Small sample molecule attribute prediction method and device based on meta-training task sampling optimization

Abstract

The invention discloses a small sample molecule attribute prediction method and device based on meta-training task sampling optimization. The method comprises the following steps: constructing a molecule-attribute relation graph; constructing a candidate meta-task pool and a small sample molecule attribute model based on the obtained relation graph, and obtaining sub-graph representation; constructing a task sampler, sampling a target task, randomly sampling an auxiliary task, and reconstructing a sub-graph; performing representation learning on the sub-graphs by using a graph neural network, calculating loss, and updating model parameters; and performing parameter adaptation of target properties on the trained model to obtain a target attribute-oriented generalization model for molecular attribute prediction. By optimizing the meta-training task sampling module, the relation between meta-tasks is deeply mined, the training efficiency of the small sample molecule attribute prediction model is improved, and then the model prediction performance is optimized.

Description

Technical Field

[0001] The present invention relates to a small sample molecular property prediction method and device based on meta-training task sampling optimization, belonging to the field of molecular property prediction and graph neural network. Background Art

[0002] Molecular property prediction is one of the classic tasks widely used in many fields such as bioinformatics and quantum chemistry. Today, drug discovery is of great significance to public health, and deep learning plays an important role in it. In the early stages of drug development, researchers need to select a large number of candidate molecules and perform virtual screening to determine whether they have specific properties. In fact, only a small number of molecules can be prepared and tested in wet labs. The limited amount of labeled data usually hinders the generalization ability of deep learning models in practical applications. This is essentially a small sample learning problem. In recent years, researchers have explored the problem of small sample molecular property prediction and proposed a series of small sample molecular property prediction models and optimization methods, such as designing novel neural network structures, training frameworks, or using parameter updates and attention to optimize algorithms. Due to the success of meta-training tasks in meta-learning, this training mechanism has gradually been applied to small sample molecular property prediction.

[0003] Regarding the application of meta-training tasks in the field of small sample molecular property prediction, researchers believe that meta-training tasks play an important role in prediction. Some studies believe that meta-training tasks are independent of each other and randomly sample meta-training tasks with equal probability. However, other studies believe that meta-training tasks are not independent, and have designed a meta-training task sampler with reinforcement learning and introduced contrastive learning to enhance task representation learning. However, these studies have failed to capture the correlation between tasks. Task correlation is not only an important factor in meta-training task scheduling, but also plays a vital role in the construction of relationship subgraphs. Existing small sample molecular property prediction methods, such as a method for predicting molecular properties disclosed in Publication No. CN117219186A, which proposes a 3D-metaMRG small sample molecular property prediction framework, uses contrastive and generative self-supervised learning pre-training and iterative molecular relationship graph learning based on meta-learning to predict molecular properties; Publication No. CN115831260A discloses a small sample molecular toxicity prediction method based on transfer learning strategy and graph neural network model, which converts compounds in toxicity data into graph structure data, and performs model pre-training and model migration based on a multi-classification module graph neural network model to obtain a small sample molecular toxicity prediction model; Publication No. CN118506902A discloses a molecular property prediction method based on a metric small sample learning method, which uses a prototype network framework combined with a graph pre-training strategy, low-rank representation, contrastive learning and an optimized Sinkhorn K means algorithm to predict molecular properties. The above-mentioned technologies fail to explore the correlation between tasks, resulting in insufficient information utilization of data, thereby reducing the prediction performance of molecular property prediction. Therefore, a small-sample molecular property prediction method based on meta-training task sampling optimization is needed to capture the relationship between tasks, make full use of relevant information, and optimize the model.

[0004] Therefore, a small sample molecular property prediction method based on meta-training task sampling optimization is needed. This method optimizes the meta-training task sampling module, and the optimized task sampler captures the relationship between tasks. Summary of the invention

[0005] The present invention aims to overcome the above-mentioned shortcomings of the prior art and provides a method and device for predicting molecular properties of small samples based on meta-training task sampling optimization. The present invention optimizes the task sampler to improve the accuracy of the molecular property prediction model.

[0006] A first aspect of the present invention relates to a small sample molecular property prediction method based on meta-training task sampling optimization, comprising the following steps:

[0007] S1: constructing a molecule-attribute relationship diagram;

[0008] S2: Based on the relationship graph obtained in step S1, a candidate meta-task pool and a small sample molecular attribute model are constructed, and a subgraph representation is obtained;

[0009] S3: Build a task sampler to sample the target task and randomly sample auxiliary tasks to reconstruct the subgraph;

[0010] S4: Use graph neural network to learn the representation of subgraphs, calculate losses, and update model parameters;

[0011] S5: Adapt the parameters of the target attribute to the model trained in step S4, and then obtain a generalized model oriented to the target attribute for molecular attribute prediction.

[0012] Preferably, in step S1, a molecule-attribute relationship graph is constructed for each molecule, specifically as follows: molecules and attributes are taken as nodes, and the molecules and their corresponding known attributes are connected by edges to form a relationship graph represents the node set, ε represents the edge set, in Represented as a molecular node set, x m is the molecular node, Represents a set of attribute nodes, p t For attribute nodes.

[0013] In step S2, based on the relationship graph obtained in step S1, a candidate meta-task pool and a small sample molecular attribute model are constructed, and a subgraph representation is obtained, as follows:

[0014] S2.1: Associate the sampled subgraphs in the molecule-attribute relationship graph with meta-tasks; associate each meta-task E t Represented as a task subgraph Will support set and queryset are expressed as corresponding subgraphs, supporting the set Expressed as Query Set Expressed as

[0015] S2.2: Next, merge the support set and queryset The subgraph of meta-task E t Represented as a task subgraph

[0016]

[0017] S2.3: Construct a pool of candidate meta-task subgraphs. According to the contrastive learning setting, randomly sample the pool of candidate meta-task subgraphs to form a pool of candidate meta-tasks. Each task contains two subgraphs with the same target attribute, namely: in Refers to two different representations of the target task t; then a small sample molecular property model is constructed, including a graph neural network and a multi-layer perceptron, and the graph neural network is used to obtain the initial node embedding and And further perform representation learning to update node embedding:

[0018]

[0019] in is the embedding of node i at iteration l, is the edge embedding of the lth iteration according to the edge type, is the set of neighbor nodes of node t, is the weight of the edge; after L layers of iteration, we get and Then obtain the subgraph embedding representation by node aggregation

[0020]

[0021] in yes The node set of , σ is the activation function;

[0022] Preferably, in step S3, a task sampler is constructed to sample the target task and randomly sample auxiliary tasks to reconstruct the subgraph, as follows:

[0023] S3.1: According to step S2.3, we can get the set of task subgraph representations Learn a similarity matrix A pool To model the relationship between meta-tasks in the task pool:

[0024]

[0025] in is the cosine similarity function, and then through message passing we can get:

[0026] Z′ pool =ReLU(A pool Z pool W pool ) (3)

[0027] where Z′ pool is the updated task subgraph representation, W poolis the learnable parameter matrix.

[0028] S3.1: Then calculate the sampling probability The updated representation As input, the sampling probability calculation formula is obtained:

[0029]

[0030] in It is a task subgraph The sampling probability, f meta is a function composed of multi-layer perceptrons; the sampling probability of the target task t is calculated by the following formula:

[0031]

[0032] The target task is selected according to the sampling probability. In addition, the NT-Xent loss is used for contrastive learning between subgraphs to optimize subgraph representation learning. In order to effectively accumulate meta-knowledge, the contrastive loss is used as the optimization reward R in combination with the reinforcement learning idea. meta , optimize the task sampler, the optimization process is as follows:

[0033]

[0034] where φ is the parameter of the meta-training task sampler, lr meta is the learning rate of the selected meta-training task, P meta is the sampling probability of the selected meta-training task, b meta Indicates dynamically calculating the average value of rewards.

[0035] S3.3: Randomly sample auxiliary tasks, combine the target task subgraph with the auxiliary tasks, and reconstruct the subgraph. Specifically: other available properties of the molecule can be regarded as auxiliary properties. By randomly sampling auxiliary tasks and combining them with the target task subgraph, an auxiliary task node set is added to the subgraph. Constitute the final task subgraph; each task subgraph has a target attribute, N a auxiliary attributes, a molecule sampled from the query set, and 2K supporting molecules in the support set;

[0036] Preferably, in step S4, a graph neural network is used to perform representation learning on the subgraph, calculate the loss, and update the model parameters, as follows:

[0037] S4.1: After obtaining the final task subgraph, the subgraph is represented and learned through the graph neural network. The node embedding update formula is as follows:

[0038]

[0039] in is the embedding of node i at iteration l, is the edge embedding of the lth iteration according to the edge type, is the set of neighbor nodes of node i, is the weight of the edge; after L layers of iteration, we get and The predicted value can be obtained:

[0040]

[0041] in is the final predicted value of molecule i for attribute t, f cls is a multilayer perceptron and σ is the activation function.

[0042] S4.2: Calculate the loss and update the model parameters. The loss on the support set can be calculated using the following formula:

[0043]

[0044] Use the loss of the support set to update so that the model can quickly adapt to the target task; use the updated model to predict the query set and calculate the query set loss:

[0045]

[0046] Accumulate the query set losses of different tasks, and update the parameters through the gradient descent strategy to optimize the model.

[0047] Preferably, in step S5, the model trained in step S4 is adapted to the parameters of the target attribute, thereby obtaining a generalized model oriented to the target attribute for molecular attribute prediction, specifically as follows: the model trained in step S4 is fine-tuned using the labeled molecules of the target attribute, so that it can be quickly generalized to the prediction of the target attribute, and then the model can be used to predict the molecules to be tested of the target attribute.

[0048] The second aspect of the present invention relates to a small sample molecular property prediction device based on meta-training task sampling optimization, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the small sample molecular property prediction method based on meta-training task sampling optimization of the present invention.

[0049] The third aspect of the present invention relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the small sample molecular property prediction method based on meta-training task sampling optimization of the present invention.

[0050] The advantages and effects of the present invention compared with the prior art are:

[0051] Although the existing technical solutions all pay attention to the importance of the relationship between molecules and attributes, they fail to realize the importance of attribute-attribute relationships. The present invention complements the guidance of the relationship between properties of the many-to-many relationship between molecules and properties by constructing task similarity relationships, enhances the feature interaction between properties, optimizes the task sampling module, and uses the contrastive learning mechanism to minimize the semantic gap between different views of the same attribute, amplifies the semantics between different target attribute subgraphs, and maintains the particularity of the molecular properties themselves while enhancing the interaction, so that attribute prediction can be performed more effectively. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0053] Figure 1 It is a framework diagram of the present invention;

[0054] Figure 2 It is a task sampling module diagram of the present invention;

[0055] Figure 3 It is a schematic diagram of model optimization and property prediction of the present invention.

[0056] Figure 4 It is a schematic diagram of the device of the present invention. DETAILED DESCRIPTION

[0057] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as limiting the present invention, but should be understood as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0058] It should be understood that the terms described in the present invention are only for describing special embodiments and are not intended to limit the present invention. In addition, for the numerical range in the present invention, it should be understood that each intermediate value between the upper and lower limits of the scope is also specifically disclosed. Each smaller range between the intermediate value in any stated value or stated range and any other stated value or intermediate value in the described range is also included in the present invention. The upper and lower limits of these smaller ranges can be independently included or excluded in the scope.

[0059] Unless otherwise indicated, all technical and scientific terms used herein have the same meanings as those generally understood by those skilled in the art. Although the present invention describes only preferred methods and materials, any methods and materials similar or equivalent to those described herein may also be used in the implementation or testing of the present invention. All documents mentioned in this specification are incorporated by reference to disclose and describe the methods and / or materials associated with the documents. In the event of a conflict with any incorporated document, the content of this specification shall prevail.

[0060] It will be apparent to those skilled in the art that various modifications and variations may be made to the specific embodiments of the present invention description without departing from the scope or spirit of the present invention. Other embodiments derived from the present invention description will be apparent to those skilled in the art. The present application description and examples are exemplary only.

[0061] The words “include,” “including,” “have,” “contain,” etc. used in this document are open-ended terms, meaning including but not limited to.

[0062] Unless otherwise specified, the "parts" described in the present invention are all based on mass parts.

[0063] Example 1

[0064] This embodiment provides a small sample molecular property prediction method based on meta-training task sampling optimization. The present invention uses the ToxCAST dataset, such as Figure 1 As shown, the implementation steps include the following:

[0065] Construct a molecule-property relationship diagram:

[0066] Step S1: Constructing a relationship diagram between molecules and attributes represents a node set and ε represents the edge set; molecules and attributes are regarded as nodes in the graph, and molecule nodes are represented as where x m is a molecular node, and the attribute node set is represented as where p t By connecting all molecule nodes and attribute nodes, a graph is formed; each edge in the graph represents the relationship between a molecule and an attribute, and all nodes and edges together constitute a molecule-attribute relationship graph.

[0067] Build a candidate meta-task pool and a small sample molecular property model:

[0068] S2.1: Reconstruct subgraphs from the molecule-attribute graph and associate them with meta-tasks; each meta-task contains a query set and support set In order to transform each meta-task E t Represented as a task subgraph Will support set and queryset are all represented as a subgraph; the support set It can be expressed as:

[0069]

[0070] in Yes t Adjacent molecular nodes, e i,t is the edge between the molecular node i and the attribute node t, and the query set It can be expressed as:

[0071]

[0072] S2.2: By merging and The meta-task E t Represented as a task subgraph

[0073]

[0074] S2.3: Construct a pool of candidate meta-task subgraphs; according to the contrastive learning setting, randomly sample the pool of candidate meta-task subgraphs to form a pool of candidate meta-tasks in Refers to two different representations of the target task t; each It contains a target task, a query molecule and 2K supporting molecules; each meta-task in the task pool corresponds to each molecular property prediction problem; then a small sample molecular property model is constructed, including graph neural network and multi-layer perceptron; the graph neural network is used to obtain the initial node embedding and

[0075]

[0076] where f mol represents the encoder, Embedding(·) represents the embedding layer, and further representation learning is performed to update the node embedding:

[0077]

[0078] in is the embedding of node i at iteration l, is the edge embedding of the lth iteration according to the edge type, is the set of neighbor nodes of node i, is the weight of the edge; after L layers of iteration, we get and Obtaining subgraph embedding representation by node aggregation

[0079]

[0080] in yes The node set of

[0081] S3. Meta-task sampler construction, sampling auxiliary tasks and subgraph reconstruction, such as Figure 2 :

[0082] S3.1: In view of the remarkable achievements of meta-tasks in the field of meta-learning, this training mechanism has gradually been applied to small sample molecular property prediction tasks. In this invention, a meta-training task sampler based on two aspects is proposed. On the one hand, each subgraph centered on the target attribute has different molecules, which can be regarded as different views of the target attribute, so their semantic differences should be minimized; accordingly, the semantic differences of different target attribute subgraphs should be maximized; on the other hand, by capturing the relationship between tasks, the sampler can be guided to sample meta-training tasks and optimize meta-knowledge accumulation; according to step S2.3, the set of task subgraph representations can be obtained. Learn a similarity matrix A pool To model the relationship between meta-tasks in the task pool:

[0083]

[0084] in is the cosine similarity function, which can be obtained through the message passing layer:

[0085] Z′ pool =ReLU(A pool Z pool W pool ) (3)

[0086] where Z′ pool is the updated task subgraph representation, W pool is the learnable parameter matrix.

[0087] S3.2: The updated representation is then As input, the sampling probability is expressed as:

[0088]

[0089] in It is a task subgraph The sampling probability, f metais a function composed of multi-layer perceptrons; the sampling probability of the target task t is calculated by the following formula:

[0090]

[0091] The target task is selected for meta-training based on the sampling probability; in addition, NT-Xent loss is used for contrastive learning between subgraphs:

[0092]

[0093] Where σ is the temperature parameter, N meta represents the number of meta-training tasks. Next, combined with the idea of ​​reinforcement learning, the contrast loss is used as the optimization reward R meta , optimize the task sampler, the optimization process is as follows:

[0094]

[0095] where φ is the parameter of the meta-training task sampler, lr meta is the learning rate of the selected meta-training task, P meta is the sampling probability of the selected meta-training task, b meta Indicates dynamically calculating the average value of rewards.

[0096] S3.3: Randomly sample auxiliary tasks, combine the target task subgraph with the auxiliary tasks, and reconstruct the subgraph; use other available attributes of the molecule as auxiliary attributes, add some other attribute nodes connected to the molecule node in the subgraph, and construct the auxiliary task subgraph

[0097]

[0098] Where N a is the number of auxiliary attributes, p a is an auxiliary attribute node, e i,a is the edge between the molecular node i and the auxiliary attribute node a, yes The molecular node in the target task subgraph With auxiliary subgraph Combined and reconstructed into the final task subgraph:

[0099]

[0100] Each task subgraph With a target attribute, N a auxiliary attributes, a query molecule, and 2K supporting molecules in the support set;

[0101] S4. Model training and attribute prediction, such as Figure 3 :

[0102] Step S4.1: Use the message passing model to learn the representation of the subgraph, and the node embedding update formula is as follows:

[0103]

[0104] in is the embedding of node i at the lth iteration layer, is the initial edge embedding according to edge type, is the set of neighbor nodes of node i, is the weight of the edge; finally, we can use the classifier f built by the multi-layer perceptron cls The predicted results can be expressed by the following formula:

[0105]

[0106] in is the final predicted value of molecule i for attribute t, is the final embedding of molecule i, is the final embedding of attribute t and σ is the activation function.

[0107] Step S4.2: Calculate the loss and update the model parameters. The loss on the support set can be calculated using the following formula:

[0108]

[0109] Use the loss of the support set to update so that the model can quickly adapt to the target task; use the updated model to predict the query set and calculate the query set loss:

[0110]

[0111] Accumulate the query set losses of different tasks, and update the parameters through the gradient descent strategy to optimize the model.

[0112] S5: Based on the model trained in step S4, fine-tune it to focus on the prediction of target attributes; during the fine-tuning process, the model learns from the support set of the target task and adjusts its parameters. After fine-tuning, the model can adapt to the prediction of molecular attributes of the target properties. For a new input molecule i, the molecule-attribute relationship subgraph can be constructed in combination with the support set and input into the graph neural network for learning to obtain the molecule and target attribute representation z query and z τ , and then use the classifier to predict the prediction results of the input molecule

[0113]

[0114] The present invention can be applied to molecular property prediction tasks in real scenarios and can greatly improve the accuracy of prediction results.

[0115] In summary, the present invention proposes a small sample molecular property prediction method based on meta-training task sampling optimization, which makes up for the defect that the existing methods fail to capture the similarities between molecular properties, and solves the problem of limiting model performance due to failure to fully mine this key information. This method significantly improves the accuracy and efficiency of the molecular property prediction model.

[0116] Example 2

[0117] like Figure 4 This embodiment relates to a small sample molecular property prediction device based on meta-training task sampling optimization, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the small sample molecular property prediction method based on meta-training task sampling optimization of Example 1.

[0118] Example 3

[0119] This embodiment relates to a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the method for predicting molecular properties of small samples based on meta-training task sampling optimization of Embodiment 1 is implemented.

[0120] The embodiments described above are only descriptions of the preferred modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A small sample molecular property prediction method based on meta-training task sampling optimization, characterized in that: The following steps are involved: S1: constructing a molecule-attribute relationship diagram; S2: Based on the relationship graph obtained in step S1, a candidate meta-task pool and a small sample molecular attribute model are constructed, and a subgraph representation is obtained; S3: Build a task sampler to sample the target task and randomly sample auxiliary tasks to reconstruct the subgraph; S4: Use graph neural network to learn the representation of subgraphs, calculate losses, and update model parameters; S5: Adapt the parameters of the model trained in step S4 to the target properties, and then obtain a generalized model oriented to the target properties for molecular property prediction.

2. According to claim 1, the small sample molecular property prediction method based on meta-training task sampling optimization is characterized by: In the step S1, a molecule-attribute relationship graph is constructed, specifically as follows: molecules and attributes are both taken as nodes, and molecules and their corresponding known attributes are connected by edges to form a molecule-attribute relationship graph.

3. The small sample molecular property prediction method based on meta-training task sampling optimization according to claim 1 is characterized by: In step S2, based on the relationship graph obtained in step S1, a candidate meta-task pool and a small sample molecular attribute model are constructed, and a subgraph representation is obtained, specifically as follows: the sampled subgraph in the molecule-attribute relationship graph is linked to the meta-task; each meta-task is expressed as a task subgraph; a candidate meta-task subgraph pool is constructed, and according to the setting of contrastive learning, the candidate meta-task subgraph pool is randomly sampled to form a candidate meta-task pool; then a small sample molecular attribute model is constructed, including a graph neural network and a multi-layer perceptron, and the graph neural network is used to obtain the initialization node embedding and And further perform representation learning to update node embedding to obtain and Then obtain the subgraph embedding representation by node aggregation in yes The node set of .

4. The small sample molecular property prediction method based on meta-training task sampling optimization according to claim 1 is characterized by: In step S3, a task sampler is constructed to sample the target task and randomly sample auxiliary tasks to reconstruct the subgraph. Specifically, a similarity matrix A is learned. pool Model the relationship between meta-tasks in the task pool: in, is the cosine similarity function; then the task subgraph representation is updated through message passing: Among them, Z pool is the set of task subgraphs, Z′ pool is the updated task subgraph representation, W pool is a learnable parameter matrix; then represented by the updated subgraph As input, the sampling probability of the task subgraph is calculated: where f meta It is a function composed of multi-layer perceptrons; then the sampling probability of the target task t is calculated: The target task is selected for meta-training based on the sampling probability. The NT-Xent loss is introduced in the task sampling module for contrastive learning between subgraphs and optimization of subgraph representation learning: Among them, σ is the temperature parameter, N meta is the number of meta-training tasks, is the cosine similarity function; in addition, based on the idea of ​​reinforcement learning, the contrast loss is used as the optimization reward R meta , optimize the task sampler, the optimization process is as follows: where φ is the parameter of the meta-training task sampler, lr meta is the learning rate of the selected meta-training task, P meta is the sampling probability of the selected meta-training task, b meta Indicates the average value of the dynamically calculated reward. Next, randomly sample auxiliary tasks, combine the target task subgraph with the auxiliary tasks, and reconstruct the subgraph into the final task subgraph: Each task subgraph With a target attribute, N a auxiliary attributes, a query molecule, and 2K supporting molecules in the support set.

5. The small sample molecular property prediction method based on meta-training task sampling optimization according to claim 1 is characterized by: In step S4, a graph neural network is used to learn the representation of the subgraph, calculate the loss, and update the model parameters, as follows: a message passing model is used to learn the representation of the subgraph. After the node embedding is updated, a classifier f constructed by a multi-layer perceptron is used. cls Make predictions about attributes: in is the final predicted value of molecule i for attribute t, is the final embedding of molecule i, is the final embedding of attribute t, σ is the activation function; then calculate the loss of the support set: Use the loss of the support set to update so that the model can quickly adapt to the target task; use the updated model to predict the query set and calculate the query set loss: Accumulate the query set losses of different tasks, and update the parameters through the gradient descent strategy to optimize the model.

6. The small sample molecular property prediction method based on meta-training task sampling optimization according to claim 1 is characterized by: In step S5, the model trained in step S4 is adapted to the parameters of the target property, thereby obtaining a generalized model oriented to the target property for molecular property prediction, specifically as follows: the model trained in step S4 is fine-tuned using the labeled molecules of the target property, so that it can be quickly generalized to the prediction of the target property, and then the model is used to predict the molecules to be tested of the target property.

7. A small sample molecular property prediction device based on meta-training task sampling optimization, characterized in that: It comprises a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the small sample molecular property prediction method based on meta-training task sampling optimization as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the small sample molecular property prediction method based on meta-training task sampling optimization described in any one of claims 1-6 is implemented.

Images