule sample is obtained.

In this step, any given object sample is processed to determine the triplet representation of the given object sample and each candidate object sample. The specific implementation method is similar to the specific implementation method of the above-mentioned step S405. For details, refer to the above-mentioned embodiment. The relevant content of this embodiment will not be repeated here.

Step S603, the characterization information of any given object sample, the characterization information of multiple candidate object samples of the given object sample, and the triplet characterization of the given object sample and the candidate object samples to form a training sample.

In this embodiment, each training sample used in the training sequence model training set includes: characterization information of a given object sample, characterization information of multiple candidate object samples, characterization information of a given object sample and each candidate object sample triplet representation.

Exemplarily, in the scenario of predicting multi-target protein sequences for a given drug, the training samples include: structural characterization of drug molecule samples, structural characterization of multiple candidate target protein samples, protein family group characterization, and protein function Describe the characterization, the triplet characterization corresponding to the drug molecule sample and each candidate target protein sample.

Exemplarily, in the scenario of predicting applicable drug molecules for a given target protein, the training samples include: structural characterization of target protein samples, protein family group characterization and protein function description characterization, multiple candidate drug molecule samples The structural characterization of the target protein sample and the triplet characterization corresponding to each candidate drug molecule sample.

Step S604, acquiring matching result annotation information of the training samples, and the training samples and the matching result annotation information of the training samples constitute a training set.

Wherein, the matching result annotation information of the training samples includes: multiple candidate object sample sequences of a given object sample. Exemplarily, for each training sample, the labeling information of the matching result of the training sample may be obtained through manual labeling.

For example, in the scenario of predicting a multi-target protein sequence for a given drug, the matching result annotation information includes: the multi-target protein sample sequence of the drug molecule sample. In the scenario of predicting applicable drug molecules for a given target protein, the matching result annotation information includes: the multi-drug molecule sample sequence of the target protein sample.

Step S605, using the training set to train the sequence model to obtain a trained sequence model.

Specifically, based on the training set, the representation information of the given object sample in the training sample, the representation information of multiple candidate object samples of the given object sample, and the triplet representation of the given object sample and the candidate object samples are input as The sequence model, through the encoding module of the sequence model, combines the representation information of the given object sample, the representation information of the candidate object sample, and the triplet representation of the given object sample and the candidate object sample to obtain the fusion of candidate object samples Representation; through the sorting module of the sequence model, according to the fusion representation of the candidate object samples, the candidate object samples are sorted, and the matching prediction result is determined; the loss function value is calculated according to the matching prediction result and the matching result labeling information of the training sample, And optimize the model parameters of the encoding module and the sorting module according to the loss function value to obtain a trained sequence model.

Taking the scenario of predicting a multi-target protein sequence for a given drug as an example, the effect of this embodiment is described. Based on the pre-built molecular target relationship knowledge map, multiple drug molecule samples are obtained by using the trained molecular target relationship prediction. A model that predicts multiple candidate target protein samples for each drug molecule sample. Based on the pre-constructed molecular target relationship knowledge map, characterize the triplets composed of each drug molecule sample, each candidate target protein sample and the protein family group of the candidate target protein sample, and obtain a triplet representation , the structural characterization of the drug molecule sample, the structural characterization of each candidate target protein sample, the characterization of protein family group and protein function description, and the characterization information such as triplet characterization, constitute the training sample, and obtain the annotation of the training sample information to get the training set. The sequence model is trained based on the training set to improve the recall rate of the training result and obtain a trained sequence model. Through the trained sequence model, the candidate target proteins can be accurately sorted according to the characterization information of the input multimodal feature information, and the multi-target protein sequence of the drug molecule can be determined accurately, so that the predicted Short multi-target protein sequences have higher target recall.

The drug multi-target prediction scheme provided by this application uses the latest knowledge map technology to carry out rough sorting and rough screening of known target proteins (or drug molecules), and obtain a small number of candidate target proteins (or candidate drug molecules). Select drug molecules), combined with transformer coding module and ListWise sorting algorithm and other characterization information that fuses multi-modal feature information to fine-sort and fine-screen candidate target proteins (or candidate drug molecules), and obtain multi-target drug molecules Protein sequence (or multi-drug molecular sequence matched with target protein), compared with other techniques for predicting drug-target relationship, can more efficiently match drug molecules with target proteins, by introducing comparative learning To train the molecular target relationship prediction model for rough sorting, it can greatly narrow the scope of searching for multi-target proteins (or multi-drug molecules), and the molecular target relationship prediction model obtained in this way does not only stop at the force of drug molecular targets In terms of prediction tasks, it can also be used to evaluate compounds with good properties for disease medication, provide efficient and accurate target protein reference sequences for drug development, and can also be used in drug search tasks for target proteins corresponding to major diseases. At the same time, the "rough sorting" + "fine sorting" method is adopted, which is superior to other target screening models and drug screening models in terms of matching accuracy and computational performance between drug molecules and target proteins.

Through the scheme provided by this application, verified by the results of wet laboratory experiments, it has the ability to predict molecular multi-target protein tasks, and efficiently provides screening results of multi-target protein sequences of drug molecules for drug researchers. Drug development no longer requires drug developers to manually perform high-throughput screening from tens of thousands of known target proteins for multi-target protein localization, only the multi-target protein sequence of the drug molecule automatically generated based on the scheme of this application , a small amount of verification experiments are carried out from the multi-target protein sequence, which greatly reduces the workload of target protein screening, saves manpower, material resources and time costs, and can quickly obtain multi-target protein lead compounds and drugs that meet the research diseases. The efficiency of research and development has been greatly improved, and the process of new drug research and development has been promoted.

Fig. 7 is a flowchart of a method for matching a drug molecule with a target protein provided in another exemplary embodiment of the present application. The method provided in this embodiment is applied to the scenario of predicting a multi-target protein sequence for a given drug. As shown in Figure 7, the specific steps of the method are as follows:

Step S701, in response to the drug multi-target protein prediction instruction, obtain the structural information of a given drug molecule, as well as the structural information, protein family group information and protein function description information of the target protein to be screened.

Among them, the drug multi-target prediction instruction refers to the instruction sent by the end-side device to the server based on the user's drug multi-target prediction request, which is used to instruct the server to obtain the information of a given drug molecule and the target protein to be screened, and execute Matching method of drug molecule and target protein.

Optionally, the drug multi-target protein prediction instruction includes identification information of a given drug molecule and/or identification information of a target protein to be screened. According to the received drug multi-target protein prediction instruction, the server extracts the identification information of a given drug molecule and/or the identification information of the target protein to be screened from the instruction. Further, if the identification information of the given drug molecule is extracted, the structural information of the given drug molecule is obtained according to the identification information of the given drug molecule. If the identification information of the target protein to be screened is extracted, according to the identification information of the target protein to be screened, the structural information, protein family group information and protein function description information of the target protein to be screened are obtained.

Optionally, the drug multi-target protein prediction instruction may include identification information of a given drug molecule. The server obtains the structural information of a given drug molecule according to the identification information of the given drug molecule; takes the known target protein as the target protein to be screened, and obtains the structural information of the known target protein, protein family group Identification information and protein function description information.

Optionally, the drug multi-target protein prediction instruction may also include structural information of a given drug molecule. The server takes the known target protein as the target protein to be screened, and obtains the structural information, protein family group information and protein function description information of the known target protein.

In this embodiment, the storage method of the various information of the drug molecule and the target protein can adopt the storage method provided in any of the above-mentioned method embodiments. Correspondingly, the method of obtaining various information of the drug molecule and the target protein is the same as For the storage method, refer to the relevant content in the foregoing method embodiments for details, and details are not repeated here.

Step S702, determine the characterization information of the drug molecule and the characterization information of the target protein, the characterization information of the drug molecule includes the structural characterization of the drug molecule, and the characterization information of the target protein includes the structural characterization of the target protein, the characterization of the protein family group and the protein Functional description characterization.

The specific implementation manner of this step is similar to the implementation manner in the above-mentioned step S202 or S402. For details, refer to the relevant content of the above-mentioned step S202 or S402, which will not be repeated here.

Step S703: According to the structural characterization of the drug molecule, the structural characterization of the target protein, and the characterization of the protein family group, predict the correlation information between the target protein and the drug molecule, and screen from the target proteins to be screened according to the correlation information Alternative target proteins.

In this step, the drug molecule is used as the given object, and the target protein is used as the object to be screened, which is realized by a method similar to the above steps S403-S404. For details, please refer to the relevant content of the above steps S403-S404, which will not be repeated here.

Step S704, sort the candidate target proteins according to the characterization information of the drug molecule and the characterization information of the target protein, and output the sequence of the multi-target protein that matches the drug molecule according to the sorting result.

In this step, the drug molecule is used as a given object, the target protein is used as an object to be screened, and the candidate target protein is used as a candidate object. The matching result of the dot protein is the multi-target protein sequence that matches the given drug molecule. For details, please refer to the relevant content of the above steps S405-S409, which will not be repeated here.

The method provided in this example can be applied to the scenario of predicting a multi-target protein sequence for a given drug, and obtain the structural information of a given drug molecule and the target protein to be screened in response to the drug multi-target protein prediction instruction structure information, protein family group information, and protein function description information, and use different pre-training models to generate structural representations of drug molecules, target protein structural representations, protein family group representations, and protein Representation information such as functional description and characterization; according to the characterization information of multimodal feature information, use the trained molecular target relationship prediction model to predict the correlation information between the target protein and the drug molecule, and screen out the high correlation information with the drug molecule Multiple target proteins can be used as candidate target proteins, so as to achieve coarse screening of candidate target proteins and greatly reduce the number of candidate target proteins. Further, through the established molecular-target relationship knowledge map, based on the molecular-target relationship knowledge map, the triplet composed of the drug molecule and each candidate target protein and the group of the candidate target protein is characterized, and the obtained The triplet characterization of the drug molecule and each candidate target protein obtains the characterization information of another modal feature information. According to the characterization information of the multi-modal feature information, using the trained sequence model, each candidate The structural characterization of the target protein, the characterization of the protein family group, the characterization of the protein function, and the corresponding triplet characterization are fused with the structural characterization of the drug molecule to obtain the fusion characterization of each candidate target protein, and according to each candidate Fusion characterization of selected target proteins, precise sequencing of candidate target proteins, and determination of multi-target protein sequences matching drug molecules based on the precise sequencing results, which can quickly and accurately screen possible targets for a given drug molecule The protein sequence greatly improves the efficiency of drug target screening, and saves a lot of manpower and experimental resources through artificial intelligence, thereby shortening the cycle of multi-target drug development and promoting the process of new drug development.

FIG. 8 is a flowchart of a method for matching a drug molecule with a target protein provided in another exemplary embodiment of the present application. The method provided in this embodiment is applied to the scenario of predicting an applicable drug molecule for a given target protein. As shown in Figure 8, the specific steps of the method are as follows:

Step S801, in response to the matching instruction of the drug molecule applicable to the target protein, obtain the structural information, protein family group information and protein function description information of a given target protein, as well as the structural information of the drug molecule to be screened.

Among them, the matching instruction of the drug molecule applicable to the target protein refers to the instruction sent by the end-side device to the server based on the user's request to match the applicable drug for the target protein, and is used to instruct the server to obtain the given target protein and the target protein to be screened. The information of the drug molecule, and the matching method of the drug molecule and the target protein.

Optionally, the matching instruction for a drug molecule applicable to a target protein includes identification information of a given target protein and/or identification information of a drug molecule to be screened. The server extracts the identification information of the given target protein and/or the identification information of the drug molecule to be screened from the matching instruction according to the received matching instruction of the drug molecule applicable to the target protein. Further, if the identification information of the given target protein is extracted, according to the identification information of the given target protein, the structural information, protein family group information and protein function description information of the given target protein are obtained. If the identification information of the drug molecule to be screened is extracted, the structural information of the drug molecule to be screened is obtained according to the identification information of the drug molecule to be screened.

Optionally, the matching instruction for a drug molecule applicable to a target protein may include identification information of a given target protein. According to the identification information of the given target protein, the server obtains the structural information, protein family group information and protein function description information of the given target protein; takes the known drug molecule as the drug molecule to be screened, and obtains the known structural information of drug molecules. Optionally, the matching instruction for the drug molecule applicable to the target protein may include structural information, protein family group information and protein function description information of a given target protein. The server uses the known drug molecule as the drug molecule to be screened, and obtains the structure information of the known drug molecule.

In this embodiment, the storage method of the various information of the drug molecule and the target protein can adopt the storage method provided in any of the above-mentioned method embodiments. Correspondingly, the method of obtaining various information of the drug molecule and the target protein is the same as For the storage method, refer to the relevant content in the foregoing method embodiments for details, and details are not repeated here.

Step S802. Determine the characterization information of the target protein and the characterization information of the drug molecule. The characterization information of the target protein includes the structural characterization of the target protein, the characterization of the protein family group and the description of the protein function. The characterization information of the drug molecule includes the drug molecule structural characterization.

The specific implementation manner of this step is similar to the implementation manner in the above-mentioned step S202 or S402. For details, refer to the relevant content of the above-mentioned step S202 or S402, which will not be repeated here.

Step S803: According to the structural characterization of the drug molecule, the structural characterization of the target protein and the characterization of the protein family group, predict the correlation information between the target protein and the drug molecule, and screen out the drug molecules to be screened according to the correlation information Alternative Drug Molecules.

In this step, the target protein is used as a given object, and the drug molecule is used as an object to be screened, which is realized by a method similar to the above-mentioned steps S403-S404. For details, please refer to the relevant content of the above-mentioned steps S403-S404, which will not be repeated here.

Step S804, sort the candidate drug molecules according to the characterization information of the drug molecule and the characterization information of the target protein, and output the multi-drug molecule sequence matching the target protein according to the sorting result.

In this step, the target protein is used as a given object, the drug molecule is used as an object to be screened, and the candidate drug molecule is used as a candidate object, and a method similar to the above steps S405-S409 is used to achieve the obtained drug molecule and target The protein matching result is the multi-drug molecular sequence matching the given target protein. For details, please refer to the relevant content of the above steps S405-S409, which will not be repeated here.

The method provided in this embodiment can be applied to the scene of predicting the applicable drug molecule for a given target protein, and in response to the matching instruction of the target protein applicable drug molecule, the structural information and protein family group of the given target protein can be obtained. The specific information and protein function description information, as well as the structural information of the drug molecule to be screened, and use different pre-training models for the feature information of different modalities to generate the structural representation of the drug molecule, the structural representation of the target protein, and the protein family group Characterization information such as characterization and protein function description characterization; according to the characterization information of multimodal feature information, use the trained molecular target relationship prediction model to predict the correlation information between the target protein and the drug molecule, and screen the given target protein Multiple drug molecules with high correlation are used as candidate drug molecules, so as to realize the coarse screening of drug molecules and greatly reduce the number of drug molecules to be screened. Further, through the established molecular-target relationship knowledge graph, based on the molecular-target relationship knowledge graph, the triplet composed of each candidate drug molecule of the target protein, the target protein and the group of the target protein is characterized , to obtain the triplet characterization corresponding to each target protein and each candidate drug molecule, and to obtain another modal feature information characterization information, according to the multi-modal feature information characterization information, using the trained sequence model, The characterization information of the multimodal feature information is fused to obtain the fusion characterization of each candidate drug molecule, and according to the fusion characterization of each candidate drug molecule, the candidate drug molecules are accurately sorted, and the target is determined according to the precise sorting result The multi-drug molecular sequence matching the protein can quickly and accurately screen out the possible drug molecular sequence that can effectively act on a given target protein, which greatly improves the efficiency of drug screening.

Fig. 9 is a flowchart of a method for matching a drug molecule with a target protein according to another exemplary embodiment of the present application. The method provided in this embodiment is applied to the end-side devices in the above system architecture. As shown in Figure 9, the processing flow of the end-side device is as follows:

Step S901 , in response to the matching request between the drug molecule and the target protein, obtain the information of a given object to be matched and an object to be screened, one of which is a drug molecule and the other is a target protein.

The user can specify the given object for matching and the object to be screened through the operation of the end-to-end device, and send a matching request between the drug molecule and the target protein to the end-to-end device, such as predicting the multi-target protein sequence for a given drug requests, or requests for predicting applicable drug molecules for a given target protein, etc. The request may contain information about a given object to be matched and an object to be filtered. Optionally, the user may also input characteristic information of a given object through the end-side device.

Step S902, sending the information of the given object to be matched and the object to be screened to the server.

Optionally, the matching request between the drug molecule and the target protein may include the identification information of the given object and the identification information of the object to be screened. According to the identification information of the given object and the identification information of the object to be screened, the server can acquire the feature information of the given object and the object to be screened.

Optionally, the matching request between a drug molecule and a target protein may only contain the identification information of a given drug molecule. According to the identification information of the given object, the server can obtain the characteristic information of the given object, and the server obtains the characteristic information of the existing objects to be screened.

Optionally, the drug multi-target protein prediction instruction may also include structural information of a given drug molecule, and the server obtains characteristic information of existing objects to be screened.

In this embodiment, the matching result is determined by the server in the following manner:

According to the structural information of the drug molecule, the structural information of the target protein, the information of the protein family group and the description information of the protein function, the characterization information of the drug molecule and the characterization information of the target protein are determined. The characterization information of the drug molecule includes the structural characterization of the drug molecule , the characterization information of the target protein includes the structural characterization of the target protein, the characterization of the protein family group and the description and characterization of the protein function; according to the structural characterization of the drug molecule, the structural characterization of the target protein and the characterization of the protein family group, predict the target protein Correlation information with the drug molecule, and according to the correlation information, select candidate objects from the objects to be screened, the object to be screened is the target protein or drug molecule; according to the characterization information of the drug molecule and the characterization information of the target protein, The candidate objects are sorted, and the matching results of drug molecules and target proteins are output according to the sorting results.

In addition, the server, based on the information of the given object and the object to be screened, as well as the structural information of the drug molecule, the structural information of the target protein, the information of the protein family group and the description information of the protein function, through the matching results of the drug molecule and the target protein, Specifically, it can be realized by using the method provided by any of the foregoing method embodiments. For specific implementation schemes and technical effects, refer to the foregoing method embodiments, which will not be repeated here.

Step S903, receiving the matching result of the drug molecule and the target protein sent by the server.

Step S904, outputting the matching result of the drug molecule and the target protein.

After receiving the matching result of the drug molecule and the target protein, the end-side device can visualize and output the matching result of the drug molecule and the target protein, so that the information of the multi-target object sequence matching the given object, and the target Object-related information (such as structure information, identification information) and other information are displayed visually, thereby providing reference data to users.

For example, in the scenario of predicting a multi-target protein sequence for a given drug, the matching result is the multi-target protein sequence that matches the given drug molecule, and the relevant information of each target protein in the multi-target protein sequence is output through visualization , to provide reference data to relevant drug developers to select druggable drug molecules from the perspective of drug design through manual interactive screening, and to derivate drug structures for target proteins of specific diseases, so as to put them into subsequent drug development links.

In this embodiment, the end-side device may be a terminal, a client, an application program (APP), a software product, or a database that can interact with the server. Here, the specific product form of the end-side device and the form of interaction with the server Not specifically limited.

Fig. 10 is a schematic structural diagram of a device for matching drug molecules and target proteins according to an exemplary embodiment of the present application. The device for matching drug molecules and target proteins provided in the embodiments of the present application can execute the processing procedures provided in the embodiments of the methods for matching drug molecules and target proteins. As shown in FIG. 10 , the device 100 for matching drug molecules and target proteins includes: a multimodal information acquisition module 101 , a multimodal characterization module 102 , a rough screening module 103 and a fine screening module 104 .

Among them, the multimodal information acquisition module 101 is used to acquire the structural information of the drug molecule to be matched, the structural information of the target protein, the protein family group information and the protein function description information.

The multimodal characterization module 102 is used to determine the characterization information of the drug molecule and the characterization information of the target protein. The characterization information of the drug molecule includes the structural characterization of the drug molecule, and the characterization information of the target protein includes the structural characterization of the target protein, protein family Group characterization and protein function characterization.

The rough screening module 103 is used to predict the correlation information between the target protein and the drug molecule according to the structural characterization of the drug molecule, the structural characterization of the target protein and the characterization of the protein family group, and to screen the target protein from the objects to be screened according to the correlation information Candidate objects are selected, and the objects to be screened are target proteins or drug molecules.

The fine screening module 104 is used to sort the candidate objects according to the characterization information of the drug molecule and the characterization information of the target protein, and output the matching result of the drug molecule and the target protein according to the sorting result.

In an optional embodiment, the target protein is the object to be screened, the drug molecule is the given object, the candidate object is the candidate target protein, and the matching result includes the target protein matching the drug molecule.

In an optional embodiment, the drug molecule is used as the object to be screened, the target protein is used as the given object, the candidate object is the candidate drug molecule, and the matching result includes the drug molecule matching the target protein.

In an optional embodiment, when predicting the correlation information between the target protein and the drug molecule based on the structural characterization of the drug molecule, the structural characterization of the target protein, and the characterization of the protein family group, the coarse screening module 103 is also used to :

The structural characterization of the drug molecule, the structural characterization of the target protein, and the characterization of the protein family group are input into the molecular-target relationship prediction model, and the correlation information between the target protein and the drug molecule is predicted through the molecular-target relationship prediction model.

In an optional embodiment, when the candidate objects are sorted according to the characterization information of the drug molecule and the characterization information of the target protein, the fine screening module 104 is also used to: , characterize the triplet containing the given object and the candidate object, and obtain the triplet representation of the given object and the candidate object, and the molecular target relationship knowledge map contains the known drug molecule and the known target protein The interaction relationship, and the protein family group information of the known target protein, the triplet includes the drug molecule, the target protein and the protein family group of the target protein; the characterization information of the given object, the characterization of the candidate object The information and the triplet representation of the given object and the candidate object are input into the sequence model, and the representation information of the given object, the representation information of the candidate object, and the triplet of the given object and the candidate object are input into the sequence model through the encoding module of the sequence model. The group representations are fused to obtain the fusion representation of the candidate objects; through the sorting module of the sequence model, the candidate objects are sorted according to the fusion representation of the candidate objects to obtain the first ranking result.

In an optional embodiment, when the candidate objects are screened out according to the correlation information, the coarse screening module 103 is also used to: sort the objects to be screened according to the correlation information with the given object to obtain the second sorting Result: determining a second preset number of objects to be screened according to the second sorting result as a plurality of candidate objects related to the given object.

In an optional embodiment, when outputting the matching result of the drug molecule and the target protein according to the sorting result, the fine screening module 104 is also used to: determine a first preset number of candidates according to the first sorting result, As the target object matched with the given object, the matching result of the drug molecule and the target protein is obtained, and the information of the target object matched with the given object is output. Wherein, the second preset number is greater than the first preset number.

In an optional embodiment, the device 100 for matching drug molecules and target proteins further includes: a first model training module, configured to: obtain a pre-established knowledge map of molecular-target relationship, the knowledge map of molecular-target relationship contains known The interaction relationship between the drug molecule and the target protein, as well as the protein family group information of the target protein; according to the knowledge map of the molecular target relationship, a comparative learning training set is generated by sampling, and the samples in the comparative learning training set include drug molecules, drug molecules Corresponding to the target protein and the triplet of the protein family group information of the target protein, the comparative learning training set contains positive samples and negative samples; according to the comparative learning training set, the molecular target relationship prediction model is subjected to comparative learning training to obtain the training Good molecular target relationship prediction model.

In an optional embodiment, when realizing sampling and generating a comparative learning training set according to the molecular-target relationship knowledge graph, the first model training module is also used to: sample drug molecules from the molecular-target relationship knowledge graph, and the The target protein with interaction relationship with the drug molecule and the protein family group information of the target protein form a positive sample; from the molecular target relationship knowledge map, sample the drug molecule and the target protein with no interaction relationship with the drug molecule. The target protein and the protein family group information of the target protein constitute a negative sample; positive samples and negative samples constitute a comparative learning training set.

In an optional embodiment, in realizing the molecular-target relationship knowledge map, sampling drug molecules, target proteins that do not have an interaction relationship with the drug molecule, and protein family group information of the target protein form a negative When sampling, the first model training module is also used to: sample the drug molecule, the first target protein that has an interaction relationship with the drug molecule, and the protein family group of the first target protein from the molecular-target relationship knowledge map According to the protein family group information of the first target protein, the second target protein with the same group as the first target protein is sampled from the molecular target relationship knowledge map; the drug molecule, the second target protein The protein family group information of the second target protein and the second target protein constitutes a negative sample.

In an optional embodiment, the device 100 for matching drug molecules and target proteins further includes: a second model training module, configured to: sample a given object sample corresponding to a given object sample from the molecular-target relationship knowledge map Use the trained molecular target relationship prediction model to predict the correlation information between the given object sample and the object sample to be screened according to the representation information of the given object sample and the corresponding representation information of the object sample to be screened , and select multiple candidate object samples related to the given object sample from the object samples to be screened according to the correlation information; The group is characterized to obtain the triplet representation of the given object sample and the candidate object sample. The triplet includes the drug molecule, the target protein and the protein family group of the target protein; the characterization information of any given object sample, The characterization information of multiple candidate object samples of a given object sample and the triplet characterization of the given object sample and the candidate object samples constitute a training sample; obtain the matching result annotation information of the training sample, the training sample and the training sample The matching result annotation information constitutes a training set; the training sequence model is trained using the training set to obtain a trained sequence model.

In an optional embodiment, when using the training set to train the sequence model to obtain the trained sequence model, the second model training module is also used for: based on the training set, the representation information of the given object sample in the training sample, The characterization information of multiple candidate object samples of a given object sample and the triplet characterization of the given object sample and the candidate object samples are input into the sequence model, and the characterization information of the given object sample, The characterization information of the candidate object samples and the triplet representations of the given object samples and the candidate object samples are fused to obtain the fusion representation of the candidate object samples; through the sorting module of the sequence model, according to the fusion representation of the candidate object samples, the The candidate object samples are sorted, and the matching prediction results are determined; the loss function value is calculated according to the matching prediction result and the matching result of the training sample, and the model parameters of the encoding module and the sorting module are optimized according to the loss function value to obtain a good training result. sequence model.

The device provided in the embodiment of the present application can be specifically used to execute the method provided in any one of the above method embodiments, and the specific functions and technical effects achieved will not be repeated here.

Fig. 11 is a schematic structural diagram of a server provided by an example embodiment of the present application. As shown in FIG. 11 , the server 110 includes: a processor 1101 , and a memory 1102 communicatively connected to the processor 1101 , and the memory 1102 stores computer-executable instructions. Wherein, the processor executes the computer-executed instructions stored in the memory to implement the solution provided by any of the above method embodiments, and the specific functions and technical effects that can be achieved will not be repeated here.

The embodiment of the present application also provides a computer-readable storage medium, in which computer-executable instructions are stored, and when the computer-executable instructions are executed by a processor, they are used to implement the solutions and specific functions provided by any of the above-mentioned method embodiments And the technical effects that can be achieved will not be repeated here.

The embodiment of the present application also provides a computer program product, the computer program product includes: a computer program, the computer program is stored in a readable storage medium, at least one processor of the electronic device can read the computer program from the readable storage medium, at least A processor executes the computer program so that the electronic device executes the solution provided by any of the above method embodiments, and the specific functions and technical effects that can be achieved are not repeated here.

In addition, in some of the processes described in the above embodiments and accompanying drawings, multiple operations appearing in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear herein or executed in parallel , is only used to distinguish different operations, and the serial number itself does not represent any execution order. Additionally, these processes can include more or fewer operations, and these operations can be performed sequentially or in parallel. It should be noted that the descriptions of "first" and "second" in this article are used to distinguish different messages, devices, modules, etc. are different types. "Multiple" means two or more, unless otherwise clearly and specifically defined.

Other embodiments of the present application will be readily apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any modification, use or adaptation of the application, these modifications, uses or adaptations follow the general principles of the application and include common knowledge or conventional technical means in the technical field not disclosed in the application . The specification and examples are to be considered exemplary only, with a true scope and spirit of the application indicated by the following claims.

It should be understood that the present application is not limited to the precise constructions which have been described above and shown in the accompanying drawings, and various modifications and changes may be made without departing from the scope thereof. The scope of the application is limited only by the appended claims.

Claims (14)

1. A method for matching drug molecules and target proteins, comprising: Obtain the structural information of the drug molecule to be matched, the structural information of the target protein, the protein family group information and the protein function description information; Determine the characterization information of the drug molecule and the characterization information of the target protein, the characterization information of the drug molecule includes the structural characterization of the drug molecule, and the characterization information of the target protein includes the structural characterization of the target protein, protein family Group characterization and protein function description characterization; According to the structural characterization of the drug molecule, the structural characterization of the target protein, and the characterization of protein family groups, predict the correlation information between the target protein and the drug molecule, and based on the correlation information, from the An alternative object is selected from the screening object, and the object to be screened is the target protein or the drug molecule; According to the characterization information of the drug molecule and the characterization information of the target protein, the candidate objects are sorted, and a matching result between the drug molecule and the target protein is output according to the sorting result. 2. The method of claim 1, wherein, The target protein is an object to be screened, the drug molecule is a given object, the candidate object is an alternative target protein, and the matching result includes a target protein that matches the drug molecule; or, The drug molecule is an object to be screened, the target protein is a given object, the candidate object is an alternative drug molecule, and the matching result includes a drug molecule matching the target protein. 3. The method according to claim 2, characterized in that, according to the structural characterization of the drug molecule, the structural characterization of the target protein and the characterization of protein family groups, predict the relationship between the target protein and the Related information on drug molecules, including: The structural characterization of the drug molecule, the structural characterization of the target protein and the characterization of the protein family group are input into the molecular target relationship prediction model, and the molecular target relationship prediction model is used to predict the relationship between the target protein and the drug molecule related information. 4. The method according to claim 2, wherein the sorting of the candidate objects according to the characterization information of the drug molecule and the characterization information of the target protein comprises: According to the pre-established molecular target relationship knowledge map, characterize the triplets containing the given object and the candidate object, and obtain the triplet representation of the given object and the candidate object, so The molecular-target relationship knowledge graph includes the interaction relationship between known drug molecules and known target proteins, as well as the protein family group information of known target proteins. The triplet includes drug molecules, target proteins and The protein family group of the target protein; Inputting the characterization information of the given object, the characterization information of the candidate object, and the triplet characterization of the given object and the candidate object into the sequence model, and through the coding module of the sequence model, the The characterization information of the given object, the characterization information of the candidate object, and the triplet characterization of the given object and the candidate object are fused to obtain the fusion characterization of the candidate object; Using the sorting module of the sequence model, the candidate objects are sorted according to the fusion representation of the candidate objects to obtain a first sorting result. 5. The method according to claim 4, wherein the filtering out candidate objects according to the correlation information comprises: According to the correlation information with the given object, sort the objects to be screened to obtain a second sorting result; determine a second preset number of objects to be screened according to the second sorting result as the objects related to the given object related multiple candidates; The matching result of outputting the drug molecule and the target protein according to the sorting result includes: According to the first sorting result, determine a first preset number of candidate objects as target objects matching the given object, obtain a matching result between the drug molecule and the target protein, and output the Information about the target object matching the given object; Wherein, the second preset number is greater than the first preset number. 6. The method according to claim 4, further comprising: Obtain a pre-established molecular-target relationship knowledge graph, which contains known interaction relationships between drug molecules and target proteins, as well as protein family group information of target proteins; According to the molecular target relationship knowledge map, sampling is generated to generate a comparative learning training set, and the samples in the comparative learning training set are triplets containing drug molecules, target proteins corresponding to drug molecules, and protein family group information of target proteins, The contrastive learning training set includes positive samples and negative samples; According to the comparative learning training set, the molecular target relationship prediction model is subjected to comparative learning training to obtain a trained molecular target relationship prediction model. 7. The method according to claim 6, characterized in that, according to the molecular target relationship knowledge map, sampling and generating a comparative learning training set includes: From the molecular-target relationship knowledge map, sample the drug molecule, the target protein that has an interaction relationship with the drug molecule, and the protein family group information of the target protein to form a positive sample; From the molecular-target relationship knowledge map, sample drug molecules, target proteins that do not have an interaction relationship with the drug molecules, and protein family group information of the target proteins to form a negative sample; The positive samples and the negative samples constitute a contrastive learning training set. 8. The method according to claim 7, characterized in that, from the molecular-target relationship knowledge map, sampling drug molecules, target proteins that do not have an interaction relationship with the drug molecules, and the target protein The protein family group information to form a negative sample, including: From the molecular-target relationship knowledge graph, sampling drug molecules, first target proteins that have an interaction relationship with the drug molecules, and protein family group information of the first target proteins; According to the protein family group information of the first target protein, sampling a second target protein having the same group as the first target protein from the molecular target relationship knowledge map; The drug molecule, the second target protein, and the protein family group information of the second target protein are used to form a negative sample. 9. The method according to any one of claims 4-8, further comprising: Sampling a given object sample and an object sample to be screened corresponding to the given object sample from the molecular target relationship knowledge map; Using the trained molecular target relationship prediction model, according to the characterization information of the given object sample and the corresponding characterization information of the object sample to be screened, predict the correlation information between the given object sample and the object sample to be screened, and Screening out a plurality of candidate object samples related to the given object sample from the object samples to be screened according to the correlation information; According to the molecular target relationship knowledge map, characterize the triplet containing the given object sample and the candidate object sample, and obtain the triplet of the given object sample and the candidate object sample Characterization, the triad includes a drug molecule, a target protein, and a protein family group of the target protein; The characterization information of any given object sample, the characterization information of multiple candidate object samples of the given object sample, and the triplet characterization of the given object sample and the candidate object samples constitute a Training samples; Acquiring the matching result annotation information of the training samples, the training samples and the matching result annotation information of the training samples constitute a training set; Using the training set to train the sequence model to obtain a trained sequence model. 10. The method according to claim 9, wherein said using said training set to train said sequence model to obtain a trained sequence model comprises: Based on the training set, the characterization information of the given object sample in the training samples, the characterization information of multiple candidate object samples of the given object sample, and the given object sample and the candidate The triplet characterization of the object sample is input into the sequence model, and the characterization information of the given object sample, the characterization information of the candidate object sample, and the given object sample are combined with the sequence model through the coding module of the sequence model The triplet representation fusion of the candidate object sample is obtained to obtain the fusion representation of the candidate object sample; sorting the candidate object samples according to the fusion representation of the candidate object samples through the sorting module of the sequence model, and determining matching prediction results; Calculate the loss function value according to the matching prediction result and the matching result annotation information of the training sample, and optimize the model parameters of the encoding module and the sorting module according to the loss function value to obtain a trained sequence model . 11. A method for matching drug molecules and target proteins, comprising: In response to the drug multi-target protein prediction instruction, obtain the structural information of the given drug molecule, as well as the structural information, protein family group information and protein function description information of the target protein to be screened; Determine the characterization information of the drug molecule and the characterization information of the target protein, the characterization information of the drug molecule includes the structural characterization of the drug molecule, and the characterization information of the target protein includes the structural characterization of the target protein, protein family Group characterization and protein function description characterization; According to the structural characterization of the drug molecule, the structural characterization of the target protein, and the characterization of protein family groups, predict the correlation information between the target protein and the drug molecule, and based on the correlation information, from the Alternative target proteins were screened out from the screened target proteins; According to the characterization information of the drug molecule and the characterization information of the target protein, the candidate target proteins are sorted, and a multi-target protein sequence matching the drug molecule is output according to the sorting result. 12. A method for matching a drug molecule with a target protein, comprising: In response to the matching instruction of the drug molecule applicable to the target protein, obtain the structural information, protein family group information and protein function description information of the given target protein, as well as the structural information of the drug molecule to be screened; Determine the characterization information of the target protein and the characterization information of the drug molecule, the characterization information of the target protein includes the structural characterization of the target protein, the characterization of the protein family group and the description and characterization of the protein function, the characterization of the drug molecule Characterization information includes structural characterization of drug molecules; According to the structural characterization of the drug molecule, the structural characterization of the target protein, and the characterization of protein family groups, predict the correlation information between the target protein and the drug molecule, and based on the correlation information, from the Screening out candidate drug molecules from the screened drug molecules; According to the characterization information of the drug molecule and the characterization information of the target protein, the candidate drug molecules are sorted, and a multi-drug molecule sequence matching the target protein is output according to the sorting result. 13. A method for matching drug molecules and target proteins, characterized in that it is applied to end-to-end devices, including: Responding to a matching request between a drug molecule and a target protein, obtaining information on a given object to be matched and an object to be screened, wherein one of the given object and the object to be screened is a drug molecule and the other is a target protein; Sending the information of the given object to be matched and the object to be screened to the server, wherein the matching result is the information of the given object to be matched and the object to be screened by the server according to claim 1 - Determined by the method described in any one of 12; receiving the matching result of the drug molecule and the target protein sent by the server; Output the matching result of the drug molecule and the target protein. 14. A server, comprising: a processor, and a memory communicatively connected to the processor; the memory stores computer-executable instructions; The processor executes the computer-implemented instructions stored in the memory to implement the method according to any one of claims 1-12.

Images