Method and device for training molecular property prediction model, electronic equipment and medium

Abstract

The application provides a training method and device of a molecular property prediction model, an electronic device and a medium. The method comprises: masking first non-geometric information of a sample molecule to obtain second non-geometric information; adding noise to first geometric information of the sample molecule to obtain second geometric information; combining the first non-geometric information of an atom that has not undergone masking, the second non-geometric information, the first geometric information of an atom that has not undergone noise addition, and the second geometric information into a pre-training data set; based on the pre-training data set, calling a molecular property prediction model to perform a first training task to obtain a trained molecular property prediction model; based on an original data set of the sample molecule, performing a second training task of molecular property prediction processing on the trained molecular property prediction model to obtain a trained molecular property prediction model. Through the application, the accuracy of the molecular property prediction model in molecular property prediction can be improved through a phased training process.

Description

Technical Field

[0001] This application relates to artificial intelligence technology, and more particularly to a training method, apparatus, electronic device, and medium for a molecular property prediction model. Background Technology

[0002] Molecular property prediction tasks utilize computer models and algorithms to predict various physicochemical properties and biological activities of molecules. Molecular properties may include water solubility, lipophilicity, toxicity, biological activity, and pharmacokinetic characteristics. In practical applications, the molecular data to be predicted may be missing, and the molecular property prediction function trained solely on the molecular property model cannot cover these situations, leading to inaccurate prediction results.

[0003] Currently, there is no good way to improve the accuracy of molecular property prediction in related technologies. Summary of the Invention

[0004] This application provides a training method, apparatus, electronic device, and medium for a molecular property prediction model, which can improve the accuracy of molecular property prediction through a phased training process.

[0005] The technical solution of this application embodiment is implemented as follows:

[0006] This application provides a method for training a molecular property prediction model, the method comprising:

[0007] The first non-geometric information of at least one atom in the sample molecule is masked to obtain the second non-geometric information;

[0008] The first geometric information of at least one atom of the sample molecule is subjected to noise processing to obtain the second geometric information;

[0009] The first non-geometric information of the atoms that have not undergone the masking process, the second non-geometric information, the first geometric information of the atoms that have not undergone the noise-adding process, and the second geometric information are combined into a pre-training data set.

[0010] Based on the pre-trained dataset, the molecular property prediction model is invoked for the first training task to obtain the trained molecular property prediction model.

[0011] Based on the original data set of the sample molecules, a second training task is performed on the trained molecular property prediction model to obtain the trained molecular property prediction model.

[0012] This application provides a training device for a molecular property prediction model, comprising:

[0013] The sample acquisition module is used to mask the first non-geometric information of at least one atom of the sample molecule to obtain the second non-geometric information.

[0014] The sample acquisition module is further configured to add noise to the first geometric information of at least one atom of the sample molecule to obtain the second geometric information;

[0015] The sample acquisition module is further configured to combine the first non-geometric information, the second non-geometric information, the first geometric information, and the second geometric information of atoms that have not undergone the masking process into a pre-training data set;

[0016] The first training module is used to call the molecular property prediction model to perform a first training task based on the pre-training data set, so as to obtain the trained molecular property prediction model.

[0017] The second training module is used to perform a second training task on the trained molecular property prediction model based on the original data set of the sample molecules, so as to obtain the trained molecular property prediction model.

[0018] This application provides an electronic device, the electronic device comprising:

[0019] Memory is used to store executable instructions or computer programs.

[0020] The processor, when executing computer-executable instructions or computer programs stored in the memory, implements the training method for the molecular property prediction model provided in the embodiments of this application.

[0021] This application provides a computer-readable storage medium storing computer-executable instructions or computer programs, which, when executed by a processor, implement the training method for the molecular property prediction model provided in this application.

[0022] This application provides a computer program product, including a computer program or computer executable instructions. When the computer program or computer executable instructions are executed by a processor, they implement the training method for the molecular property prediction model provided in this application.

[0023] The embodiments of this application have the following beneficial effects:

[0024] By masking the non-geometric information of sample molecules and adding noise to their geometric information, these samples are combined into a pre-training dataset. This increases the replicability of the sample data. Based on this processed data, the model can perform the processing of actual data for prediction. This enhances the molecular property prediction model's ability to identify and predict both geometric and non-geometric information of atoms in molecules during the first training task. Furthermore, training the pre-trained molecular property prediction model's parameters using molecular property prediction processing improves its accuracy. Simultaneously, since the model has learned the basic knowledge of geometric and non-geometric information of atoms in molecules after the first training task, the convergence speed of the molecular property prediction model is accelerated when the second training task specifically trains the molecular property prediction function, thus saving training time and computational resources. Attached Figure Description

[0025] Figure 1 This is a schematic diagram illustrating the application mode of the training method for the molecular property prediction model provided in the embodiments of this application;

[0026] Figure 2 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application;

[0027] Figure 3A This is a schematic diagram of the first process of training the molecular property prediction model provided in the embodiments of this application;

[0028] Figure 3B This is a schematic diagram of the second process of the training method for the molecular property prediction model provided in the embodiments of this application;

[0029] Figure 3C This is a schematic diagram of the third process of the training method for the molecular property prediction model provided in the embodiments of this application;

[0030] Figure 3D This is a schematic diagram of the fourth step in the training method of the molecular property prediction model provided in the embodiments of this application;

[0031] Figure 4A This is a schematic diagram of the molecular structure provided in the embodiments of this application;

[0032] Figure 4B This is a schematic diagram illustrating the principle of processing molecular data provided in the embodiments of this application;

[0033] Figure 5 This is a schematic diagram of the fifth step in the training method of the molecular property prediction model provided in the embodiments of this application;

[0034] Figure 6A This is a schematic diagram illustrating the principle of molecular property prediction provided in the embodiments of this application;

[0035] Figure 6B This is a schematic diagram illustrating the principle of the training molecular property prediction model provided in the embodiments of this application;

[0036] Figure 6C This is a schematic diagram illustrating the principle of the molecular property prediction model provided in the embodiments of this application.

[0037] Figure 7 This is a schematic diagram of the molecular property prediction model provided in the embodiments of this application. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0039] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0040] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0041] It should be noted that the data collection and processing in this application should strictly comply with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0042] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0043] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0044] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.

[0045] 1) Graph Neural Network (GNN): A class of deep learning models used to process graph-structured data. GNNs learn node representations and graph representations by performing message passing and aggregation on the nodes and edges of the graph. GNNs have wide applications in social network analysis, knowledge graphs, molecular biology, and other fields. In a GNN, the graph is represented as G = (V, E), where V represents the set of nodes and E represents the set of edges. Given a node feature matrix X, the goal of a GNN is to learn a function f that maps the node feature matrix to a node representation matrix H, i.e., H = f(X).

[0046] 2) Molecules: The smallest relatively stable unit in a substance that can exist independently and retain the substance's physicochemical properties. Molecules are composed of atoms, which are bound together by certain forces in a specific order and arrangement to form molecules.

[0047] 3) Chemical bond: A chemical bond is a collective term for the strong interaction forces between two or more adjacent atoms (or ions) within a pure molecule or crystal. The forces that bind ions or atoms together are collectively called chemical bonds. Types of chemical bonds include ionic bonds, covalent bonds, and metallic bonds.

[0048] 4) Molecular structure diagram: A graphical representation of the connections and spatial arrangement of atoms in a molecule. It graphically shows how atoms within a molecule are connected by chemical bonds and their relative positions in space. In the embodiments of this application, the edges in the molecular structure diagram represent the chemical bonds between atoms, and the nodes in the molecular structure diagram represent atoms.

[0049] 5) Molecular structural formula: This is a graphical representation of the connections between atoms in a molecule, using chemical symbols and line segments. The molecular structural formula reveals the arrangement of atoms in a molecule and the chemical bonds between them. Specifically, it shows how atoms are connected to each other through single, double, or triple bonds, and their relative positions in space.

[0050] 6) Equivariant Map: A map that is commutative with group actions. Given a map... Let X→Y, where X and Y are the input vector space and output vector space, respectively. If for a group G, we have... Then mapping It is a group G that is variable.

[0051] 7) Multilayer Perceptron (MLP): A type of feedforward neural network, typically consisting of an input layer, one or more hidden layers, and an output layer, hereinafter referred to as MLP. Each layer contains multiple neurons, and neurons in adjacent layers are connected by weights. MLP is a widely used supervised learning model for classification and regression problems.

[0052] 8) Equivariant Graph Neural Network (EPN): Equivariant Graph Neural Networks satisfy equivariance based on the message-passing framework of Graph Neural Networks (GNNs). An Equivariant Graph Neural Network contains several equivariant message-passing layers. In an Equivariant Graph Neural Network, each equivariant message-passing layer is used not only to update node features but also to maintain the structural invariance of the graph. This is typically achieved by introducing specific orthogonal transformations or equivariant mappings to ensure that the structural properties of the graph (such as cycles, trees, etc.) are preserved after the transformation.

[0053] 9) Molecular properties: These refer to the chemical and physical characteristics determined by the inherent properties of the molecule itself. These properties typically include the size, shape, composition, structure, chemical reactivity, stability, kinetic behavior, and intermolecular interactions. Macroscopic molecular properties include: solubility, boiling point, melting point, acidity / basicity, optical activity, magnetism, toxicity, and affinity.

[0054] 10) Atom: The smallest indivisible particle in a chemical reaction. The type of atom varies depending on the number of protons and neutrons: the number of protons determines which element the atom belongs to, while the number of neutrons determines which isotope of that element it is. Atoms form molecules, and molecules form matter. In matter, like charges repel each other, while unlike charges attract each other.

[0055] 11) Masking: This generally refers to the selective masking or ignoring of certain information during data processing. The key to masking is selectively modifying data to achieve a specific purpose or meet specific requirements. This processing method has wide applications in many fields, such as data preprocessing, feature extraction, privacy protection, and signal enhancement. In the embodiments of this application, masking is used to mask the types of atoms in the molecule to improve the effectiveness of training the molecular property prediction model.

[0056] 12) Noise Addition: This is a data preprocessing technique commonly used in machine learning and data analysis to improve the generalization ability and robustness of models. Its core idea is to artificially introduce random noise into coordinate data to simulate the uncertainty of data in the real world. Noise types can include Gaussian noise, salt and pepper noise, or more complex physics-based perturbations. The magnitude and type of noise can be adjusted according to the specific task and data.

[0057] 13) Pre-training: A common method in deep learning, it involves training a model on a large-scale unlabeled dataset to learn the general features or patterns of the data. The main purpose of pre-training is to improve model performance using unlabeled data, especially when labeled data is scarce. The process of pre-training a model typically includes two phases: the pre-training phase and the fine-tuning phase. In the pre-training phase, the model is trained on a large-scale unlabeled dataset to learn the underlying structure and features of the data. This process often involves unsupervised or self-supervised learning tasks, such as predicting the next word, predicting missing words, predicting future frames, etc.

[0058] This application provides a training method, a training device, an electronic device, a computer-readable storage medium, and a computer program product for a molecular property prediction model, which can improve the accuracy of molecular property prediction through a phased training process.

[0059] The following describes exemplary applications of the electronic devices provided in the embodiments of this application. These electronic devices can be implemented as terminal devices, such as laptops, tablets, desktop computers, set-top boxes, smart TVs, in-vehicle terminals, virtual reality (VR) devices, augmented reality (AR) devices, and other various types of terminals. They can also be implemented as servers. The following will describe exemplary applications when the electronic device is implemented as a terminal device or a server.

[0060] refer to Figure 1 , Figure 1 This is a schematic diagram illustrating the application mode of the training method for the molecular property prediction model provided in the embodiments of this application; for example, Figure 1 The system involves server 200, network 300, terminal device 400, and database 500. Terminal device 400 is connected to server 200 through network 300, which can be a wide area network, a local area network, or a combination of both.

[0061] For example, terminal device 400 can be a terminal device used by a user, such as a computer or mobile phone; the user can be a technician; server 200 can be a server of a molecular property analysis platform; and database 500 can be a molecular database that stores a large amount of molecular-related data.

[0062] For example, a user sends a training request through terminal device 400, which is then sent to server 200 via network 300. Server 200 extracts a large number of sample molecular structure diagrams from database 500 based on the detection request, and calls the training method of the molecular property prediction model provided in this embodiment to train the initialized molecular property prediction model, thus obtaining a trained molecular property prediction model. The trained molecular property prediction model can be stored in server 200.

[0063] The user sends molecular data to be processed through terminal device 400. The molecular data can be represented in the form of a molecular structure diagram. Each node in the molecular structure diagram is an atom, and each atom is associated with corresponding geometric and non-geometric information. The molecular data to be processed is uploaded to server 200 through network 300. Server 200 calls a trained molecular property prediction model based on the molecular data to predict the molecular properties of the molecules corresponding to the molecular data. Server 200 sends the predicted molecular properties to terminal device 400 so that the user can determine the molecular properties corresponding to the molecular data and further study the substances associated with the molecular structure diagram.

[0064] In some embodiments, the training method of the molecular property prediction model of this application can also be applied in the following application scenarios:

[0065] 1. Drug Discovery and Design: Based on the training method of the molecular property prediction model provided in the embodiments of this application, the molecular property prediction model is trained, and based on the trained molecular property prediction model, the molecular properties such as the biological activity and toxicity of compounds are predicted to screen potential therapeutic drugs. The molecular property prediction model can help scientists evaluate the biological activity, toxicity, solubility, metabolic stability, and other properties of candidate drug molecules, thereby screening more promising candidate drugs at an early stage and accelerating the drug development process. Alternatively, for existing drugs, the predicted molecular properties can guide the structural optimization of drug molecules to improve efficacy and reduce side effects.

[0066] 2. Chemistry and Materials Science: Based on the training method of the molecular property prediction model provided in the embodiments of this application, the molecular property prediction model is trained, and based on the trained molecular property prediction model, the physical and chemical properties of new materials, such as conductivity, thermal stability, and optical properties, are predicted. The predicted molecular properties can be used to design novel electronic components, catalysts, energy storage materials, etc.

[0067] 3. Food Science: Based on the training method of the molecular property prediction model provided in the embodiments of this application, the molecular property prediction model is trained, and based on the trained molecular property prediction model, the molecular properties of compound molecules in food are predicted to determine whether there are harmful substances in the food, such as toxins, pesticide residues, etc., so as to ensure food safety.

[0068] 4. Agriculture: Based on the training method of the molecular property prediction model provided in the embodiments of this application, the molecular property prediction model is trained, and based on the trained molecular property prediction model, the properties of molecules in pesticides are predicted. In pesticide research and development, the molecular property prediction model can predict the activity, residue and environmental behavior of pesticides, which is of great significance for the development of environmentally friendly pesticides.

[0069] 5. Environmental Science: Based on the training method of the molecular property prediction model provided in the embodiments of this application, the molecular property prediction model is trained, and based on the trained molecular property prediction model, the molecular properties of molecules in water and air collected in the environment are determined, so as to determine whether there are pollutants in the environment based on the molecular properties.

[0070] This application embodiment can be implemented using database technology. A database, simply put, can be viewed as an electronic filing cabinet storing electronic files, where users can perform operations such as adding, querying, updating, and deleting data. A "database" is a collection of data stored together in a certain way, capable of being shared by multiple users, having minimal redundancy, and being independent of application programs.

[0071] A Database Management System (DBMS) is a computer software system designed to manage databases, generally possessing basic functions such as storage, retrieval, security, and backup. DBMSs can be classified according to the database model they support, such as relational or XML (Extensible Markup Language); or according to the type of computer they support, such as server clusters or mobile devices; or according to the query language used, such as Structured Query Language (SQL) or XQuery; or according to performance priorities, such as maximum scale or maximum operating speed; or other classification methods. Regardless of the classification method used, some DBMSs can cross categories, for example, simultaneously supporting multiple query languages.

[0072] In some embodiments, the server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. Electronic devices can be smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, etc., but are not limited to these. Terminal devices and servers can be directly or indirectly connected via wired or wireless communication, which is not limited in this embodiment.

[0073] See Figure 2 , Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may be... Figure 1 Server 200 in the middle, Figure 2 The server 200 shown includes at least one processor 410, memory 450, and at least one network interface 420. The various components of server 200 are coupled together via a bus system 440. It is understood that the bus system 440 is used to implement communication between these components. In addition to a data bus, the bus system 440 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 2 The general labeled all buses as Bus System 440.

[0074] The processor 410 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0075] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 450 may optionally include one or more storage devices physically located away from the processor 410.

[0076] The memory 450 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 450 described in this application embodiment is intended to include any suitable type of memory.

[0077] In some embodiments, memory 450 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0078] Operating system 451 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks;

[0079] The network communication module 452 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 420, exemplary network interfaces 420 including: Bluetooth, WiFi, and Universal Serial Bus (USB), etc.

[0080] In some embodiments, the apparatus provided in this application can be implemented in software. Figure 2 A training device 455 for a molecular property prediction model stored in memory 450 is shown. This device can be software in the form of programs and plugins, and includes the following software modules: a sample acquisition module, a first training module 4552, and a second training module 4553. These modules are logically connected and can therefore be arbitrarily combined or further separated according to their implemented functions. The functions of each module will be described below.

[0081] The training method of the molecular property prediction model provided in this application will be described in conjunction with the exemplary application and implementation of the terminal device provided in the embodiments of this application.

[0082] The training method for the molecular property prediction model provided in this application embodiment will be described below. As mentioned earlier, the electronic device implementing the training method for the molecular property prediction model in this application embodiment can be a terminal device or a server, or a combination of both. Therefore, the executing entity of each step will not be described again below.

[0083] It should be noted that the examples of molecular property prediction processing using the training method of the molecular property prediction model in the following text are based on the example of organic molecules. Based on the understanding of the following text, those skilled in the art can apply the training method of the molecular property prediction model provided in the embodiments of this application to the processing of other types of molecules, such as: performing molecular property prediction for molecules of multiple components in proteins, or performing molecular property prediction for molecules of metallic materials.

[0084] See Figure 3A , Figure 3A This is a flowchart illustrating the training method for the molecular property prediction model provided in this application embodiment, which will be combined with... Figure 3AThe steps shown are explained. Figure 3A The entity responsible for executing the steps is Figure 1 Server 200 in the middle.

[0085] In step 301, the first non-geometric information of at least one atom of the sample molecule is masked to obtain the second non-geometric information.

[0086] To facilitate understanding, the application scenario of the training method for the molecular property prediction model provided in this application embodiment is explained. The data representation of the sample molecule is a molecular structure diagram. Each node in the molecular structure diagram represents an atom, and the edges between nodes represent the chemical bonds between atoms. Each atom in the molecular structure diagram carries geometric and non-geometric information. The geometric information of the molecule mainly refers to the spatial structure of the atomic arrangement within the molecule, which determines the shape and spatial orientation of the molecule. The non-geometric information of the molecule covers properties other than its spatial structure, which are usually related to the chemical and physical properties of the molecule. The geometric information corresponding to atoms in the molecule refers to the spatial arrangement of atoms and the relative positions between atoms, including: atomic coordinates, chemical bond length, and chemical bond angle. The non-geometric information of atoms does not directly describe the spatial structure of the molecule, but rather information related to the chemical and physical properties of the molecule. The non-geometric information of atoms includes: atomic charge, atom type, atom chemical bond type, and atom hybridization state.

[0087] For example, masking is the process of replacing the original data with a mask. In this embodiment of the application, masking for non-geometric information can be implemented by replacing the original type with a pre-configured type.

[0088] In some embodiments, reference Figure 3B , Figure 3B This is a schematic diagram of the second process of the training method for the molecular property prediction model provided in the embodiments of this application; step 301 can be achieved through... Figure 3B Steps 3011 to 3012 are implemented, and the details are explained below.

[0089] In step 3011, each atom in the sample molecule is traversed, and the following processing is performed on the traversed atom: the traversed atom is selected based on a first pre-configured probability.

[0090] For example, each atom encountered is randomly selected based on a first pre-configured probability, which can be set according to the needs of the actual application scenario.

[0091] In some embodiments, selecting the traversed atoms based on the first pre-configured probability can be achieved in the following way: generating a random probability for the traversed atoms to obtain a first probability; when the first probability is less than the first pre-configured probability, determining that the atom is selected as the first target atom; when the first probability is greater than or equal to the first pre-configured probability, determining that the atom is not selected.

[0092] For example, during the random probability generation process, the generated probability ranges from 0 to 1. If the first probability is less than the first pre-configured probability, the current atom is used as the first target atom. Conversely, if the first probability is greater than or equal to the first pre-configured probability, the current atom is not masked.

[0093] In some embodiments, the first pre-configured probability can be replaced by a first pre-configured value. A random number is generated within a given preset value range to obtain a first random number. When the first random number is less than the first pre-configured value, the atom is determined to be selected as the first target atom; when the first random number is greater than or equal to the first pre-configured value, the atom is determined not to be selected. For example, if the value range is set to 0 to 100, and the first pre-configured value is 20, and a random number is generated for the traversed atom, resulting in a random number of 15, which is less than the first pre-configured value of 20, then the traversed atom is taken as the first target atom.

[0094] In step 3012, in response to the atom being selected as the first target atom, the first atom type mask in the first non-geometric information of the first target atom is set to a pre-configured type to obtain the second non-geometric information of the first target atom.

[0095] For example, in response to an atom not being selected as the first target atom, the first non-geometric information of the atom is maintained. In response to an atom being selected as the first target atom, the first atom type mask of the first target atom is set to a pre-configured type. The masking method is to replace the atom type label in the data corresponding to the first target atom with the label of the pre-configured type. The label of the pre-configured type can be set to 1 or 0. That is, the atom type label in the first non-geometric information is replaced with 0 to form the second non-geometric information.

[0096] In this embodiment, the atom types are masked randomly. By masking the atom types, the molecular property prediction model is forced to learn to rely on other contextual information to predict the masked portion. This helps the model capture more generalized patterns in the molecule, thereby improving the model's generalization ability on unseen data. Atom type masking can be used as part of unsupervised learning, where the model does not need to rely on labeled data but learns directly from the intrinsic structure of the data itself, saving computational resources required for training. By randomly selecting the masked atoms to use different masking patterns, the diversity of training data is effectively increased, helping the model see more diverse molecular representations.

[0097] Continue to refer to Figure 3A In step 302, the first geometric information of at least one atom of the sample molecule is subjected to noise processing to obtain the second geometric information.

[0098] As illustrated above, geometric information includes atomic coordinates, chemical bond lengths, and chemical bond angles. Noise addition processing involves adding noise to the original data. In this embodiment, noise addition processing for geometric information can be achieved by superimposing the original atomic coordinates with the coordinates generated by calling a noise function to obtain the noise-added atomic coordinates.

[0099] In some embodiments, reference Figure 3C , Figure 3C This is a schematic diagram of the third step in the training method of the molecular property prediction model provided in this application embodiment; step 302 can be achieved through... Figure 3C Steps 3021 to 3022 are implemented, and the details are explained below.

[0100] In step 3021, each atom in the sample molecule is traversed, and the following processing is performed on the traversed atom: the traversed atom is selected based on the second pre-configured probability.

[0101] For example, each atom encountered is randomly selected based on a second pre-configured probability, which can be set according to the needs of the actual application scenario.

[0102] In some embodiments, selecting the traversed atoms based on the second pre-configured probability can be achieved in the following way: generating a random probability for the traversed atoms to obtain a second probability; when the second probability is less than the second pre-configured probability, determining that the atom is selected as the second target atom; when the second probability is greater than or equal to the second pre-configured probability, determining that the atom is not selected.

[0103] For example, during the random probability generation process, the generated probability ranges from 0 to 1. If the second probability is less than the value of the second pre-configured probability, then the current atom is used as the second target atom. Conversely, if the second probability is greater than or equal to the value of the second pre-configured probability, no noise processing is performed on the current atom.

[0104] In some embodiments, the second pre-configured probability can be replaced by a second pre-configured value. A random number is generated within a given preset value range to obtain a second random number. When the second random number is less than the second pre-configured value, the atom is determined to be selected as the second target atom; when the second random number is greater than or equal to the second pre-configured value, the atom is determined not to be selected. For example, if the value range is set to 0 to 100, and the second pre-configured value is 20, and a random number is generated for the traversed atom, resulting in a random number of 15, which is less than the second pre-configured value of 20, then the traversed atom is selected as the second target atom.

[0105] In step 3022, in response to the atom being selected as the second target atom, pre-configured noise is added to the first coordinates in the first geometric information of the second target atom to obtain the second geometric information of the second target atom.

[0106] For example, in response to an atom not being selected as the second target atom, the atom's first geometric information is maintained. In response to an atom being selected as the second target atom, the first coordinates of the second target atom are superimposed with pre-configured noise, which can be a Gaussian function. For example: atomic geometric information. Characterized by the three-dimensional coordinates of atoms, i.e. coordinates of atom i Add Gaussian noise ∈ i ,Right now Gaussian noise is superimposed on the first coordinates in the first geometric information to form the second geometric information.

[0107] In this embodiment, noise is added to the atomic coordinates in a random manner. This noise addition allows the molecular property prediction model to adapt to random fluctuations in the input data, making it more robust when processing real-world data and reducing prediction errors caused by small noise introduced during data acquisition or processing. By introducing noise, the molecular property prediction model is forced to learn more generalized features, rather than simply memorizing specific details from the training data, thus improving its performance on new data. Furthermore, in practical applications, the atomic coordinates obtained experimentally may contain certain errors. Noise addition can simulate these experimental errors to some extent, making the trained model closer to real experimental conditions.

[0108] Continue to refer to Figure 3AIn step 303, the first geometric information, the first non-geometric information, the second non-geometric information of the atoms that have not undergone masking, the first geometric information and the second geometric information of the atoms that have not undergone noise processing are combined into a pre-training data set.

[0109] For example, atoms used to perform masking are used as mask atoms, and atoms used to perform noise-adding are used as noise-adding atoms; where the noise-adding atom and the mask atom are the same atom, or the noise-adding atom and the mask atom are different atoms. Then, step 303 can be interpreted as combining the first geometric information of the atoms other than the noise-adding atoms, the second geometric information of the noise-adding atoms, the first non-geometric information of the atoms other than the mask atoms, and the second non-geometric information of the mask atoms into a pre-training data set. For ease of understanding, refer to... Figure 4A as well as Figure 4B , Figure 4A This is a schematic diagram of the molecular structure diagram provided in the embodiments of this application; the molecular formula of the benzene ring is (C6H6), and the molecular structure formula of the benzene ring is converted into a molecular structure diagram. Each node in the molecular structure diagram is an atom, and the edges between atoms represent chemical bonds. Figure 4B This is a schematic diagram of processing molecular data provided in the embodiments of this application; an atomic node A is randomly selected from the benzene ring molecular structure diagram, the atom type of atomic node A is masked (atomic node A represents the masked atom), and the coordinates of atomic node B are noise-added (atomic node B represents the noise-added atom) to obtain a sample molecular structure diagram. The sample molecular structure diagram (that is, a molecular structure diagram in the pre-training dataset) can be used to pre-train a molecular property prediction model to obtain a pre-trained molecular property prediction model.

[0110] In step 304, based on the pre-trained dataset, the molecular property prediction model is invoked to perform the first training task, resulting in the trained molecular property prediction model.

[0111] For example, the molecular property prediction model includes: a graph neural network, a first prediction module, a second prediction module, and a molecular property prediction module. The graph neural network is used to extract features from molecular data, which includes geometric and non-geometric information. The first prediction module predicts the first non-geometric information corresponding to the second non-geometric information, the second prediction module predicts the first geometric information corresponding to the second geometric information, and the molecular property prediction module predicts molecular properties. The first training task is to pre-train the molecular property prediction model to train its prediction capabilities for both geometric and non-geometric information, improving the model's robustness and generalization. The loss function for the first training task is the sum of the loss functions corresponding to the two prediction functions. For the prediction of geometric information, the loss function can be cross-entropy loss or information entropy loss. For the prediction of non-geometric information, the loss function can be regression loss.

[0112] For ease of understanding, the following explanation is provided with reference to the accompanying drawings, illustrating the structural schematic diagram of the molecular prediction model provided in the embodiments of this application. Figure 7 This is a schematic diagram of the molecular property prediction model provided in this application embodiment. The graph neural network in the molecular property prediction model can be implemented as an equivariant graph neural network. The first prediction module is implemented as a type prediction module, which can be a convolutional neural network model for performing classification processing. The second prediction module is implemented as a coordinate prediction module, which can be a recurrent neural network model for performing regression processing. The molecular property prediction module includes a pooling layer and a multilayer perceptron.

[0113] The molecular property prediction model 700 includes: an equivariant graph neural network 701, a molecular property detection module 702, a type prediction module 703, and a coordinate prediction module 704. The molecular property detection module 702 includes: a pooling layer 7021 and a property prediction multilayer perceptron 7022. The equivariant graph neural network 701 includes multiple message passing layers, for example: message passing layer 1 to message passing layer N, where N is a positive integer, and each message passing layer includes a multilayer perceptron. The type prediction module 703 is used to determine the predicted coordinate information of the noisy atoms based on the output of the equivariant graph neural network 701, and the coordinate prediction module 704 is used to determine the predicted atom type of the masked atoms based on the output of the equivariant graph neural network 701.

[0114] In some embodiments, reference Figure 3D , Figure 3D This is a schematic diagram of the fourth step in the training method of the molecular property prediction model provided in this application embodiment; step 304 can be achieved through... Figure 3D Steps 3041 to 3044 are implemented, and the details are explained below.

[0115] In step 3041, based on each first non-geometric information and each second non-geometric information, the molecular property prediction model is invoked to perform atom type prediction processing, and the first cross-entropy loss of the molecular property prediction model is determined based on the result of the atom type prediction processing.

[0116] For example, atomic type prediction processing includes feature extraction and classification based on the extracted features. Cross-entropy loss, a commonly used loss function in machine learning and deep learning, is suitable for classification problems. Its core purpose is to measure the difference between the probability distribution predicted by the model and the probability distribution of the true labels.

[0117] In some embodiments, step 3041 can be implemented as follows: based on each first non-geometric information and each second non-geometric information, a molecular property prediction model is invoked to perform feature extraction processing to obtain the first geometric features of the sample molecule; based on the first geometric features, the mask atoms in the sample molecule are classified to obtain the prediction type of each mask atom, wherein the mask atom is an atom used to perform mask processing; based on the prediction type of the mask atom and the first atom type, the first cross-entropy loss of the molecular property prediction model is determined.

[0118] For example, the first geometric feature is extracted using the graph neural network described above. In the equivariant graph neural network, the output of the previous layer is used as the input of the current layer. (Where, the superscript l represents the value after l iterations), the geometric information of the current layer is calculated using the following formulas (2.1) to (2.4). Non-geometric information

[0119]

[0120] Where σ1, σ2, and σ3 are multilayer perceptrons. ij It represents the update value corresponding to the non-geometric information and geometric information between atoms i and j.

[0121] In some embodiments, the equivariant graph neural network can also introduce more complex update forms. For example, an attention mechanism can be introduced into the equivariant graph neural network, and the update formula of the equivariant graph neural network can be modified to the following formulas (3.1) to (3.6):

[0122]

[0123] Where, α ij It is the attention coefficient, W Q W K It is a learnable weight matrix.

[0124] In some embodiments, the non-geometric representation of the output of a given isomorphic graph neural network is: Using multilayer perceptrons to Perform mapping and predict the probability p using the atomic type of the normalization function (softmax). i This can be expressed as formulas (4.1) and (4.2) below:

[0125]

[0126] Among them, s i ∈R C C represents the number of atomic types, (·) kRepresents the k-th element of the vector, (p i ) k Let p represent the probability that atom i is of type k. The atom type with the highest predicted probability is selected as the predicted atom type, and the cross-entropy loss between the predicted type and the actual type is determined. The predicted probability p of the predicted atom type is then used. i and actual type of tag l i Calculate the first cross-entropy loss L for the predicted mask atom type. tp This can be expressed as the following formula (5.1):

[0127] L tp =-∑ i∈A ∑ k (l i ) k log(p i ) k (5.1)

[0128] Where A is the set of atoms whose type is masked, (·) k Let l represent the k-th element of the vector. i It is the one-hot encoding of the type of the atom before it was masked. For example, if the type of the atom before it was masked is 'c', then...

[0129] In step 3042, based on each first geometric information and each second geometric information, the molecular property prediction model is invoked to perform atomic position prediction processing, and the first regression loss of the molecular property prediction model is determined based on the result of the atomic position prediction processing.

[0130] For example, atomic type prediction processing includes feature extraction and position regression based on the extracted features. Regression loss is a metric used to measure the difference between the predicted value and the actual value of a regression model. In regression tasks, the goal of the model is to predict a continuous numerical value, rather than a discrete class label as in classification tasks. Therefore, the regression loss function is used to evaluate how close the model's predicted value is to the actually observed value.

[0131] In some embodiments, step 3042 can be implemented as follows: based on each first geometric information and each second geometric information, the molecular property prediction model is invoked to perform feature extraction processing to obtain the first non-geometric features of the sample molecule; based on the first non-geometric features, the positions of the noisy atoms in the sample molecule are predicted to obtain the predicted coordinates of each mask atom, wherein the noisy atom is an atom used to perform the noisy processing; based on the predicted coordinates of the noisy atoms and the first coordinates, the first regression loss of the molecular property prediction model is determined.

[0132] Example, based on the predicted coordinates. and the label of the actual first coordinate Calculate the regression loss L for coordinate prediction coo This can be represented by the following formula (5.2):

[0133]

[0134] In step 3043, the first total loss of the molecular property prediction model is determined based on the first cross-entropy loss and the first regression loss.

[0135] For example, the first total loss can be the sum of the first cross-entropy loss and the first regression loss, for instance: the first total loss is represented as L pre =L tp +L coo Alternatively, the result of a weighted summation.

[0136] In step 3044, based on the first total loss, the parameters of the molecular property prediction model are updated through backpropagation to obtain the trained molecular property prediction model.

[0137] For example, the parameters of each part of the molecular property prediction model can be updated through backpropagation. This process updates the parameters of each part of the molecular property prediction model to reduce the value of the first total loss. This process is repeated multiple times until the number of updates reaches the corresponding pre-configured iteration number, resulting in a molecular property prediction model that has completed the training process for the first training task.

[0138] In this embodiment, a phased training method is used to add masks and noise to the original data. Introducing masks and noise forces the model to learn more robust and generalizable features, rather than simply memorizing specific details from the training data. This helps the model perform better on new, unseen data. Masks and noise increase the diversity of the data, making it more difficult for the model to achieve a perfect fit on the training data, thereby reducing the risk of overfitting.

[0139] Continue to refer to Figure 3A In step 305, based on the original data set of sample molecules, a second training task is performed on the trained molecular property prediction model to obtain the trained molecular property prediction model.

[0140] For example, the second training task is used to train the molecular property prediction model to perform molecular property prediction processing. This belongs to the fine-tuning stage of the training process. Based on the pre-trained model, a small amount of task-specific data is used to further train the model to adapt it to a specific task or domain. The loss function used in the second training task can be cross-entropy loss, and the training method used is supervised training.

[0141] In some embodiments, the original dataset includes sample molecule data and the actual molecular properties of the sample molecules; step 305 can be implemented as follows: based on the sample molecule data, the trained molecular property prediction model is called for classification processing to obtain the first predicted probability of the sample molecule for different molecular properties; the molecular property with the highest first predicted probability is selected as the predicted molecular property of the sample molecule; the second cross-entropy loss of the molecular property prediction model is determined based on the predicted molecular property and the actual molecular property; based on the second cross-entropy loss, the parameters of the molecular property prediction model are updated through backpropagation to obtain the trained molecular property prediction model.

[0142] For example, the classification process is implemented through a graph neural network and a property prediction module in the molecular property prediction model. The graph neural network is used to extract features from the sample molecular data, and the pooling layer of the property prediction module is used to pool the feature extraction results. The multilayer perceptron of the property prediction module calls the normalization function based on the pooling results to predict the first prediction probability of the sample molecule for different molecular properties.

[0143] For example, taking non-geometric information as an example, suppose the non-geometric information of atom i output by the equivariant graph neural network is: Pooling layers aggregate the non-geometric information of all atomic nodes in the sample molecular structure diagram through summation to obtain global information h. global This can be represented by the following formula (7.1):

[0144]

[0145] Where V is the set of nodes in the molecular structure diagram.

[0146] In some embodiments, the pooling layer can simultaneously aggregate non-geometric information of points and edges, that is, the calculation formula of the pooling layer can be modified to the following formula (7.2):

[0147]

[0148] Based on the aforementioned global non-geometric information h global The property detection multilayer perceptron is invoked, and the calculation formula corresponding to the multilayer perceptron is as follows: Formula (8):

[0149]

[0150] Among them, the non-geometric information h based on the global global The multilayer perceptron (MLP) is invoked to predict the predicted values ​​of molecular properties, where Act(·) is the activation function. These are the weights (W) and biases (b) of the two linear layers of the multilayer perceptron, respectively, and y is the predicted value of the molecular properties.

[0151] Second cross-entropy loss L pred It can be characterized as Assuming there are multiple samples, the total loss of the model can be the mean absolute error (MAE), which is represented by the following formula (9):

[0152]

[0153] in, It is a predicted value. Compared with the actual value y i The error between the values ​​is calculated, where n is a positive integer and n is the number of molecules used as samples. The molecular property prediction model is then fine-tuned overall using the total loss.

[0154] In some embodiments, the original dataset further includes sample molecule data, the actual molecular properties of the sample molecules, and the actual properties of the atomic groups in the sample molecules; step 305 can be implemented as follows: based on the sample molecule data, the trained molecular property prediction model is invoked for feature extraction processing to obtain the molecular features of the sample molecules; based on the molecular features, molecular property classification processing is performed to obtain the predicted molecular properties of the sample molecules, and the second cross-entropy loss of the molecular property prediction model is determined based on the predicted molecular properties and the actual molecular properties; based on the molecular features, the properties of atomic groups are classified to obtain the predicted properties of the atomic groups in the sample molecules, and the third cross-entropy loss of the molecular property prediction model is determined based on the predicted properties of the atomic groups and the actual properties; based on the second cross-entropy loss and the third cross-entropy loss, the second total loss is determined; based on the second total loss, the parameters of the molecular property prediction model are updated through backpropagation to obtain the trained molecular property prediction model.

[0155] For example, an atomic group refers to a group of atoms in a molecule that typically participates in a chemical reaction as a whole. Atomic groups can be neutral or charged. The study of atomic groups is crucial for understanding the properties and reaction mechanisms of compounds. By identifying and modifying specific atomic groups, molecular properties can be determined. The method for determining the predicted properties of atomic groups can refer to the molecular property prediction process described above. This process includes: feature extraction of the atomic groups; using a multilayer perceptron to predict the probability of different properties based on the feature extraction results; and selecting the property with the highest prediction probability as the predicted property. The method for obtaining the cross-entropy loss can be referred to above and will not be elaborated upon here. The second total loss can be the sum of the second and third cross-entropy losses, or a weighted summation.

[0156] In this embodiment, based on the second cross-entropy loss, a third cross-entropy loss corresponding to the predicted properties of atomic groups is combined. This allows both molecular-level and atomic group-level properties to be considered simultaneously, enabling the model to learn molecular features at different scales and facilitating a more comprehensive understanding of molecular structure and properties. The properties of atomic groups have a significant impact on the overall properties of molecules. Combining group-level losses enhances the model's ability to identify local molecular features, thereby improving the accuracy of predicting overall molecular properties. By modeling the properties of atomic groups, the model can learn more generalized features that may be consistent across different types of molecules, enhancing the model's generalization ability across different molecular structures.

[0157] In some embodiments, based on sample molecule data, a trained molecular property prediction model is invoked to perform feature extraction processing to obtain the molecular features of the sample molecule. This can be achieved in the following way: based on sample molecule data, a trained molecular property prediction model is invoked to perform feature extraction processing at multiple levels to obtain the feature information of each atom in the sample molecule, wherein the output of each level of feature extraction processing is the input of the next level of feature extraction processing; the feature information of each atom is aggregated to obtain the molecular features of the sample molecule.

[0158] For example, the feature extraction processing of the trained molecular property prediction model is implemented through an equivariant graph neural network. The principle of the equivariant graph neural network is described above and will not be repeated here.

[0159] In some embodiments, after step 305, molecular property prediction processing can be performed as follows: based on the trained molecular property prediction model, feature extraction processing is performed on the molecular data of the molecule to be detected to obtain the second geometric feature and the second non-geometric feature of the molecule to be detected; based on the second geometric feature and the second non-geometric feature, the molecule to be detected is classified to obtain the second prediction probability of the molecule to be detected for different molecular properties; at least one molecular property with the highest second prediction probability is selected as the target molecular property of the molecule to be detected, wherein the types of molecular properties include: physical properties and chemical properties.

[0160] For example, the principles of feature extraction and classification can be found above, and will not be repeated here. Physical properties include electrical conductivity, thermal stability, and optical properties. Chemical properties include redox properties, acidity / alkalinity, toxicity, and biological activity. In practical applications, the same molecule may exhibit multiple different properties; the actual number of at least one molecular property selected will be determined based on the specific scenario.

[0161] In this embodiment, the non-geometric information of the sample molecules is masked, the geometric information of the sample molecules is noise-added, and the two are combined into a pre-training dataset. This increases the replicability of the sample data. Based on the processed sample data, the model can perform the processing to predict actual data. This enhances the molecular property prediction model's ability to identify and predict the geometric and non-geometric information of atoms in molecules during the first training task. Furthermore, training the parameters of the pre-trained molecular property prediction model to perform molecular property prediction processing improves the accuracy of the model's molecular property predictions. Simultaneously, since the molecular property prediction model has learned the basic knowledge of the geometric and non-geometric information of atoms in molecules after the first training task, the convergence speed of the molecular property prediction model is accelerated when the second training task specifically trains the molecular property prediction function, thereby saving training time and computational resources.

[0162] The following will illustrate an exemplary application of the training method for the molecular property prediction model of this application in a practical application scenario.

[0163] In related technologies, the model is pre-trained by processing molecular property prediction data. For example, masked coordinate prediction pre-training or coordinate denoising pre-training is performed on a large amount of unlabeled molecular data, and then fine-tuned on a small amount of labeled molecular property data before being applied to the molecular property prediction task. Masked atom type prediction pre-training methods first mask the atom types, and then use the classification task of predicting the masked atom types as the pre-training task for the molecular model. Coordinate denoising pre-training methods first add noise to the atom coordinates, and then use the regression task of predicting the noise in the atom coordinates as the pre-training task for the molecular model. Masked atom type prediction pre-training methods can only allow the model to learn the non-geometric properties of molecules, and coordinate denoising pre-training methods can only allow the model to learn the geometric properties of molecules. Neither of these pre-training methods allows the molecular model to learn both geometric and non-geometric properties simultaneously, which limits the performance gain of pre-training for the molecular model.

[0164] This application provides a training method for a molecular property prediction model, which combines masked atom type prediction and coordinate denoising joint pre-training. At the same time, the atom type is masked and the coordinates are denoised. Atom type recovery and noise prediction are used as pre-training tasks, which enables the molecular model to learn both the geometric and non-geometric properties of the molecule. This improves the generalization ability of the molecular model through pre-training and thus enhances the prediction accuracy of the model in downstream tasks of molecular property prediction.

[0165] The training method of the molecular property prediction model provided in the embodiments of this application will be explained below with reference to the accompanying drawings. Figure 5 , Figure 5This is a schematic diagram of the fifth step in the training method of the molecular property prediction model provided in the embodiments of this application; Figure 5 The implementing body is Figure 1 Server 200 in the middle.

[0166] In step 501, the non-geometric feature information of the original molecular map data is masked, and the geometric information of the original molecular map data is noise-added to obtain sample molecular map data.

[0167] For example, molecular graph data is data represented as a molecular structure diagram. To facilitate understanding, the molecular structure diagram is explained as follows: molecules are composed of atoms and chemical bonds. Atoms can be regarded as nodes and chemical bonds as edges, thus abstracting the graph structure of molecules. Figure 4A This is a schematic diagram of the molecular structure provided in the embodiments of this application; the molecular formula of the benzene ring is (C6H6). The molecular structure of the benzene ring has been converted into a molecular structure diagram. Each node in the molecular structure diagram represents an atom, and the edges between atoms represent chemical bonds. Each atom a in the molecular structure diagram... i Both contain geometric and non-geometric information, and the definitions of each type of information are as follows:

[0168] Non-geometric information h of atoms i : Some properties of the atom itself, such as the number of charges, the number of protons, the number of neutrons, etc. i ∈R d , where d is the characteristic dimension of the atom.

[0169] Atomic geometric information The three-dimensional coordinates of an atom, i.e.

[0170] Molecular representation M: Representing a molecule as a set of atomic geometric and non-geometric information, i.e.

[0171] For example, the geometric information of a molecule mainly refers to the spatial structure of the atomic arrangement within the molecule, which determines the molecule's shape and spatial orientation. Specifically, it includes the following aspects: Molecular shape: determined by the bond angles and bond lengths between atoms in the molecule, it can be linear, triangular, tetrahedral, planar quadrilateral, etc. Molecular symmetry: the atomic arrangement in a molecule may have different symmetry elements, such as rotation axes, reflection planes, inversion centers, etc. Molecular configuration: in stereochemistry, molecular configuration refers to the specific spatial arrangement of atoms in a molecule, and these arrangements cannot be interchanged between molecules through rotational operations.

[0172] Non-geometric information about molecules encompasses properties other than their spatial structure. These properties are typically related to the chemical and physical properties of molecules, including: molecular mass and volume; molecular polarity: the non-uniform distribution of charge within a molecule, resulting in positive and negative poles; intermolecular forces: the interactions between molecules, such as van der Waals forces, hydrogen bonds, and ionic bonds; molecular motion: the movement of molecules in space, including translation, rotation, and vibration; and chemical properties: the ability of a molecule to participate in chemical reactions, including its reactivity and stability. In chemical and physical research, both of these characteristics are fundamental to understanding the properties of matter and reaction mechanisms. By studying the geometric properties of molecules, we can predict their physical and chemical properties; while non-geometric properties directly relate to the actual behavior and function of molecules.

[0173] In this embodiment, masking the non-geometric feature information of the original molecular diagram data refers to masking the atom types of atoms in the molecule. The masking method can be to set the atom type to a preset atom type. For example, masking the type of atom i means treating the masked atom type as an atom type and using it to initialize the non-geometric representation h of atom i. i .

[0174] Adding noise to the geometric information of the original molecular diagram data refers to adding Gaussian noise to the coordinates of atoms in the molecule's geometry. For example, adding Gaussian noise to the coordinates of atom i. Add Gaussian noise ∈ i ,Right now

[0175] Figure 4B This is a schematic diagram of processing molecular data provided in the embodiments of this application; an atomic node A is randomly selected from the benzene ring molecular structure diagram, the atom type of atomic node A is masked, and the coordinates of atomic node B are noise-added to obtain a sample molecular structure diagram. The sample molecular structure diagram can be used to pre-train a molecular property prediction model to obtain a pre-trained molecular property prediction model.

[0176] Different molecules are input for masking and noise addition. The algorithm steps for this process are as follows:

[0177] (1) Input molecular structure diagram feature information: The point set of the given input molecular structure diagram is V.

[0178] (2) Randomly select the atomic nodes to be masked: For each node i∈V in the point set, perform a random trial with probability p, and determine whether to mask the type of atom i based on the result of the random trial. That is, treat the masked atom type as an atom type and use it to initialize the non-geometric representation h of atom i. i .

[0179] (3) Randomly select the atom nodes to be noise-added: Traverse each node i∈V in the point set, perform a random trial with probability p, and determine whether to add noise to the coordinates of atom i based on the result of the random trial. Add Gaussian noise ∈ i ,Right now

[0180] (4) Output: Returns a molecular structure diagram that can be applied to the pre-training process.

[0181] In step 502, the molecular property prediction model is pre-trained as a whole based on the sample molecular map data.

[0182] For example, the joint pre-training method of masked atom type prediction and coordinate denoising can be combined with any molecular property prediction model. In this embodiment, a molecular property prediction model based on an isovariant graph neural network is used as an example for illustration. For ease of understanding, the structure of the molecular property prediction model provided in this embodiment is explained. Figure 7 This is a schematic diagram of the molecular property prediction model provided in this application embodiment. The molecular property prediction model 700 includes: an equivariant graph neural network 701, a molecular property detection module 702, a type prediction module 703, and a coordinate prediction module 704. The molecular property detection module 702 includes: a pooling layer 7021 and a property prediction multilayer perceptron 7022. The equivariant graph neural network 701 includes multiple message passing layers, for example: message passing layer 1 to message passing layer N, where N is a positive integer, and the message passing layers include multilayer perceptrons. The type prediction module 703 is used to determine the predicted coordinate information of the noisy atom based on the output of the equivariant graph neural network 701, and the coordinate prediction module 704 is used to determine the predicted atom type of the masked atom based on the output of the equivariant graph neural network 701.

[0183] Commonly used graph neural networks include graph convolutional networks (GCNs), graph attention networks (GATs), and graph isomorphism networks (GINs). The basic operation of graph neural networks is message passing, and a graph neural network can be represented by the following formula (1):

[0184]

[0185] in, This represents the representation of node v at level l. It is a non-linear activation function, where N(v) represents the set of neighboring nodes of node v, and M is a message function used to compute messages between node u and node v.

[0186] In a graph neural network, the output of the previous layer is used as the input of the current layer. (Where, the superscript l represents the value after l iterations), the geometric information of the current layer is calculated using the following formulas (2.1) to (2.4). Non-geometric information

[0187]

[0188] Where σ1, σ2, and σ3 are multilayer perceptrons. ij It represents the update value corresponding to the non-geometric information and geometric information between atoms i and j.

[0189] In some embodiments, the equivariant graph neural network can also introduce more complex update forms. For example, an attention mechanism can be introduced into the equivariant graph neural network, and the update formula of the equivariant graph neural network can be modified to the following formulas (3.1) to (3.6):

[0190]

[0191]

[0192] Where, α ij It is the attention coefficient, W Q W K It is a learnable weight matrix.

[0193] Based on the feature information of each atom output by the equivariant graph neural network (the first output feature information above), the type prediction module is used to predict the type of the masked atom, and the coordinate prediction module is used to predict the coordinates of the atom with added noise.

[0194] In some embodiments, the non-geometric representation of the output of a given isomorphic graph neural network is: First, a multilayer perceptron is used to... Perform mapping and predict the probability p using the atomic type of the normalization function (softmax). i This can be expressed as formulas (4.1) and (4.2) below:

[0195]

[0196] Among them, s i ∈R C C represents the number of atomic types, (·) k Represents the k-th element of the vector, (p i ) k This represents the probability that atom i is of type k.

[0197] For example, the molecular map with masked atom types and added coordinate noise is input into the molecular model, and the output p of the masked atom type prediction module is used. i and actual label l i Calculate the cross-entropy loss L for the predicted mask atom type tp This can be expressed as the following formula (5.1):

[0198] L tp =-∑ i∈A ∑ k (l i ) k log(p i ) k (5.1)

[0199] Where A is the set of atoms whose type is masked, (·) k Let l represent the k-th element of the vector. i It is the one-hot encoding of the type of the atom before it was masked. For example, if the type of the atom before it was masked is 'c', then...

[0200] In some embodiments, the coordinate prediction module uses the geometric representation output by the isomorphic graph neural network as the predicted value of the noisy coordinates. That is, given the geometric representation of atom i output by the last layer of the isomorphic graph neural network, the prediction module uses the geometric representation of atom i as the predicted value of the noisy coordinates. Then the coordinate prediction value The calculation formula is as follows (4.3):

[0201]

[0202] For example, the noisy molecular structure map is input into the molecular property prediction model, and the output of the mask prediction module is used. and actual label Calculate the regression loss L for coordinate prediction coo This can be represented by the following formula (5.2):

[0203]

[0204] Regression loss L predicted using coordinates coo and cross-entropy loss L tp The molecular property prediction model is pre-trained holistically. The overall loss function is represented as L. pre =L tp +L coo .

[0205] For example, using the total pre-training loss L prePre-training the molecular property prediction model can be performed using the backpropagation algorithm, which optimizes the weights through gradient descent to reduce the loss function. Backpropagation is an algorithm used to train neural networks; it updates the weights by calculating the gradient of the loss function with respect to the weights. In backpropagation, forward propagation is first performed to calculate the output of each neuron, and then the loss value of the output layer is calculated. Next, the error is propagated backward from the output layer to the input layer, updating the weights of each neuron. Assume the loss of the j-th neuron in the (l+1)-th layer is... If the derivative of the activation function is f′, then the following formula (6.1) holds:

[0206]

[0207] in, This is the target output of the j-th neuron in the (l+1)-th layer. Based on this, the loss of the i-th neuron in the l-th layer is calculated, represented by formula (6.2):

[0208]

[0209] Finally, the weights are updated using the following formula (6.3).

[0210]

[0211] Where η is the learning rate.

[0212] In step 503, the pre-trained molecular property prediction model is fine-tuned based on the original molecular graph data to obtain the trained molecular property prediction model.

[0213] For example, after pre-training, the molecular property prediction model will receive pre-trained weights, which contain general features and knowledge that distinguish different molecule types. During the fine-tuning phase, these pre-trained weights are used as the initial weights of the molecular property prediction model and trained on specific molecular property prediction tasks to adjust the model's weights to suit the task.

[0214] Step 503 can be implemented as follows: The sample molecular structure diagram is input into the pre-trained molecular property prediction model. The isomorphic graph neural network in the molecular property prediction model performs feature extraction processing on the sample molecular structure diagram to obtain the first output feature information of each atom in the sample molecular structure diagram. The output feature information of atom i is represented as... In the molecular property prediction model, the pooling layer aggregates the output feature information of each atom to obtain global feature information. Based on the global feature information, the property prediction multilayer perceptron in the molecular property prediction model is called to obtain the predicted molecular properties. The loss function is determined based on the predicted molecular properties and the actual molecular properties. The molecular property prediction model is then backpropagated based on the loss function to obtain the trained molecular property prediction model.

[0215] refer to Figure 6B , Figure 6B This is a schematic diagram illustrating the principle of training a molecular property prediction model provided in this application embodiment. The training samples include: sample geometric information and sample non-geometric information of molecules (taking benzene ring molecules as an example), and molecular property labels. The sample geometric information and sample non-geometric information are input into a molecular property prediction model that is pre-trained based on masked atom type prediction and coordinate denoising. The molecular property prediction model that is pre-trained based on masked atom type prediction and coordinate denoising outputs predicted molecular properties. The predicted molecular properties are compared with the actual molecular properties corresponding to the molecular property labels to determine the loss of the model to be trained. The pre-trained molecular property prediction model is then fine-tuned based on the loss to obtain the trained molecular property prediction model.

[0216] For example, suppose the non-geometric information of atom i output by the isomorphic graph neural network is: Pooling layers aggregate the non-geometric information of all atomic nodes in the sample molecular structure diagram through summation to obtain the global non-geometric information h. global (That is, the global feature information mentioned above), which is represented by the following formula (7.1):

[0217]

[0218] Where V is the set of nodes in the molecular structure diagram.

[0219] In some embodiments, the pooling layer can simultaneously aggregate non-geometric information of points and edges, that is, the calculation formula of the pooling layer can be modified to the following formula (7.2):

[0220]

[0221] Based on the aforementioned global non-geometric information h global The property detection multilayer perceptron is invoked, and the calculation formula corresponding to the multilayer perceptron is as follows: Formula (8):

[0222]

[0223] Among them, the non-geometric information h based on the global global The multilayer perceptron (MLP) is invoked to predict the predicted values ​​of molecular properties, where Act(·) is the activation function. These are the weights (W) and biases (b) of the two linear layers of the multilayer perceptron, respectively, and y is the predicted value of the molecular properties.

[0224] If there are multiple sample molecular structure diagrams, then the sub-loss L corresponding to each sample molecular structure diagram is... pred It can be characterized as The total loss of the model can be the mean absolute error (MAE), which is represented by the following formula (9):

[0225]

[0226] in, It is a predicted value. Compared with the actual value y i The error between the values ​​is calculated, where n is a positive integer and n is the number of molecules used as samples. The molecular property prediction model is then fine-tuned overall using the total loss.

[0227] In step 504, based on the molecular structure diagram to be processed, the trained molecular property prediction model is invoked to make predictions and obtain the predicted molecular properties.

[0228] For example, molecular property prediction is defined as follows: given a large unlabeled molecular dataset {M} i} and a small number of labeled molecular datasets {M i ,y i}(y i (Associated with molecular property labels), the task is to predict the properties of a given molecule M. The prediction principle has been explained in step 503 and will not be repeated here. (See reference...) Figure 6A , Figure 6A This is a schematic diagram illustrating the principle of molecular property prediction provided in the embodiments of this application; molecular properties are predicted based on the geometric and non-geometric information of the molecule (the benzene ring is used as an example in the embodiments of this application).

[0229] refer to Figure 6C , Figure 6C This is a schematic diagram illustrating the principle of the molecular property prediction model provided in this application embodiment. The molecular structure diagram of the benzene ring molecule is input into the molecular property prediction model, which is pre-trained based on masked atom type prediction and coordinate denoising. The molecular property prediction model outputs the predicted molecular properties for the benzene ring molecule.

[0230] In some embodiments, if the input molecular structure diagram contains geometric representations of masked or missing atoms, the model can call the mask prediction module to determine the predicted geometric information of the masked atoms and predict molecular properties based on the predicted geometric information and the actual feature information of the unmasked atoms. If the molecular structure diagram input to the model is a complete molecular structure diagram without any masked or missing parts, the mask prediction module branch of the model does not need to be used, and only the pooling layer and the property prediction multilayer perceptron are used to predict the properties of the molecular structure diagram.

[0231] In this embodiment, the molecular property prediction model based on the joint pre-training of masked atom type prediction and coordinate denoising can be applied to molecular property prediction tasks; it is universal and can be extended to other microstructures, such as proteins and RNA; it is versatile and can be applied in different scenarios, such as industrial applications, where predicting molecular properties can promote drug discovery and material design.

[0232] The following description continues to illustrate the exemplary structure of the training device 455 for the molecular property prediction model provided in the embodiments of this application as a software module. In some embodiments, such as... Figure 2 As shown, the software modules in the training device 455 of the molecular property prediction model stored in the memory 450 may include:

[0233] The sample acquisition module 4551 is used to mask the first non-geometric information of at least one atom of the sample molecule to obtain second non-geometric information; the sample acquisition module 4551 is also used to add noise to the first geometric information of at least one atom of the sample molecule to obtain second geometric information; the sample acquisition module 4551 is also used to combine the first non-geometric information of the atom that has not undergone the masking process, the second non-geometric information, the first geometric information of the atom that has not undergone the noise addition process, and the second geometric information into a pre-training data set; the first training module 4552 is used to call the molecular property prediction model to perform a first training task based on the pre-training data set to obtain the trained molecular property prediction model; the second training module 4553 is used to perform a second training task of molecular property prediction processing on the trained molecular property prediction model based on the original data set of the sample molecule to obtain the trained molecular property prediction model.

[0234] In some embodiments, the sample acquisition module 4551 is configured to traverse each atom in the sample molecule and perform the following processing on the traversed atom: select the traversed atom based on a first pre-configured probability; in response to the atom being selected as a first target atom, set the first atom type mask in the first non-geometric information of the first target atom to a pre-configured type to obtain the second non-geometric information of the first target atom.

[0235] In some embodiments, the sample acquisition module 4551 is configured to generate a first probability by performing random probability generation on the traversed atoms; when the first probability is less than the first pre-configured probability, the atom is determined to be selected as the first target atom; when the first probability is greater than or equal to the first pre-configured probability, the atom is determined not to be selected.

[0236] In some embodiments, the sample acquisition module 4551 is configured to traverse each atom in the sample molecule and perform the following processing on the traversed atom: select the traversed atom based on a second pre-configured probability; in response to the atom being selected as a second target atom, add pre-configured noise to the first coordinates in the first geometric information of the second target atom to obtain the second geometric information of the second target atom.

[0237] In some embodiments, the sample acquisition module 4551 is configured to generate a second probability by performing random probability generation on the traversed atoms; when the second probability is less than the second pre-configured probability, the atom is determined to be selected as the second target atom; when the second probability is greater than or equal to the second pre-configured probability, the atom is determined not to be selected.

[0238] In some embodiments, the first training module 4552 is configured to: invoke the molecular property prediction model to perform atom type prediction processing based on each of the first non-geometric information and each of the second non-geometric information; determine a first cross-entropy loss of the molecular property prediction model based on the result of the atom type prediction processing; invoke the molecular property prediction model to perform atom position prediction processing based on each of the first and second geometric information; determine a first regression loss of the molecular property prediction model based on the result of the atom position prediction processing; determine a first total loss of the molecular property prediction model based on the first cross-entropy loss and the first regression loss; and update the parameters of the molecular property prediction model through backpropagation based on the first total loss to obtain the trained molecular property prediction model.

[0239] In some embodiments, the first training module 4552 is configured to: call the molecular property prediction model to perform feature extraction processing based on each of the first non-geometric information and each of the second non-geometric information to obtain the first geometric features of the sample molecule; classify the mask atoms in the sample molecule based on the first geometric features to obtain the prediction type of each mask atom, wherein the mask atom is an atom used to perform the masking processing; and determine the first cross-entropy loss of the molecular property prediction model based on the prediction type of the mask atom and the first atom type.

[0240] In some embodiments, the first training module 4552 is configured to: call the molecular property prediction model to perform feature extraction processing based on each of the first geometric information and each of the second geometric information to obtain the first non-geometric features of the sample molecule; perform position prediction on the noisy atoms in the sample molecule based on the first non-geometric features to obtain the predicted coordinates of each mask atom, wherein the noisy atom is an atom used to perform the noisy processing; and determine the first regression loss of the molecular property prediction model based on the predicted coordinates of the noisy atom and the first coordinates.

[0241] In some embodiments, the sample acquisition module 4551 is configured to use the atom used to perform the masking process as the mask atom and the atom used to perform the noise-adding process as the noise-adding atom; wherein the noise-adding atom and the mask atom are the same atom, or the noise-adding atom and the mask atom are different atoms.

[0242] In some embodiments, the original dataset includes sample molecule data and the actual molecular properties of the sample molecules; the second training module 4553 is used to, based on the sample molecule data, call the trained molecular property prediction model for classification processing to obtain the first predicted probability of the sample molecule for different molecular properties; select the molecular property with the highest first predicted probability as the predicted molecular property of the sample molecule; determine the second cross-entropy loss of the molecular property prediction model based on the predicted molecular property and the actual molecular property; and update the parameters of the molecular property prediction model through backpropagation based on the second cross-entropy loss to obtain the trained molecular property prediction model.

[0243] In some embodiments, the original dataset further includes sample molecule data, the actual molecular properties of the sample molecules, and the actual properties of the atomic groups in the sample molecules; the second training module 4553 is used to, based on the sample molecule data, call the trained molecular property prediction model to perform feature extraction processing to obtain the molecular features of the sample molecules; perform molecular property classification processing based on the molecular features to obtain the predicted molecular properties of the sample molecules, and determine the second cross-entropy loss of the molecular property prediction model based on the predicted molecular properties and the actual molecular properties; perform atomic group property classification processing based on the molecular features to obtain the predicted properties of the atomic groups in the sample molecules, and determine the third cross-entropy loss of the molecular property prediction model based on the predicted properties of the atomic groups and the actual properties; determine the second total loss based on the second cross-entropy loss and the third cross-entropy loss; and update the parameters of the molecular property prediction model through backpropagation based on the second total loss to obtain the trained molecular property prediction model.

[0244] In some embodiments, the second training module 4553 is used to call the trained molecular property prediction model to perform multi-level feature extraction processing based on the sample molecule data to obtain feature information of each atom in the sample molecule, wherein the output of each level of feature extraction processing is the input of the next level of feature extraction processing; and to perform aggregation processing on the feature information of each atom to obtain the molecular features of the sample molecule.

[0245] In some embodiments, the second training module 4553 is used to perform a second training task on the trained molecular property prediction model based on the original data set of the sample molecules to obtain a trained molecular property prediction model. After obtaining the trained molecular property prediction model, the module performs feature extraction processing on the molecular data of the molecule to be detected based on the trained molecular property prediction model to obtain a second geometric feature and a second non-geometric feature of the molecule to be detected. Based on the second geometric feature and the second non-geometric feature, the module performs classification processing on the molecule to be detected to obtain a second prediction probability of the molecule to be detected for different molecular properties. The module selects at least one molecular property with the highest second prediction probability as the target molecular property of the molecule to be detected, wherein the types of the molecular property include physical properties and chemical properties.

[0246] In some embodiments, the molecular property prediction model includes: a graph neural network, a first prediction module, a second prediction module, and a molecular property prediction module, wherein the graph neural network is used to extract features from molecular data, and the molecular data includes geometric information and non-geometric information; the first prediction module is used to predict the first non-geometric information corresponding to the second non-geometric information, the second prediction module is used to predict the first geometric information corresponding to the second geometric information, and the molecular property prediction module is used to predict molecular properties.

[0247] This application provides a computer program product comprising a computer program or computer-executable instructions stored in a computer-readable storage medium. An electronic device's processor reads the computer program or computer-executable instructions from the computer-readable storage medium and executes the computer program or computer-executable instructions, causing the electronic device to perform the molecular property prediction model training method described above in this application.

[0248] This application provides a computer-readable storage medium storing computer-executable instructions or a computer program. When the computer-executable instructions or the computer program are executed by a processor, the processor will execute the training method of the molecular property prediction model provided in this application. For example, ... Figure 3A The training method for the molecular property prediction model is shown.

[0249] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.

[0250] In some embodiments, computer-executable instructions may take the form of programs, software, software modules, scripts, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as stand-alone programs or as modules, components, subroutines, or other units suitable for use in a computing environment.

[0251] As an example, computer-executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple co-located files (e.g., files that store one or more modules, subroutines, or code sections).

[0252] As an example, computer-executable instructions can be deployed to execute on a single electronic device, or on multiple electronic devices located at one location, or on multiple electronic devices distributed across multiple locations and interconnected via a communication network.

[0253] In summary, by masking the non-geometric information of sample molecules and adding noise to the geometric information of sample molecules through the embodiments of this application, and combining them into a pre-training dataset, the replication degree of the sample data is increased. Based on the processed sample data, the processing of actual data can be performed to predict the actual data. This enhances the molecular property prediction model's ability to identify and predict the geometric and non-geometric information of atoms in molecules during the first training task. Furthermore, by training the parameters of the pre-trained molecular property prediction model to perform molecular property prediction processing, the accuracy of the molecular property prediction model in predicting molecular properties can be improved. Simultaneously, since the molecular property prediction model has learned the basic knowledge of the geometric and non-geometric information of atoms in molecules after the first training task, the convergence speed of the molecular property prediction model is accelerated when the second training task specifically trains the molecular property prediction function, thereby saving training time and computational resources required for training.

[0254] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A training method for a molecular property prediction model, characterized in that, The method includes: The first non-geometric information of at least one atom in the sample molecule is masked to obtain the second non-geometric information; The first geometric information of at least one atom of the sample molecule is subjected to noise processing to obtain the second geometric information; The first non-geometric information of the atoms that have not undergone the masking process, the second non-geometric information, the first geometric information of the atoms that have not undergone the noise-adding process, and the second geometric information are combined into a pre-training data set. Based on the pre-trained dataset, the molecular property prediction model is invoked for the first training task to obtain the trained molecular property prediction model. Based on the original data set of the sample molecules, a second training task is performed on the trained molecular property prediction model to obtain the trained molecular property prediction model. Based on the trained molecular property prediction model, feature extraction processing is performed on the molecular data of the molecule to be detected to obtain the second geometric feature and the second non-geometric feature of the molecule to be detected. The molecules to be detected are classified based on the second geometric feature and the second non-geometric feature to obtain the second predicted probability of the molecules to be detected for different molecular properties. At least one molecular property with the highest second predicted probability is selected as the target molecular property of the molecule to be detected, wherein the types of the molecular property include physical properties and chemical properties.

2. The method according to claim 1, characterized in that, The process of masking the first non-geometric information of at least one atom of the sample molecule to obtain the second non-geometric information includes: For each atom in the sample molecule, the following processing is performed on the encountered atom: The atoms that are traversed are selected based on a first pre-configured probability; In response to the atom being selected as the first target atom, the first atom type mask in the first non-geometric information of the first target atom is set to a pre-configured type to obtain the second non-geometric information of the first target atom.

3. The method according to claim 2, characterized in that, The selection operation of the traversed atoms based on the first pre-configured probability includes: For each atom encountered during the traversal, a random probability is generated to obtain a first probability; When the first probability is less than the first pre-configured probability, the atom is determined to be selected as the first target atom; When the first probability is greater than or equal to the first pre-configured probability, it is determined that the atom is not selected.

4. The method according to claim 1, characterized in that, The step of adding noise to the first geometric information of at least one atom of the sample molecule to obtain second geometric information includes: For each atom in the sample molecule, the following processing is performed on the encountered atom: The atoms that are traversed are selected based on a second pre-configured probability; In response to the atom being selected as the second target atom, pre-configured noise is added to the first coordinates in the first geometric information of the second target atom to obtain the second geometric information of the second target atom.

5. The method according to claim 4, characterized in that, The selection of the traversed atoms based on the second pre-configured probability includes: A second probability is obtained by randomly generating a probability for each of the atoms that have been traversed. When the second probability is less than the second pre-configured probability, the atom is determined to be selected as the second target atom; When the second probability is greater than or equal to the second pre-configured probability, it is determined that the atom is not selected.

6. The method according to claim 1, characterized in that, The step of calling the molecular property prediction model based on the pre-trained data set to perform the first training task, and obtaining the trained molecular property prediction model, includes: Based on each of the first non-geometric information and each of the second non-geometric information, the molecular property prediction model is invoked to perform atom type prediction processing, and the first cross-entropy loss of the molecular property prediction model is determined based on the result of the atom type prediction processing. Based on each of the first geometric information and each of the second geometric information, the molecular property prediction model is invoked to perform atomic position prediction processing, and the first regression loss of the molecular property prediction model is determined based on the result of the atomic position prediction processing. Based on the first cross-entropy loss and the first regression loss, the first total loss of the molecular property prediction model is determined; Based on the first total loss, the parameters of the molecular property prediction model are updated through backpropagation to obtain the trained molecular property prediction model.

7. The method according to claim 6, characterized in that, The step of invoking the molecular property prediction model to perform atom type prediction processing based on each of the first non-geometric information and each of the second non-geometric information, and determining the first cross-entropy loss of the molecular property prediction model based on the result of the atom type prediction processing, includes: Based on each of the first non-geometric information and each of the second non-geometric information, the molecular property prediction model is invoked to perform feature extraction processing to obtain the first geometric features of the sample molecule. Based on the first geometric feature, the mask atoms in the sample molecule are classified to obtain the predicted type of each mask atom, wherein the mask atom is an atom used to perform the masking process; Based on the predicted type of the masked atom and the first atom type, the first cross-entropy loss of the molecular property prediction model is determined.

8. The method according to claim 6, characterized in that, The step of invoking the molecular property prediction model to perform atomic position prediction processing based on each of the first geometric information and each of the second geometric information, and determining the first regression loss of the molecular property prediction model based on the result of the atomic position prediction processing, includes: Based on each of the first geometric information and each of the second geometric information, the molecular property prediction model is invoked to perform feature extraction processing to obtain the first non-geometric features of the sample molecule; Based on the first non-geometric feature, the position prediction of the noisy atoms in the sample molecule is performed to obtain the predicted coordinates of each mask atom, wherein the noisy atom is the atom used to perform the noisy processing; Based on the predicted coordinates of the noisy atom and the first coordinate, the first regression loss of the molecular property prediction model is determined.

9. The method according to any one of claims 1 to 8, characterized in that, The method further includes: The atom used to perform the masking process is used as the mask atom, and the atom used to perform the noise-adding process is used as the noise-adding atom; wherein the noise-adding atom and the mask atom are the same atom, or the noise-adding atom and the mask atom are different atoms.

10. The method according to any one of claims 1 to 8, characterized in that, The original dataset includes sample molecule data and the actual molecular properties of the sample molecules; The second training task, which involves processing the trained molecular property prediction model with molecular property prediction based on the original dataset of the sample molecules, to obtain the trained molecular property prediction model, includes: Based on the sample molecule data, the trained molecular property prediction model is invoked for classification processing to obtain the first predicted probability of the sample molecule for different molecular properties. The molecular property with the highest predicted probability is selected as the predicted molecular property of the sample molecule; The second cross-entropy loss of the molecular property prediction model is determined based on the predicted molecular properties and the actual molecular properties. Based on the second cross-entropy loss, the parameters of the molecular property prediction model are updated through backpropagation to obtain the trained molecular property prediction model.

11. The method according to any one of claims 1 to 8, characterized in that, The original dataset also includes sample molecule data, the actual molecular properties of the sample molecules, and the actual properties of the atomic groups in the sample molecules; The second training task, which involves processing the trained molecular property prediction model with molecular property prediction based on the original dataset of the sample molecules, to obtain the trained molecular property prediction model, includes: Based on the sample molecule data, the trained molecular property prediction model is invoked to perform feature extraction processing to obtain the molecular features of the sample molecules. Based on the molecular features, molecular property classification is performed to obtain the predicted molecular properties of the sample molecules, and the second cross-entropy loss of the molecular property prediction model is determined based on the predicted molecular properties and the actual molecular properties. Based on the molecular characteristics, the properties of atomic groups are classified to obtain the predicted properties of atomic groups in the sample molecule, and the third cross-entropy loss of the molecular property prediction model is determined based on the predicted properties of the atomic groups and the actual properties. Based on the second cross-entropy loss and the third cross-entropy loss, a second total loss is determined; Based on the second total loss, the parameters of the molecular property prediction model are updated through backpropagation to obtain the trained molecular property prediction model.

12. The method according to claim 11, characterized in that, The step of using the sample molecule data to call the trained molecular property prediction model for feature extraction to obtain the molecular features of the sample molecules includes: Based on the sample molecule data, the trained molecular property prediction model is invoked to perform feature extraction processing at multiple levels to obtain the feature information of each atom in the sample molecule. The output of each level of feature extraction processing is the input of the next level of feature extraction processing. The feature information of each atom is aggregated to obtain the molecular features of the sample molecule.

13. The method according to any one of claims 1 to 8, characterized in that, The molecular property prediction model includes: a graph neural network, a first prediction module, a second prediction module, and a molecular property prediction module. The graph neural network is used to extract features from molecular data, which includes geometric information and non-geometric information. The first prediction module is used to predict the first non-geometric information corresponding to the second non-geometric information. The second prediction module is used to predict the first geometric information corresponding to the second geometric information. The molecular property prediction module is used to predict molecular properties.

14. A training device for a molecular property prediction model, characterized in that, The device includes: The sample acquisition module is used to mask the first non-geometric information of at least one atom of the sample molecule to obtain the second non-geometric information. The sample acquisition module is further configured to add noise to the first geometric information of at least one atom of the sample molecule to obtain the second geometric information; The sample acquisition module is further configured to combine the first non-geometric information, the second non-geometric information, the first geometric information, and the second geometric information of atoms that have not undergone the masking process into a pre-training data set; The first training module is used to call the molecular property prediction model to perform a first training task based on the pre-training data set, so as to obtain the trained molecular property prediction model. The second training module is used to perform a second training task on the trained molecular property prediction model based on the original data set of the sample molecules, so as to obtain the trained molecular property prediction model. The second training module is used to perform feature extraction processing on the molecular data of the molecule to be detected based on the trained molecular property prediction model, to obtain the second geometric feature and the second non-geometric feature of the molecule to be detected; to classify the molecule to be detected based on the second geometric feature and the second non-geometric feature, to obtain the second prediction probability of the molecule to be detected for different molecular properties; and to select at least one molecular property with the highest second prediction probability as the target molecular property of the molecule to be detected, wherein the types of the molecular property include: physical properties and chemical properties.

15. An electronic device, characterized in that, The electronic device includes: Memory is used to store executable instructions or computer programs. A processor, when executing computer-executable instructions or computer programs stored in the memory, implements the training method for the molecular property prediction model according to any one of claims 1 to 13.

16. A computer-readable storage medium storing computer-executable instructions or a computer program, characterized in that, When the computer-executable instructions or computer program are executed by a processor, the training method for the molecular property prediction model according to any one of claims 1 to 13 is implemented.

17. A computer program product comprising computer-executable instructions or a computer program, characterized in that, When the computer-executable instructions or computer program are executed by a processor, the training method for the molecular property prediction model according to any one of claims 1 to 13 is implemented.

Images