Molecular attribute prediction method and system based on multistage knowledge distillation

Abstract

The invention discloses a molecular attribute prediction method and system based on multistage knowledge distillation, and belongs to the technical field of molecular representation learning and molecular property prediction. Comprising the steps of obtaining molecular representation symbols, and determining molecular text prompt features by using a large language model based on the molecular representation symbols; obtaining a molecular topological graph, and processing the molecular topological graph and the molecular text prompt features through a trained molecular attribute prediction model to obtain a molecular attribute prediction result; wherein when the encoder in the molecular attribute prediction model is trained, multi-stage knowledge distillation is carried out in combination with multi-stage actions of molecules. The molecular structure is comprehensively analyzed, comprehensive understanding of molecules can be deepened, and the problem that existing molecular structure analysis is limited by task requirements and local information processing is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of molecular representation learning and molecular property prediction, and in particular to a molecular property prediction method and system based on multi-level knowledge distillation. Background Art

[0002] The statements in this section merely mention background art related to the present invention and do not necessarily constitute prior art.

[0003] Accurate prediction of molecular properties (such as biological activity and toxicity) is crucial in the drug development process and determines whether candidate compounds can successfully pass clinical trials. Traditional wet experimental methods are not only time-consuming and lengthy, but also require a lot of money and resource investment. Today, with the help of artificial intelligence technology, researchers can develop molecular property prediction models to quickly screen and evaluate the therapeutic potential and clinical trial success rate of candidate drugs. This AI-based approach not only reduces the time and cost of relying on experiments, but also significantly speeds up the drug development process and reduces the risk of high-cost failures in the later stages due to ineffectiveness or toxicity.

[0004] The inventors found that although existing methods have made significant progress in the field of molecular property prediction, there are still many challenges:

[0005] (1) Traditional methods usually rely only on the geometric shape of a single molecule when obtaining molecular representations. Although these methods can utilize the structural properties of molecules for analysis, they are difficult to provide personalized hints required to adapt to specific tasks.

[0006] (2) Existing methods are usually limited to processing local information in molecular graphs, such as node or edge-level features. Therefore, these methods tend to localize molecules when processing molecular structures and fail to fully consider the overall characteristics and global structure of molecules.

[0007] (3) Current methods often rely on a single learner to complete downstream tasks, but this single learning strategy is prone to overfitting, thus limiting the effective generalization ability of the model. Summary of the invention

[0008] In order to address the deficiencies of the prior art, the present invention provides a molecular property prediction method, system, electronic device, computer-readable storage medium and computer program product based on multi-level knowledge distillation, which can effectively integrate the information of atoms, bonds, substructures and overall molecules, and make full use of the text prompts provided by ChatGPT, so as to better understand the structure and properties of molecules.

[0009] In a first aspect, the present invention provides a molecular property prediction method based on multi-level knowledge distillation;

[0010] A molecular property prediction method based on multi-level knowledge distillation, comprising:

[0011] Acquire molecular representation symbols, and determine molecular text prompt features using a large language model based on the molecular representation symbols;

[0012] Obtaining a molecular topology map, and processing the molecular topology map and the molecular text prompt features through a trained molecular property prediction model to obtain a molecular property prediction result;

[0013] Wherein, when training the encoder in the molecular property prediction model, multi-level knowledge distillation is performed in combination with the multi-level actions of the molecules.

[0014] In some embodiments, the method of determining the molecular text prompt feature based on the molecular representation symbol using the large language model is specifically as follows: the molecular representation symbol and a preset user input template are input into ChatGPT, the text representation is obtained and encoded using the word2vec encoder to generate the molecular text prompt feature.

[0015] In some embodiments, the processing of the molecular topology map and the segmented text prompt features by the trained molecular property prediction model includes:

[0016] The chemical bonds in the molecular topology are encoded through the edge-level module, the atoms in the molecular topology are encoded through the point-level module, the functional groups in the molecular topology are encoded through the subgraph-level module, the encoding results are integrated across levels through the graph-level module, and the molecular-level features are updated in combination with the molecular text prompt features;

[0017] The updated molecular features are processed by the sigmoid function to obtain the molecular property prediction results.

[0018] In some embodiments, encoding chemical bonds in a molecule by an edge-level module, encoding atoms in a molecule by a point-level module, and encoding functional groups in a molecule by a subgraph-level module include:

[0019] Update the bond-level features using the atomic-level features of the source and target atoms and aggregating information from the updated graph-level features;

[0020] Using the updated bond-level features and the features of the corresponding source atoms, combined with the updated molecular-level features, the atomic-level features are updated;

[0021] Combine the neighbor subgraph structure and update the substructure level features.

[0022] In some embodiments, the cross-level fusion of the encoding results through the graph-level module and the updating of the molecular-level features in combination with the molecular text prompt features are specifically as follows: based on the atomic-level initial features, the bond-level initial features, the substructure-level initial features and the molecular-level initial features, the messages transmitted by the atomic-level features, the bond-level features and the substructure-level features are aggregated, and in combination with the molecular text prompt features, the molecular-level features are updated through a multi-layer perceptron.

[0023] In some embodiments, when training the encoder in the molecular property prediction model, performing multi-level knowledge distillation in combination with the multi-level actions of the molecule includes:

[0024] Construct multiple encoders and connect the probability distributions of molecules, bonds, atoms, and substructures in the neural network with the cross-entropy loss of the encoder to obtain the multi-level states of the molecule;

[0025] Molecular-level actions, bond-level actions, atomic-level actions, and substructure-level actions are set up, and KL divergence is combined to measure the distance between the probability distributions of the same elements of multiple neural networks to minimize the loss of each encoder at different levels.

[0026] In a second aspect, the present invention provides a molecular property prediction system based on multi-level knowledge distillation;

[0027] A molecular property prediction system based on multi-level knowledge distillation, comprising:

[0028] The molecular feature enhancement module is configured to: obtain molecular representation symbols, and determine molecular text prompt features based on the molecular representation symbols using a large language model;

[0029] The molecular property prediction module is configured to: obtain a molecular topology map, process the molecular topology map and the molecular text prompt features through a trained molecular property prediction model, and obtain a molecular property prediction result;

[0030] Wherein, when training the encoder in the molecular property prediction model, multi-level knowledge distillation is performed in combination with the multi-level actions of the molecules.

[0031] In a third aspect, the present invention provides an electronic device;

[0032] An electronic device comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the molecular property prediction method based on multi-level knowledge distillation.

[0033] In a fourth aspect, the present invention provides a computer-readable storage medium;

[0034] A computer-readable storage medium stores a computer program / instruction, which, when executed by a processor, implements the steps of the molecular property prediction method based on multi-level knowledge distillation.

[0035] In a fifth aspect, the present invention provides a computer program product;

[0036] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned molecular property prediction method based on multi-level knowledge distillation.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. The technical solution provided by the present invention generates molecular prompts for specific task requirements through ChatGPT, which can accurately provide text descriptions of functional groups, chemical properties and potential drug applications, and can assist in generating two-dimensional molecular representations and enhance molecular feature representations.

[0039] 2. The technical solution provided by the present invention provides a novel encoder DouMPNN, which can effectively collect information across nodes, edges, subgraphs and the entire graph, allowing a comprehensive analysis of molecular structures, from microscopic atoms and bonds to macroscopic substructures; in addition, DouMPNN further deepens the comprehensive understanding of molecules by integrating text descriptions.

[0040] 3. The technical solution provided by the present invention provides a new multi-layer knowledge distillation architecture, which implements a free-direction knowledge distillation strategy through a hierarchical reinforcement learning algorithm, and can dynamically manage the knowledge transfer direction from the node, edge, subgraph and layer levels. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0042] Figure 1 An example diagram of CaP acquisition provided by an embodiment of the present invention;

[0043] Figure 2 An example diagram of the encoder DouMPNN provided in an embodiment of the present invention;

[0044] Figure 3 A schematic diagram of a flow chart of a molecular property prediction method based on multi-level knowledge distillation provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0046] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "include" and "have" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0047] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0048] Embodiment 1

[0049] Existing molecular property prediction methods are difficult to adapt to the personalized prompts required for specific tasks and do not consider the overall characteristics and global structure of molecules; therefore, the present invention provides a molecular property prediction method based on multi-level knowledge distillation, which generates molecular prompts for specific task requirements through ChatGPT, and can accurately provide text descriptions of functional groups, chemical properties and potential drug applications. The encoder is designed as DouMPNN, which can effectively integrate the information of atoms, bonds, substructures and overall molecules, and make full use of the text prompts provided by ChatGPT, so as to better understand the structure and properties of molecules; in order to improve the generalization ability of molecular representation learning, reduce overfitting and promote knowledge sharing, a multi-level distillation strategy is adopted to allow two learners to interactively learn at all levels of molecules.

[0050] Next, combine Figure 1-Figure 3 , a molecular property prediction method based on multi-level knowledge distillation disclosed in this embodiment is described in detail. The molecular property prediction method based on multi-level knowledge distillation includes the following steps:

[0051] S1. Obtain molecular representation symbols, and based on the molecular representation symbols, use a large language model to determine molecular text prompt features and associate them with the molecular representation symbols.

[0052] In this embodiment, the molecule representation symbol is a molecule SMILES symbol.

[0053] Specifically, the molecular SMILES symbols and the preset user input template are input into ChatGPT, the text representation is obtained and encoded using the word2vec encoder to generate the molecular text prompt feature.

[0054] Combination Figure 1 The user input template includes role identification and task description. The specific description of role identification is: Assume that you are a chemistry student taking an exam; the task description is: Given a series of chemical molecules represented by SMILES, please give the chemical bonds and functional groups contained in each molecule and briefly describe their chemical properties. You need to provide answers in the following format:

[0055] Original SNILES:

[0056] Chemical Bond:

[0057] Functional Group:

[0058] Structural features:

[0059] Please note that each question above is answered in one sentence. The following is the SMILES string: xxxxxxxxxx.

[0060] In subsequent applications, only the molecule SMILES string needs to be replaced.

[0061] S2. Obtain a molecular topology map, process the molecular topology map using a trained molecular property prediction model, and obtain a molecular property prediction result.

[0062] Among them, the molecular topology map includes atomic-level initial features (points in the molecular topology map), bond-level initial features (edges in the molecular topology map), substructure-level initial features (subgraphs in the molecular topology map) and molecular-level initial features (molecular topology map), which can be directly applied during processing.

[0063] In this embodiment, the molecular property prediction model includes an encoder and a sigmoid function connected in sequence, and the encoder is a novel encoder DouMPNN. Through the processing of the molecular topology graph by DouMPNN, it can effectively collect information across nodes, edges, subgraphs and the entire graph, allowing a comprehensive analysis of the molecular structure, and merging the initial molecular features with the molecular text prompt features during the molecular level feature update process to obtain the updated molecular features. As an implementation method, S2 specifically includes:

[0064] S201. Encode the chemical bonds in the molecular topology map through the edge-level module, encode the atoms in the molecular topology map through the point-level module, and encode the functional groups in the molecular topology map through the subgraph-level module; cross-level fusion of the encoding results through the graph-level module, input the cross-level fusion results and the molecular text prompt features into the multi-layer perceptron for processing, and update the molecular-level features.

[0065] For example, the specific process is as follows:

[0066] (1) Update molecular-level features

[0067] The features at the molecular level are updated by aggregating the features at each level, expressed as:

[0068]

[0069] Among them, m(G)' λ is the updated molecular-level feature, N(G)λ, F(G) λ and H(G)λ represent the messages transmitted from atomic-level features, bond-level features, and substructure-level features, respectively. The symbol ⊙ represents the connection function. n(v), f(e), h(s), and m(G)λ represent the initial features at the atomic, bond, substructure, and molecular levels, respectively. t represents the molecular text prompt feature, and G represents the molecular topology map.

[0070] (2) Update key-level features

[0071] The bond-level features are updated by using the atomic-level features of the source and target atoms and aggregating information from the updated graph-level features. In short, these feature representations are concatenated with the previous bond-level features and the concatenated representation is used as the input of a multi-layer perceptron (MLP) to complete the update of the bond-level features.

[0072] Specifically, the steps are as follows:

[0073]

[0074] where [·] denotes the concatenation function and f'(e) denotes the updated bond-level representation.

[0075] (3) Update atomic-level features

[0076] For each atom, the information it receives is calculated using the updated bond-level features of the incoming bond and the features of the corresponding source atom. This information is then combined with the updated molecular-level features to update the atomic-level features. This is expressed as:

[0077]

[0078] n'(v)λ=MLP([n(v)λ,O(v)λ,m'(G)λ]);

[0079] Among them, n'(v) λ represents the updated features of atom v, O(v)λ represents the information transferred from bond-level features to atomic-level features, and t e represents the target atom of the bond, and se represents the source atom of the bond.

[0080] (4) Update substructure level features

[0081] The information needed to update the substructure comes from the neighbor substructure, expressed as:

[0082]

[0083]

[0084] Where N(s) represents the set of neighbor substructures of substructure s, h'(s) λ represents the updated feature representation of substructure s, and P(s)λ represents the message passed from atomic-level features to substructure-level features.

[0085] As an implementation mode, before the encoder DouMPNN is applied, it also includes:

[0086] Two identical encoders DouMPNN were constructed and trained using a multi-layer knowledge distillation architecture. A free-direction knowledge distillation strategy was implemented through a hierarchical reinforcement learning algorithm, which can dynamically manage the direction of knowledge transfer from nodes, edges, subgraphs, and layers. After training, the two encoders DouMPNN used for knowledge distillation were tested separately, and the one with the best test performance was used for the final molecular property prediction.

[0087] Furthermore, the specific process is as follows:

[0088] Step 1, construct two neural networks DouMPNN_A and DouMPNN_B, and connect the probability distributions of molecules, bonds, atoms and substructures in the neural networks DouMPNN_A and DouMPNN_B with the cross entropy losses of the neural networks DouMPNN_A and DouMPNN_B at different levels to obtain the multi-level state of the molecules.

[0089] For example, the following items are concatenated as the state vectors for molecule i, bond i, atom i, and substructure i (1) The probability distribution of molecule i, bond i, atom i and substructure i in network λ; (2) The cross entropy loss of network λ; (3) The probability distribution of molecule i, bond i, atom i and substructure i in network η; (4) The cross entropy loss of network η. It can be expressed as:

[0090]

[0091]

[0092] Step 2: Set up molecular-level actions, bond-level actions, atomic-level actions, and substructure-level actions.

[0093] Specifically, use and Represents actions at the molecular level, bond level, atomic level, and substructure level, respectively.

[0094] when a G = 0, it means transferring the knowledge of molecules from network λ to network η; when a G = 1, it means transferring the knowledge of molecules from network η to network λ. When , it means transferring the knowledge of key i from network λ to network η; when When , it means transferring the knowledge of key i from network η to network λ. When , it means transferring the knowledge of node i from network λ to network η; when When , it means transferring the knowledge of node i from network η to network λ. When , it means transferring the knowledge of substructure i from network λ to network η; when , it means transferring the knowledge of substructure i from network η to network λ.

[0095] Step 3: Combine KL divergence to measure the distance between the probability distributions of the same elements of multiple neural networks to minimize the loss of each neural network at different levels.

[0096] In order to transfer multi-level knowledge simultaneously, in this embodiment, KL divergence is used to measure the distance between the probability distributions of the same elements in two networks, and a new loss function is proposed to minimize each network at different levels. It is expressed as:

[0097] (1) Molecular level loss:

[0098]

[0099] (2) Bond level loss:

[0100]

[0101] (3) Atomic level loss:

[0102]

[0103] (4) Substructure level loss:

[0104]

[0105] In the formula, Represent the probability distributions of the models λ and η at the molecular, bond, atomic, and substructure levels, respectively.

[0106]

[0107]

[0108] S202, processing the updated molecular-level features through a sigmoid function to obtain molecular property prediction results.

[0109] Here, the molecular property prediction result is the molecular property and the corresponding prediction probability value, and the molecular property includes toxicity, water solubility and fat solubility, etc. And the use of sigmoid function for probability prediction is a prior art, which is not improved in this embodiment and will not be repeated here.

[0110] Embodiment 2

[0111] This embodiment discloses a molecular property prediction system based on multi-level knowledge distillation, including:

[0112] The molecular feature enhancement module is configured to: obtain molecular representation symbols, and determine molecular text prompt features based on the molecular representation symbols using a large language model;

[0113] The molecular property prediction module is configured to: obtain a molecular topology map, process the molecular topology map and the molecular text prompt features through a trained molecular property prediction model, and obtain a molecular property prediction result;

[0114] Wherein, when training the encoder in the molecular property prediction model, multi-level knowledge distillation is performed in combination with the multi-level actions of the molecules.

[0115] It should be noted that the above-mentioned molecular feature enhancement module and molecular property prediction module correspond to the steps in Embodiment 1, and the examples and application scenarios implemented by the above-mentioned modules and the corresponding steps are the same, but are not limited to the contents disclosed in the above-mentioned Embodiment 1. It should be noted that the above-mentioned modules, as part of the system, can be executed in a computer system such as a set of computer executable instructions.

[0116] Embodiment 3

[0117] Embodiment 3 of the present invention provides an electronic device, including a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps of the above-mentioned molecular property prediction method based on multi-level knowledge distillation are completed.

[0118] Embodiment 4

[0119] Embodiment 4 of the present invention provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps of the above-mentioned molecular property prediction method based on multi-level knowledge distillation are completed.

[0120] Embodiment 5

[0121] Embodiment 5 of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned molecular property prediction method based on multi-level knowledge distillation.

[0122] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0123] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0124] These computer program instructions can also be loaded onto a computer or other programmable data processing device, and a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0125] The description of each embodiment in the above embodiments has different emphases. For parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0126] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A molecular property prediction method based on multi-level knowledge distillation, characterized in that: include: Acquire molecular representation symbols, and determine molecular text prompt features using a large language model based on the molecular representation symbols; Obtaining a molecular topology map, and processing the molecular topology map and the molecular text prompt features through a trained molecular property prediction model to obtain a molecular property prediction result; Wherein, when training the encoder in the molecular property prediction model, multi-level knowledge distillation is performed in combination with the multi-level actions of the molecules.

2. The molecular property prediction method based on multi-level knowledge distillation according to claim 1, characterized in that: The method of determining the molecular text prompt feature based on the molecular representation symbol and using the large language model is specifically as follows: the molecular representation symbol and a preset user input template are input into ChatGPT, the text representation is obtained and encoded using the wrord2vec encoder to generate the molecular text prompt feature.

3. The molecular property prediction method based on multi-level knowledge distillation according to claim 1, characterized in that: The processing of the molecular topology map and the molecular text prompt features by using the trained molecular property prediction model includes: The chemical bonds in the molecular topology are encoded through the edge-level module, the atoms in the molecular topology are encoded through the point-level module, the functional groups in the molecular topology are encoded through the subgraph-level module, the encoding results are integrated across levels through the graph-level module, and the molecular-level features are updated in combination with the molecular text prompt features; The updated molecular features are processed by the sigmoid function to obtain the molecular property prediction results.

4. The molecular property prediction method based on multi-level knowledge distillation according to claim 3, characterized in that: The chemical bonds in the molecule are encoded through the edge-level module, the atoms in the molecule are encoded through the point-level module, and the functional groups in the molecule are encoded through the subgraph-level module, including: Update the bond-level features using the atomic-level features of the source and target atoms and aggregating information from the updated graph-level features; Using the updated bond-level features and the features of the corresponding source atoms, combined with the updated molecular-level features, the atomic-level features are updated; Combine the neighbor subgraph structure and update the substructure level features.

5. The molecular property prediction method based on multi-level knowledge distillation according to claim 3, characterized in that: The cross-level fusion of the encoding results through the graph-level module and the updating of the molecular-level features in combination with the molecular text prompt features are specifically as follows: based on the atomic-level initial features, the bond-level initial features, the substructure-level initial features and the molecular-level initial features, the messages transmitted by the atomic-level features, the bond-level features and the substructure-level features are aggregated, and in combination with the molecular text prompt features, the molecular-level features are updated through a multi-layer perceptron.

6. The molecular property prediction method based on multi-level knowledge distillation according to claim 1, characterized in that: When training the encoder in the molecular property prediction model, multi-level knowledge distillation is performed in combination with the multi-level actions of the molecule, including: Construct multiple neural networks and connect the probability distributions of molecules, bonds, atoms, and substructures in the neural networks with the cross entropy loss of the encoder to obtain the multi-level states of the molecules; Molecular-level actions, bond-level actions, atomic-level actions, and substructure-level actions are set up, and KL divergence is combined to measure the distance between the probability distributions of the same elements of multiple encoders to minimize the loss of each encoder at different levels.

7. A molecular property prediction system based on multi-level knowledge distillation, characterized in that: include: The molecular feature enhancement module is configured to: obtain molecular representation symbols, and determine molecular text prompt features based on the molecular representation symbols using a large language model; The molecular property prediction module is configured to: obtain a molecular topology map, process the molecular topology map and the molecular text prompt features through a trained molecular property prediction model, and obtain a molecular property prediction result; Wherein, when training the encoder in the molecular property prediction model, multi-level knowledge distillation is performed in combination with the multi-level actions of the molecules.

8. An electronic device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the molecular property prediction method based on multi-level knowledge distillation according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the steps of the molecular property prediction method based on multi-level knowledge distillation described in any one of claims 1 to 6 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the molecular property prediction method based on multi-level knowledge distillation described in any one of claims 1 to 6 are implemented.

Images