Unsupervised prediction method and device for molecular drug-like properties based on two-dimensional molecular graph

Abstract

The present application proposes a method and device for unsupervised prediction of molecular drug-likeness based on two-dimensional molecular graphs, wherein the method comprises: first, on a database ChEMBL including a large number of drug-like and non-drug-like molecules, a teacher model is trained through pre-training tasks of atom mask prediction and chemical bond mask prediction, so that it learns the characteristics of molecular topological structure that are highly correlated with drug-likeness; then, a student model with the same structure as the teacher model is distilled out only on drug data; then, an arbitrary molecule is input into the teacher model and the student model respectively, and the difference between the feature representations output by the two models is used as the drug-likeness score of the arbitrary molecule; finally, the drug-likeness of the arbitrary molecule is predicted according to the drug-likeness score to obtain a prediction result; thus, the trained teacher model and student model can directly capture the molecular features that are highly correlated with drug-likeness, while greatly alleviating the length deviation problem of the one-dimensional sequence model.

Description

Technical Field

[0001] The present application relates to the technical field of bioinformatics, and in particular to a method for unsupervised prediction of molecular drug-like properties based on two-dimensional molecular graphs, a computer-readable storage medium, a computer device, and an apparatus for unsupervised prediction of molecular drug-like properties based on two-dimensional molecular graphs. Background Art

[0002] In the related technology, the existing molecular drug-likeness prediction methods are mainly divided into three categories, among which the first category is the molecular drug-likeness prediction based on manual rules: it is represented by quantitative estimate of drug-likeness (QED) and rule of five (RO5). Drug scientists manually summarize some rules based on existing drugs as the basis for judging or scoring molecular drug-likeness; however, the defect of this method is that the rules are too harsh, it is easy to directly screen out some potential molecules, and only cover the characteristics of drug molecules before the proposed era. The second category is supervised molecular drug-likeness prediction: in order to overcome the defects of manual rules, a method of supervised molecular drug-likeness detection in a data-driven form is proposed; this type of method usually collects drug molecules as positive examples and other molecules as negative examples, and trains a binary classifier to judge whether the molecules are drug-like or not; however, the defect of this type of method is that non-drug molecules that can be used as negative examples are inexhaustible, so the data distribution learned by the trained model is seriously dependent on the selected training data, and can only be generalized to specific types of negative example molecules. The third category is the unsupervised prediction of drug-likeness of molecules: Currently, the only unsupervised method for predicting drug-likeness of molecules is based on the Recurrent Neural Network (RNN) model, which takes the simplified molecular-input line-entry system (SMILES) sequence of the molecule as input. Because this method is based on the one-dimensional sequence representation of the molecule, it does not directly capture the topological structure of the two-dimensional molecular graph, resulting in a serious length deviation in the drug-likeness prediction indicator generated. Summary of the invention

[0003] The present application aims to solve at least one of the technical problems in the above-mentioned technology to a certain extent. To this end, one purpose of the present application is to propose an unsupervised prediction method for molecular drug-likeness based on a two-dimensional molecular graph, by taking the structure of the two-dimensional molecular graph as input to train a teacher model that can learn the molecular topological structure characteristics, and distilling the teacher model only on the drug molecule data to obtain a student model, thereby, through the trained teacher model and student model, molecular features that are highly correlated with drug-likeness can be directly captured, and the length deviation problem is greatly alleviated.

[0004] A second object of the present application is to provide a computer-readable storage medium.

[0005] The third objective of the present application is to provide a computer device.

[0006] The fourth objective of the present application is to propose an unsupervised prediction device for molecular drug-like properties based on two-dimensional molecular graphs.

[0007] To achieve the above-mentioned objectives, the first aspect of the present application proposes an unsupervised prediction method for molecular drug-likeness based on two-dimensional molecular graphs, the method comprising the following steps: obtaining a first training data set, wherein the first training data set includes two-dimensional molecular graphs of drug-like molecules and two-dimensional molecular graphs of non-drug-like molecules; inputting the first training data set into a pre-constructed teacher model, and pre-training the teacher model using mask prediction to obtain a pre-trained teacher model; obtaining a second training data set, wherein the second training data set only includes two-dimensional molecular graphs of drug molecules; inputting the second training data set into the pre-trained teacher model to obtain a first atomic representation, and inputting the second training data set into a pre-constructed student model with the same structure as the teacher model to obtain a second Atomic representation, and minimize the distance between the first atomic representation and the second atomic representation to update the parameters of the student model to obtain a trained student model; obtain a two-dimensional molecular graph of the molecule to be predicted, input the two-dimensional molecular graph of the molecule to be predicted into the pre-trained teacher model to obtain a third atomic representation, and input the two-dimensional molecular graph of the molecule to be predicted into the trained student model to obtain a fourth atomic representation, and calculate the drug-likeness score of the molecule to be predicted according to the third atomic representation and the fourth atomic representation; predict the drug-likeness of the molecule to be predicted according to the drug-likeness score to obtain a prediction result; thus, the molecular features that are highly correlated with drug-likeness can be directly captured by the trained teacher model and the student model, thereby greatly alleviating the length deviation problem.

[0008] In addition, the unsupervised prediction method of molecular drug properties based on two-dimensional molecular graphs proposed in the above embodiments of the present application may also have the following additional technical features:

[0009] Optionally, the teacher model is pre-trained using mask prediction to obtain a pre-trained teacher model, including: mapping the atoms and / or chemical bonds corresponding to each two-dimensional molecular graph in the first training data set to corresponding types in RDKit; randomly selecting some atoms and / or chemical bonds from all mapped atoms and / or chemical bonds, and replacing the types of the some atoms and / or chemical bonds with mask symbols to perform mask operations; inputting the masked molecular data into the teacher model for encoding to obtain the feature representation of each mask, and predicting the original type of the feature representation through a classifier.

[0010] Optionally, the student model is trained according to the following loss function:

[0011]

[0012]

[0013]

[0014] in, represents the total loss function, represents the Euclidean distance between the representations learned by the teacher model and the student model, represents the cosine distance between the representations learned by the teacher model and the student model, represents the feature representation of the i-th atom output by the teacher model, represents the feature representation of the i-th atom output by the student model, |B| represents the number of samples in the current training batch, and N represents the number of atoms in the molecule.

[0015] Optionally, the drug-likeness score of the molecule to be predicted is calculated according to the following formula:

[0016]

[0017] in, Indicates the drug-likeness score.

[0018] To achieve the above-mentioned objectives, the second aspect embodiment of the present application proposes a computer-readable storage medium, on which is stored a program for unsupervised prediction of molecular drug-like properties based on a two-dimensional molecular graph. When the program for unsupervised prediction of molecular drug-like properties based on a two-dimensional molecular graph is executed by a processor, the method for unsupervised prediction of molecular drug-like properties based on a two-dimensional molecular graph as described above is implemented.

[0019] According to the computer-readable storage medium of the embodiment of the present application, by storing an unsupervised prediction program for molecular drug-likeness based on a two-dimensional molecular graph, the unsupervised prediction method for molecular drug-likeness based on a two-dimensional molecular graph is implemented when the unsupervised prediction program for molecular drug-likeness based on a two-dimensional molecular graph is executed by a processor. As a result, molecular features that are highly correlated with drug-likeness can be directly captured through the trained teacher model and student model, thereby greatly alleviating the length deviation problem.

[0020] To achieve the above objectives, the third aspect of the present application proposes a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned unsupervised prediction method of molecular drug-like properties based on two-dimensional molecular graphs.

[0021] According to the computer device of the embodiment of the present application, the unsupervised prediction program of the molecular drug-likeness based on the two-dimensional molecular graph is stored in the memory, so that when the unsupervised prediction program of the molecular drug-likeness based on the two-dimensional molecular graph is executed by the processor, the above-mentioned unsupervised prediction method of the molecular drug-likeness based on the two-dimensional molecular graph is implemented. Therefore, the molecular features that are highly correlated with drug-likeness can be directly captured through the trained teacher model and student model, and the length deviation problem is greatly alleviated.

[0022] To achieve the above-mentioned objectives, the fourth aspect of the present application proposes an unsupervised prediction device for molecular drug-likeness based on two-dimensional molecular graphs, comprising: a first acquisition module, used to acquire a first training data set, wherein the first training data set includes two-dimensional molecular graphs of drug-like molecules and two-dimensional molecular graphs of non-drug-like molecules; a first training module, used to input the first training data set into a pre-constructed teacher model, and pre-train the teacher model using mask prediction to obtain a pre-trained teacher model; a second acquisition module, used to acquire a second training data set, wherein the second training data set only includes two-dimensional molecular graphs of drug molecules; a second training module, used to input the second training data set into the pre-trained teacher model to obtain a first atomic representation, and input the second training data set into a pre-constructed student model with the same structure as the teacher model. model to obtain a second atomic representation, and minimize the distance between the first atomic representation and the second atomic representation to update the parameters of the student model to obtain a trained student model; a calculation module, used to obtain a two-dimensional molecular graph of the molecule to be predicted, input the two-dimensional molecular graph of the molecule to be predicted into the pre-trained teacher model to obtain a third atomic representation, and input the two-dimensional molecular graph of the molecule to be predicted into the trained student model to obtain a fourth atomic representation, and calculate the drug-likeness score of the molecule to be predicted according to the third atomic representation and the fourth atomic representation; a drug-likeness prediction module, used to predict the drug-likeness of the molecule to be predicted according to the drug-likeness score to obtain a prediction result; thereby, the molecular features that are highly correlated with drug-likeness can be directly captured through the trained teacher model and the student model, while greatly alleviating the length deviation problem.

[0023] In addition, the unsupervised prediction device for molecular drug properties based on two-dimensional molecular graphs proposed in the above embodiments of the present application may also have the following additional technical features:

[0024] Optionally, the first training module is also used to map the atoms and / or chemical bonds corresponding to each two-dimensional molecular graph in the first training data set into corresponding types in RDKit; randomly select some atoms and / or chemical bonds from all mapped atoms and / or chemical bonds, and replace the types of the some atoms and / or chemical bonds with mask symbols to perform mask operations; input the masked molecular data into the teacher model for encoding to obtain the feature representation of each mask, and predict the original type of the feature representation through a classifier.

[0025] Optionally, the student model is trained according to the following loss function:

[0026]

[0027]

[0028]

[0029] in, represents the total loss function, represents the Euclidean distance between the representations learned by the teacher model and the student model, represents the cosine distance between the representations learned by the teacher model and the student model, represents the feature representation of the i-th atom output by the teacher model, represents the feature representation of the i-th atom output by the student model, |B| represents the number of samples in the current training batch, and N represents the number of atoms in the molecule.

[0030] Optionally, the drug-likeness score of the molecule to be predicted is calculated according to the following formula:

[0031]

[0032] in, Indicates the drug-likeness score. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of the process of unsupervised prediction method of molecular drug-like properties based on two-dimensional molecular graph according to an embodiment of the present application;

[0034] Figure 2 A schematic diagram of an overall network structure according to an embodiment of the present application;

[0035] Figure 3 It is a block diagram of an unsupervised prediction device for molecular drug-like properties based on a two-dimensional molecular graph according to an embodiment of the present application. DETAILED DESCRIPTION

[0036] Embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0037] In order to better understand the above technical solution, exemplary embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0038] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0039] like Figure 2 As shown, the method of the present application is mainly divided into three stages: (1) The first stage: pre-training the teacher model. In this stage, two pre-training tasks are designed to learn common topological structure features on a large amount of 2D molecular data; (2) The second stage: distilling the student model. In this stage, the output of the student model and the teacher model are brought closer on the drug molecules, and the student model only learns how to fit the data distribution of the drug molecules; (3) The third stage: drug-likeness detection. In this stage, the molecule is encoded by using the teacher model and the student model at the same time, and the output difference of the two models is used as the drug-likeness score of the molecule.

[0040] It should be noted that in the method of the present application, molecular data is represented by a 2D graph; the atoms in the molecule are used as nodes on the graph, the chemical bonds are used as edges on the graph, and the various properties of the atoms and chemical bonds are calculated as the initialization features of the graph. In addition, the structure of the teacher model and the student model is consistent, and can be any encoder model that can encode a 2D graph; the present application adopts the Graph Transformer model to obtain the node representation and edge representation of the molecular graph, wherein the Graph Transformer model is a deep learning model based on graph structured data, which combines the advantages of Graph Neural Networks (GNNs) and Transformer architecture.

[0041] Specifically, Figure 1 FIG. 1 is a flow chart of an unsupervised prediction method of a molecular drug-like property based on a two-dimensional molecular graph according to an embodiment of the present application. Figure 1 As shown, the unsupervised prediction method of molecular drug properties based on a two-dimensional molecular graph in the embodiment of the present application includes the following steps:

[0042] S101, obtaining a first training data set, wherein the first training data set includes two-dimensional molecular graphs of drug-like molecules and two-dimensional molecular graphs of non-drug-like molecules.

[0043] It should be noted that the first training data set is collected through the ChEMBL database, which includes a large amount of small molecule data. The ChEMBL database is a public database of biologically active molecules.

[0044] S102, inputting the first training data set into a pre-built teacher model, and pre-training the teacher model using mask prediction to obtain a pre-trained teacher model.

[0045] That is to say, in the first stage, a teacher model is pre-trained so that the general molecular topological structure features can be learned through the trained teacher model.

[0046] As an embodiment, mask prediction is used to pre-train the teacher model to obtain a pre-trained teacher model, including: mapping the atoms and / or chemical bonds corresponding to each two-dimensional molecular graph in the first training data set to the corresponding types in RDKit; randomly selecting some atoms and / or chemical bonds from all mapped atoms and / or chemical bonds, and replacing the types of some atoms and / or chemical bonds with mask symbols to perform mask operations; inputting the masked molecular data into the teacher model for encoding to obtain the feature representation of each mask, and predicting the original type of the feature representation through a classifier.

[0047] As a specific embodiment, in the first stage, two pre-training tasks based on 2D molecular graphs are designed: Molecular Atomic Masking (MAM) and Molecular Bond Masking (MBM) to learn general molecular topological structure features; the loss function in this stage is defined as:

[0048]

[0049] in, represents the hyperparameters for balancing the two tasks, represents the loss function of the pre-trained teacher model, represents the loss function for the atomic mask prediction task, Represents the loss function for the bond mask prediction task.

[0050] In the atom mask prediction task, the atoms are first mapped to a total of 119 types defined in RDKit, which is an open source toolkit that provides a series of data structures and algorithms; then some atoms in the molecule are randomly selected according to a certain ratio, and their types are replaced with the mask symbol "[MASK]"; then the masked molecular data is sent to the teacher model for encoding to obtain each masked atom representation; finally, it is classified through a linear layer. The goal of this task is to minimize the cross entropy loss between the classification result and the true atom category, so the loss function is defined as:

[0051]

[0052] in, represents the masked atoms, represents the masked numerator, represents a data set, Represents the dataset The molecules sampled in represents the probability distribution, Represents the i-th atom type of the masked molecule predicted by the model.

[0053] Thus, through this task, the teacher model can infer the masked atom types from the surrounding atoms and chemical bonding features.

[0054] In the chemical bond mask prediction task, each chemical bond is first mapped to a total of four types defined in RDKit, including single bond, double bond, triple bond and aromatic bond; the subsequent process is similar to the atom mask prediction task. After obtaining the masked chemical bond features, another classifier is used to predict the original chemical bond type. Therefore, the loss function of this task is:

[0055]

[0056] in, Represents the i-th bond type of the masked molecule predicted by the model.

[0057] It should be noted that because the atom type and chemical bond type in the same molecule are highly correlated, the type of an atom can be easily inferred from the surrounding chemical bond types, and similarly, the chemical bond type can also be easily inferred from the surrounding atom types. In order to prevent information leakage and increase the difficulty of the training task, the chemical bonds (or atoms) connected to the masked atoms (or chemical bonds) can also be masked.

[0058] S103, obtaining a second training data set, wherein the second training data set only includes two-dimensional molecular graphs of drug molecules.

[0059] That is, in the second stage, only drug molecule data are used to perform knowledge distillation on the teacher model to obtain the student model.

[0060] S104, inputting the second training data set into the pre-trained teacher model to obtain a first atomic representation, and inputting the second training data set into a pre-built student model with the same structure as the teacher model to obtain a second atomic representation, and minimizing the distance between the first atomic representation and the second atomic representation to update the parameters of the student model to obtain a trained student model.

[0061] As a specific embodiment, in the second stage, the parameters of the teacher model obtained from the first stage training are frozen, and a model with the same structure is randomly initialized as the student model; the drug molecule obtains two atomic representations through the two models, and the parameters of the student model are updated by minimizing the atomic representation distance output by the top layer; the features output by the top layer of the model reflect the semantic information of the molecule and are more related to the abstract property of drug-likeness. The loss function for training the student model is defined as: .

[0062] Specifically, represents the total loss function, represents the Euclidean distance between the representations learned by the teacher model and the student model, Represents the cosine distance between the representations learned by the teacher model and the student model; it should be noted that the reason for using these two distance functions is that the two distances measure the similarity between the outputs of the teacher model and the student model from different perspectives, that is, the cosine distance represents the similarity in direction, while the Euclidean distance represents the similarity in magnitude; in this way, the student model can be more effectively guided to imitate the representation learning of the teacher model.

[0063] More specifically, It is defined as the L2 norm between the top-level outputs of the teacher model and the student model, and is used to measure the direct difference between the two models:

[0064]

[0065] in, represents the feature representation of the i-th atom output by the teacher model, represents the feature representation of the i-th atom output by the student model, |B| represents the number of samples in the current training batch, and N represents the number of atoms in the molecule.

[0066] Because this loss function averages the number of atoms, the drug-likeness scores obtained have almost no length bias. In addition, in order to ensure that the student model can also fit the teacher model in terms of feature space, Defined as:

[0067]

[0068] S105, obtaining a two-dimensional molecular graph of the molecule to be predicted, inputting the two-dimensional molecular graph of the molecule to be predicted into a pre-trained teacher model to obtain a third atomic representation, and inputting the two-dimensional molecular graph of the molecule to be predicted into a trained student model to obtain a fourth atomic representation, and calculating the drug-likeness score of the molecule to be predicted based on the third atomic representation and the fourth atomic representation.

[0069] That is to say, in the third stage, when a new molecule is tested for drug-likeness, the new molecule to be tested is input into the teacher model and the student model respectively, and the difference between the output features of the two models is calculated as the drug-likeness score of the current molecule.

[0070] As a specific example, a given molecule is input into the teacher model and the student model respectively, the atomic representations output by the two models are obtained, and then the drug-likeness score is calculated as follows:

[0071]

[0072] Since the teacher model is trained on the ChEMBL database including both drug-like and non-drug-like molecules, while the student model is trained only on real drugs, the outputs of the teacher and student models will be very similar when the input molecules are drugs or close to drug molecule distribution; however, for non-drug-like molecules, their outputs are not guaranteed to be similar. Therefore, the representation gap between their outputs can be used as a means to quantify the drug-likeness of each input molecule.

[0073] S106, predicting the drug-likeness of the molecule to be predicted according to the drug-likeness score to obtain a prediction result.

[0074] As an embodiment, the larger the drug-likeness score, the smaller the drug-likeness possibility of the molecule to be predicted, and vice versa. If the drug-likeness score is smaller, the drug-likeness possibility of the molecule to be predicted is greater. In addition, a threshold value can also be preset through experience. If the drug-likeness score is greater than the preset threshold value, it means that the molecule to be predicted is a non-drug-like molecule. If the drug-likeness score is less than or equal to the preset threshold value, it means that the molecule to be predicted is a drug-like molecule.

[0075] In summary, according to the unsupervised prediction method of molecular drug-likeness based on two-dimensional molecular graphs according to the embodiment of the present application, first, a teacher model is trained through two designed pre-training tasks on the database ChEMBL, which includes a large number of drug-like and non-drug-like molecules; the teacher model learns the characteristics of the molecular topological structure that is highly correlated with drug-likeness. Then, a student model with the same structure is distilled out only on the drug data; since the student model only learns how to fit the output of the teacher model on the distribution of drug molecules, the outputs of the two models will be similar when drug-like molecules are input, and the student model does not fit well when non-drug-like molecules are input, and the output gap between the two models is large. Therefore, a molecule can be input into the two models respectively, and the difference in the feature representation of the output of the two models is used as the drug-likeness score of the molecule. Finally, a series of experiments on the drug-likeness detection benchmark data set demonstrated the effectiveness of the method of the present application.

[0076] In addition, an embodiment of the present application also proposes a computer-readable storage medium, on which is stored a program for unsupervised prediction of molecular drug-like properties based on a two-dimensional molecular graph. When the program for unsupervised prediction of molecular drug-like properties based on a two-dimensional molecular graph is executed by a processor, the method for unsupervised prediction of molecular drug-like properties based on a two-dimensional molecular graph as described above is implemented.

[0077] According to the computer-readable storage medium of the embodiment of the present application, by storing an unsupervised prediction program for molecular drug-likeness based on a two-dimensional molecular graph, the unsupervised prediction method for molecular drug-likeness based on a two-dimensional molecular graph is implemented when the unsupervised prediction program for molecular drug-likeness based on a two-dimensional molecular graph is executed by a processor. As a result, molecular features that are highly correlated with drug-likeness can be directly captured through the trained teacher model and student model, thereby greatly alleviating the length deviation problem.

[0078] In addition, an embodiment of the present application also proposes a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned unsupervised prediction method of molecular drug-like properties based on two-dimensional molecular graphs.

[0079] According to the computer device of the embodiment of the present application, the unsupervised prediction program of the molecular drug-likeness based on the two-dimensional molecular graph is stored in the memory, so that when the unsupervised prediction program of the molecular drug-likeness based on the two-dimensional molecular graph is executed by the processor, the above-mentioned unsupervised prediction method of the molecular drug-likeness based on the two-dimensional molecular graph is implemented. Therefore, the molecular features that are highly correlated with drug-likeness can be directly captured through the trained teacher model and student model, thereby greatly alleviating the length deviation problem.

[0080] In order to implement the above embodiment, the present application embodiment also proposes an unsupervised prediction device for molecular drug properties based on a two-dimensional molecular graph, such as Figure 3 As shown, the unsupervised prediction device for molecular drug-likeness based on two-dimensional molecular graph includes: a first acquisition module 10, a first training module 20, a second acquisition module 30, a second training module 40, a calculation module 50 and a drug-likeness prediction module 60.

[0081] Among them, the first acquisition module 10 is used to acquire a first training data set, and the first training data set includes a two-dimensional molecular graph of a drug-like molecule and a two-dimensional molecular graph of a non-drug-like molecule; the first training module 20 is used to input the first training data set into a pre-constructed teacher model, and pre-train the teacher model using mask prediction to obtain a pre-trained teacher model; the second acquisition module 30 is used to acquire a second training data set, and the second training data set only includes a two-dimensional molecular graph of a drug molecule; the second training module 40 is used to input the second training data set into the pre-trained teacher model to obtain a first atomic representation, and input the second training data set into a pre-constructed structure with the teacher model The same student model is used to obtain a second atomic representation, and the distance between the first atomic representation and the second atomic representation is minimized to update the parameters of the student model to obtain a trained student model; the calculation module 50 is used to obtain a two-dimensional molecular graph of the molecule to be predicted, input the two-dimensional molecular graph of the molecule to be predicted into the pre-trained teacher model to obtain a third atomic representation, and input the two-dimensional molecular graph of the molecule to be predicted into the trained student model to obtain a fourth atomic representation, and calculate the drug-likeness score of the molecule to be predicted according to the third atomic representation and the fourth atomic representation; the drug-likeness prediction module 60 is used to predict the drug-likeness of the molecule to be predicted according to the drug-likeness score to obtain a prediction result.

[0082] As an embodiment, the first training module 20 is also used to map the atoms and / or chemical bonds corresponding to each two-dimensional molecular graph in the first training data set into corresponding types in RDKit; randomly select some atoms and / or chemical bonds from all mapped atoms and / or chemical bonds, and replace the types of some atoms and / or chemical bonds with mask symbols to perform mask operations; input the masked molecular data into the teacher model for encoding to obtain the feature representation of each mask, and predict the original type of the feature representation through the classifier.

[0083] As an example, the student model is trained according to the following loss function:

[0084]

[0085]

[0086]

[0087] in, represents the total loss function, represents the Euclidean distance between the representations learned by the teacher model and the student model, represents the cosine distance between the representations learned by the teacher model and the student model, represents the feature representation of the i-th atom output by the teacher model, represents the feature representation of the i-th atom output by the student model, |B| represents the number of samples in the current training batch, and N represents the number of atoms in the molecule.

[0088] As an example, the drug-likeness score of the molecule to be predicted is calculated according to the following formula:

[0089]

[0090] in, Indicates the drug-likeness score.

[0091] It should be noted that the above description and examples of the unsupervised prediction method of molecular drug properties based on two-dimensional molecular graphs are also applicable to the unsupervised prediction device of molecular drug properties based on two-dimensional molecular graphs in this embodiment, and will not be repeated here.

[0092] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0093] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0094] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0095] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0096] It should be noted that in the claims, any reference signs placed between brackets shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The present application may be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third etc. does not indicate any order. These words may be interpreted as names.

[0097] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0098] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

[0099] In the description of the present application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the feature. In the description of the present application, "plurality" means two or more, unless otherwise clearly and specifically defined.

[0100] In this application, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0101] In the present application, unless otherwise clearly specified and limited, a first feature being “above” or “below” a second feature may mean that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, a first feature being “above”, “above”, and “above” a second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being “below”, “below”, and “below” a second feature may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0102] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representation of the above terms should not be understood as necessarily being directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

[0103] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. An unsupervised prediction method for molecular drug properties based on two-dimensional molecular graphs, characterized in that: The following steps are involved: Acquire a first training data set, wherein the first training data set includes two-dimensional molecular graphs of drug-like molecules and two-dimensional molecular graphs of non-drug-like molecules; Inputting the first training data set into a pre-built teacher model, and pre-training the teacher model using mask prediction to obtain a pre-trained teacher model; Acquire a second training data set, wherein the second training data set only includes two-dimensional molecular graphs of drug molecules; Inputting the second training data set into the pre-trained teacher model to obtain a first atomic representation, and inputting the second training data set into a pre-built student model with the same structure as the teacher model to obtain a second atomic representation, and minimizing the distance between the first atomic representation and the second atomic representation to update the parameters of the student model to obtain a trained student model; Obtaining a two-dimensional molecular graph of the molecule to be predicted, inputting the two-dimensional molecular graph of the molecule to be predicted into the pre-trained teacher model to obtain a third atomic representation, and inputting the two-dimensional molecular graph of the molecule to be predicted into the trained student model to obtain a fourth atomic representation, and calculating the drug-likeness score of the molecule to be predicted according to the gap between the third atomic representation and the fourth atomic representation; Predicting the drug-likeness of the molecule to be predicted according to the drug-likeness score to obtain a prediction result; The teacher model is pre-trained by using mask prediction to obtain a pre-trained teacher model, including: Mapping atoms and / or chemical bonds corresponding to each two-dimensional molecular graph in the first training data set to corresponding types in RDKit; Randomly selecting some atoms and / or chemical bonds from all mapped atoms and / or chemical bonds, and replacing the types of the some atoms and / or chemical bonds with mask symbols to perform a mask operation; The masked molecular data is input into the teacher model for encoding to obtain a feature representation of each mask, and the original type of the feature representation is predicted by a classifier.

2. The unsupervised prediction method of molecular drug properties based on two-dimensional molecular graph according to claim 1, characterized in that: The student model is trained according to the following loss function: in, represents the total loss function, represents the Euclidean distance between the representations learned by the teacher model and the student model, represents the cosine distance between the representations learned by the teacher model and the student model, represents the feature representation of the i-th atom output by the teacher model, represents the feature representation of the i-th atom output by the student model, |B| represents the number of samples in the current training batch, and N represents the number of atoms in the molecule.

3. The unsupervised prediction method of molecular drug properties based on two-dimensional molecular graphs according to claim 2, characterized in that: The drug-likeness score of the molecule to be predicted is calculated according to the following formula: in, Indicates the drug-likeness score.

4. A computer-readable storage medium, characterized in that: An unsupervised prediction program for molecular drug-likeness based on two-dimensional molecular graphs is stored thereon, and when the unsupervised prediction program for molecular drug-likeness based on two-dimensional molecular graphs is executed by a processor, an unsupervised prediction method for molecular drug-likeness based on two-dimensional molecular graphs as described in any one of claims 1-3 is implemented.

5. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the unsupervised prediction method of molecular drug-likeness based on two-dimensional molecular graph as described in any one of claims 1-3.

6. An unsupervised prediction device for molecular drug properties based on two-dimensional molecular graphs, characterized in that: include: A first acquisition module is used to acquire a first training data set, wherein the first training data set includes a two-dimensional molecular graph of drug-like molecules and a two-dimensional molecular graph of non-drug-like molecules; A first training module, used for inputting the first training data set into a pre-built teacher model, and pre-training the teacher model by using mask prediction to obtain a pre-trained teacher model; A second acquisition module, used to acquire a second training data set, wherein the second training data set only includes a two-dimensional molecular graph of drug molecules; A second training module is used to input the second training data set into the pre-trained teacher model to obtain a first atomic representation, and input the second training data set into a pre-built student model with the same structure as the teacher model to obtain a second atomic representation, and minimize the distance between the first atomic representation and the second atomic representation to update the parameters of the student model to obtain a trained student model; a calculation module, for obtaining a two-dimensional molecular graph of the molecule to be predicted, inputting the two-dimensional molecular graph of the molecule to be predicted into the pre-trained teacher model to obtain a third atomic representation, and inputting the two-dimensional molecular graph of the molecule to be predicted into the trained student model to obtain a fourth atomic representation, and calculating the drug-likeness score of the molecule to be predicted according to the gap between the third atomic representation and the fourth atomic representation; A drug-likeness prediction module, used to predict the drug-likeness of the molecule to be predicted according to the drug-likeness score to obtain a prediction result; Wherein, the first training module is also used for: Mapping atoms and / or chemical bonds corresponding to each two-dimensional molecular graph in the first training data set to corresponding types in RDKit; Randomly selecting some atoms and / or chemical bonds from all mapped atoms and / or chemical bonds, and replacing the types of the some atoms and / or chemical bonds with mask symbols to perform a mask operation; The masked molecular data is input into the teacher model for encoding to obtain a feature representation of each mask, and the original type of the feature representation is predicted by a classifier.

7. The unsupervised prediction device for molecular drug properties based on two-dimensional molecular graphs according to claim 6, characterized in that: The student model is trained according to the following loss function: in, represents the total loss function, represents the Euclidean distance between the representations learned by the teacher model and the student model, represents the cosine distance between the representations learned by the teacher model and the student model, represents the feature representation of the i-th atom output by the teacher model, represents the feature representation of the i-th atom output by the student model, |B| represents the number of samples in the current training batch, and N represents the number of atoms in the molecule.

8. The unsupervised prediction device for molecular drug properties based on two-dimensional molecular graphs according to claim 7, characterized in that: The drug-likeness score of the molecule to be predicted is calculated according to the following formula: in, Indicates the drug-likeness score.

Images