CN118536542A - Method and system for constructing optimal architecture of graph neural network - Google Patents

Abstract

The invention discloses a method and a system for constructing an optimal structure of a graph neural network, and relates to the technical field of construction of graph neural network structures. The technical key points of the invention include: predesigned search space containing multiple parameters of graphic neural network, and prompt template for guiding reply of large language model; acquiring an initially triggered prompt based on a prompt template; inputting the initially triggered prompt into a pre-trained large language model for sampling to obtain initial multi-group graph neural network configuration parameters; constructing a plurality of corresponding graph neural networks based on the initial plurality of sets of graph neural network configuration parameters; training the multiple graph neural networks respectively by using a training set, and screening by using a testing set; iteratively repeating the above processes to obtain an optimal graph neural network; and applying the constructed optimal graph neural network to a drug property prediction task. The invention can obtain the graph neural network architecture with better performance on the premise of saving training cost and resources.

Description

A method and system for constructing an optimal architecture of a graph neural network

Technical Field

The present invention relates to the technical field of graph neural network architecture construction, and in particular to a method and system for constructing an optimal graph neural network architecture and a method and system for predicting drug molecular properties.

Background Art

When processing graph data with rich relationships, graph neural networks (GNNs) can better capture the non-Euclidean properties of graph data compared to traditional neural networks such as recurrent neural networks. Therefore, there are now many methods that use GNNs to predict molecular properties based on the information of molecular graphs.

At the same time, in order to obtain the optimal GNN architecture for the corresponding task and reduce the parameter adjustment time of algorithm designers, graph neural network architecture search has been derived. This type of method is used for a given search space. and dataset , using a specific search strategy to select the best architecture , Can be maximized The accuracy of the network The main search strategies currently include algorithms based on reinforcement learning, evolutionary learning, and differentiable methods.

Although various graph architecture search algorithms have achieved good results in different domain tasks, designing architecture search algorithms also requires a lot of domain knowledge. Search methods such as random sampling can also achieve good performance results, but they require training and verification of a large number of architectures, which has certain requirements for resources and time. In addition, existing methods require modifications in the code to change different search spaces, reference indicators, etc. This is actually very inconvenient for users who are not in the computer field.

Summary of the invention

In order to solve the above technical problems, the present invention proposes a method and system for constructing an optimal graph neural network architecture and a method and system for predicting drug molecular properties, so as to automatically obtain the optimal graph neural network architecture and save time and human resource costs.

According to one aspect of the present invention, a method for constructing an optimal architecture of a graph neural network is proposed, the method comprising:

Pre-design a search space containing multiple graph neural network parameters;

A prompt template is pre-designed for guiding the large language model to respond; the prompt template includes a description of the pre-designed search space;

Acquire the prompt word of the initial trigger based on the prompt word template;

Input the initial triggering prompt into the pre-trained large language model for sampling to obtain the initial multiple sets of graph neural network configuration parameters;

Construct corresponding multiple graph neural networks based on the initial multiple groups of graph neural network configuration parameters;

Use the training set to train multiple graph neural networks separately, and use the test set to screen them;

The above sampling, construction, training, and screening processes are repeated iteratively until the maximum number of iterations is reached to obtain the final optimal graph neural network.

Furthermore, the multiple graph neural network parameters in the search space include: normalization operation, activation function, convolution layer operation, readout layer operation, number of loops, learning rate, hidden layer dimension, batch number, and ratio; each parameter contains multiple options.

Furthermore, the prompt template also includes task description, response paradigm regulations, and search strategy; wherein the task description is used to explain that this task is to build a graph neural network with an optimal architecture based on the graph neural network parameter training in the search space; the response paradigm regulations are used to specify the parameter form of the large language model response output; the search strategy is used to provide a method for obtaining the optimal graph neural network configuration by reasoning with a large language model.

Furthermore, the search strategy includes: in the exploration stage, i.e., the stage where experimental data is lacking, focusing on graph neural network configurations that have not appeared in the search space during the search; in the development stage, i.e., the stage where experimental data is available, providing common configurations of high-performance graph neural network architectures, corresponding performance indicators, performance indicator means, and deviations from high-performance graph neural network architecture performance indicators and means, retaining common points in high-performance graph neural network configurations during the search, and modifying other parts of the configuration to generate a new graph neural network architecture.

Furthermore, the iterative repetition of the above-mentioned sampling, construction, training, and screening processes to obtain the final optimal graph neural network includes: adopting a training method that combines low-precision and high-precision, and dividing the training rounds into low-precision, high-precision, and training until convergence; in the training stage using the training set, first obtain a low-precision graph neural network model through a few rounds of training, obtain the performance of the low-precision graph neural network model on the test set, screen out half of the models with the best performance and then conduct a majority of rounds of training, and then feed back the training results to the pre-trained large language model; the large language model samples and outputs the better graph neural network configuration parameters obtained by its reasoning; after multiple iterations, select the model with the best performance from all the constructed graph neural networks, train until convergence, and obtain the optimal graph neural network model based on the final performance screening.

According to another aspect of the present invention, a system for constructing an optimal architecture of a graph neural network is proposed, the system comprising:

A prompt template design module, configured to pre-design a search space including a plurality of graph neural network parameters; pre-design a prompt template for guiding a large language model to respond; the prompt template includes a description of the pre-designed search space;

An initial prompt acquisition module, configured to acquire an initial triggered prompt based on a prompt template;

The optimal model construction module is configured to input the initial triggered prompt into the pre-trained large language model for sampling to obtain the initial multiple sets of graph neural network configuration parameters; construct the corresponding multiple graph neural networks based on the initial multiple sets of graph neural network configuration parameters; use the training set to train the multiple graph neural networks separately, and use the test set to screen them; iteratively repeat the above sampling, construction, training, and screening processes until the maximum number of iterations is reached to obtain the final optimal graph neural network.

Furthermore, the multiple graph neural network parameters in the prompt template design module include: normalization operation, activation function, convolution layer operation, readout layer operation, number of loops, learning rate, hidden layer dimension, batch number, ratio; each parameter includes multiple options;

The prompt template also includes task description, response paradigm regulations, and search strategy; wherein the task description is used to explain that this task is to build a graph neural network with an optimal architecture based on the graph neural network parameter training in the search space; the response paradigm regulations are used to specify the parameter form of the large language model response output; the search strategy is used to provide a method for obtaining the optimal graph neural network configuration by inferring the large language model.

Furthermore, the search strategy includes: in the exploration stage, i.e., the stage where experimental data is lacking, focusing on graph neural network configurations that have not appeared in the search space during the search; in the development stage, i.e., the stage where experimental data is available, providing common configurations of high-performance graph neural network architectures, corresponding performance indicators, performance indicator means, and deviations from high-performance graph neural network architecture performance indicators and means, retaining common points in high-performance graph neural network configurations during the search, and modifying other parts of the configuration to generate a new graph neural network architecture.

According to another aspect of the present invention, a method for predicting drug molecular properties based on graph neural network is proposed, the method comprising:

Get the drug molecule graph dataset;

The drug molecule graph dataset is used as a training set and a test set, and the above-mentioned graph neural network optimal architecture construction method is used to obtain a molecular property prediction model based on the optimal graph neural network;

The molecular graph to be predicted is input into the molecular property prediction model based on the optimal graph neural network to obtain the predicted property results.

According to another aspect of the present invention, a drug molecular property prediction system based on graph neural network is also proposed. The system has a program module corresponding to the steps described in the above-mentioned drug molecular property prediction method based on graph neural network, and executes the steps in the above-mentioned drug molecular property prediction method during runtime.

The present invention has the following technical effects:

The present invention proposes a method and system for constructing an optimal architecture of a graph neural network and a method and system for predicting the properties of drug molecules, including: pre-designing a search space containing multiple graph neural network parameters;

Pre-designing a prompt template for guiding the large language model to reply; the prompt template includes a description of the pre-designed search space; obtaining an initial triggering prompt based on the prompt template;

Input the initial triggered prompt into the pre-trained large language model for sampling, and obtain the initial multiple groups of graph neural network configuration parameters; construct the corresponding multiple graph neural networks based on the initial multiple groups of graph neural network configuration parameters; use the training set to train the multiple graph neural networks separately, and use the test set to screen; iteratively repeat the above sampling, construction, training, and screening processes until the maximum number of iterations is reached to obtain the final optimal graph neural network; and apply the constructed optimal graph neural network to the drug property prediction task. The present invention can automatically obtain a graph neural network architecture with better performance based on different data sets of drug property prediction tasks under the premise of saving training costs and resources. The present invention is flexible in operation, and users can adjust data sets, reference indicators, search spaces, etc. as needed.

BRIEF DESCRIPTION OF THE DRAWINGS

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

FIG1 is a flow chart of a method for constructing an optimal architecture of a graph neural network according to an embodiment of the present invention.

FIG. 2 is a schematic diagram of the structure of a message passing neural network in an embodiment of the present invention.

FIG3 is a schematic diagram of the structure of a graph neural network obtained by sampling in an embodiment of the present invention.

Figure 4 is a structural diagram of a graph neural network optimal architecture construction system described in an embodiment of the present invention.

Figure 5 is a schematic diagram of the process of obtaining a molecular property prediction model based on an optimal graph neural network using the graph neural network optimal architecture construction method in an embodiment of the present invention.

DETAILED DESCRIPTION

In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the scope of protection of the present invention.

The embodiment of the present invention proposes a method for constructing an optimal architecture of a graph neural network, as shown in FIG1 , the method comprising:

S110, pre-designing a search space containing multiple graph neural network parameters;

S120, pre-designing a prompt template for guiding the large language model to reply; the prompt template includes a description of the pre-designed search space;

S130, obtaining an initial triggering prompt based on a prompt template;

S140, inputting the initial triggered prompt into the pre-trained large language model for sampling, and obtaining initial multiple sets of graph neural network configuration parameters; constructing corresponding multiple graph neural networks based on the initial multiple sets of graph neural network configuration parameters; using the training set to train the multiple graph neural networks respectively, and using the test set to screen them;

S150, iteratively repeat the sampling, construction, training, and screening processes in S140 until the maximum number of iterations is reached to obtain the final optimal graph neural network.

The method starts from S110. In S110, a search space containing multiple graph neural network parameters is pre-designed; the multiple graph neural network parameters include: normalization operation, activation function, convolution layer operation, readout layer operation, number of cycles, learning rate, hidden layer dimension, batch number, ratio; each parameter includes multiple options.

According to an embodiment of the present invention, in the designed default search space, a message passing neural network (MPNN) is used as the core backbone, as shown in Figure 2. In each layer, different model configuration options such as normalization, activation function, etc. are provided, and hyperparameters such as learning rate and hidden layer dimension are added to form a search space, as shown in Table 1.

Table 1

Then execute S120, in which a prompt template is pre-designed to guide the large language model to reply; the prompt template includes a description of the pre-designed search space, and also includes a task description, a reply paradigm specification, and a search strategy; wherein the task description is used to explain that this task is to build a graph neural network with an optimal architecture based on the graph neural network parameter training in the search space; the reply paradigm specification is used to specify the parameter form of the large language model reply output; and the search strategy is used to provide a method for the large language model to infer the optimal graph neural network configuration.

According to an embodiment of the present invention, the purpose of the prompt is mainly to make the large language model understand the predetermined search space and guide its reasoning to obtain a better model configuration. The prompt word template includes six parts: system settings, task instructions, search space description, response paradigm regulations, search strategy, and re-emphasis. An example of a prompt word template is shown below, where the content in [ ] is constructed by different variables.

In the system settings and task descriptions, explain to the large language model the role it plays in the conversation and the tasks it needs to perform. The tasks to be performed are to build a graph neural network with the optimal architecture based on the graph neural network parameter training in the search space. In the search space description, use clear text to help the large language model better understand the designed search space. In the reply paradigm specification, by giving examples, let the model reply in the format, that is, the parameter form requirements of the reply output, so that it is convenient to automatically process the output data and extract the configuration to build the model. In the search strategy, guide the large language model to reason and obtain the optimal graph neural network configuration by disassembling the reasoning method, that is, give the method for the large language model to reason and obtain the optimal graph neural network configuration. In the re-emphasis part, repeat the important information such as the paradigm that must be followed to reply and the number of configurations that need to be replied to ensure that the large language model replies as required.

Among them, some tips are added to the search strategy, including: in the exploration stage, that is, the stage where experimental data is lacking, focus on graph neural network configurations that have not appeared in the search space during the search; in the development stage, that is, the stage where experimental data is available, give the common configurations of high-performance graph neural network architectures, the corresponding performance indicators, the mean of performance indicators, and the deviation of the performance indicators and mean of high-performance graph neural network architectures. When searching, retain the common points in the high-performance graph neural network configurations, and modify other parts of the configuration to generate a new graph neural network architecture.

Specifically, the sampling process is guided in a step-by-step manner, and the iteration is divided into exploration and development stages. In the previous exploration stage, due to the lack of experimental data, the large language model is guided to focus on model configurations that have not appeared in the search space. In the development stage after accumulating a certain amount of experimental results, the large model is guided to modify some configurations based on the feedback experimental results in the prompt words to explore whether the model performance is optimized. In addition, when feeding back the verification results to the model, the previous performance is averaged according to the following formula to obtain , calculate the performance deviation value of each model , which helps to better utilize the performance of the previous model.

Where m represents the number of iterations; n represents the number of graph neural network architectures returned in one iteration; Represents the j-th graph neural network returned in the i-th iteration; Represents the average of previous performance.

Then, S130 is executed, in which the prompt word for initial triggering is obtained based on the prompt word template.

According to an embodiment of the present invention, the selected data set, the default search space, the reference index, the performance of the trained graph neural network, etc. are taken as variables, and the initial trigger prompt is automatically constructed according to the pre-designed text prompt template. The prompt can be simply understood as the text that guides the large language model to reply.

Then execute S140, in which the initial triggered prompt is input into the pre-trained large language model for sampling to obtain initial multiple sets of graph neural network configuration parameters; construct corresponding multiple graph neural networks based on the initial multiple sets of graph neural network configuration parameters; use the training set to train the multiple graph neural networks separately, and use the test set to screen them.

Then execute S150, in which the sampling, construction, training, and screening processes in S140 are iteratively repeated until the maximum number of iterations is reached to obtain the final optimal graph neural network.

According to an embodiment of the present invention, the prompt is input into a trained large language model. A large language model refers to a model with large-scale parameters for processing natural language, which often has powerful reasoning capabilities and rich built-in knowledge. Specifically, the existing trained large language model can be used in the form of a calling interface.

During the sampling process, the large language model will return the model configuration that it infers to be optimal in the form of text. All configurations are sampled from the search space. The network construction module parses these configurations to form the corresponding graph neural network model, as shown in Figure 3.

These graph neural network models are then trained and verified on the dataset using a combination of low-precision and high-precision methods. The model with the most accurate prediction results is selected from the verification results of each round and returned to the large language model to guide it to derive the next set of model configurations with more accurate prediction results. Specifically, a training method that combines low-precision and high-precision methods is adopted, and the training rounds are divided into low-precision, high-precision, and training until convergence; in the training stage, a low-precision graph neural network model is first obtained through a few rounds of training, and the performance of the low-precision graph neural network model is obtained on the test set. Half of the models with the best performance are selected and then trained for more rounds. The training results are then fed back to the large language model, and the large language model again outputs the better graph neural network configuration parameters obtained by its reasoning; after multiple iterations, the maximum number of iterations is reached, and the graph neural network model is selected from all graph neural network models. The model with the best performance is selected and trained until convergence, and the optimal graph neural network model is obtained based on the final performance.

Compared with the traditional method of training all models until convergence and then screening them, this training method can concentrate resources from models with poor performance to models with better performance, which can greatly save training resources and costs.

The present invention can automatically obtain a graph neural network architecture with better performance based on different data sets while saving training costs and resources as much as possible. In addition, the present invention is more flexible in operation, convenient for non-computer professional users to use, and can adjust data sets, reference indicators, search spaces, etc.

Another embodiment of the present invention provides a system for constructing an optimal architecture of a graph neural network, as shown in FIG4 , the system comprising:

A prompt template design module 410 is configured to pre-design a search space including multiple graph neural network parameters; pre-design a prompt template for guiding the large language model to reply; the prompt template includes a description of the pre-designed search space;

An initial prompt acquisition module 420, configured to acquire an initial triggered prompt based on a prompt template;

The optimal model construction module 430 is configured to input the initial triggered prompt into the pre-trained large language model for sampling, and obtain the initial multiple sets of graph neural network configuration parameters; construct the corresponding multiple graph neural networks based on the initial multiple sets of graph neural network configuration parameters; use the training set to train the multiple graph neural networks separately, and use the test set to screen them; iteratively repeat the above sampling, construction, training, and screening processes until the maximum number of iterations is reached to obtain the final optimal graph neural network.

In this embodiment, preferably, the multiple graph neural network parameters in the prompt template design module 410 include: normalization operation, activation function, convolution layer operation, readout layer operation, number of loops, learning rate, hidden layer dimension, batch number, ratio; each parameter contains multiple options.

In this embodiment, preferably, the prompt template in the prompt template design module 410 also includes a task description, a response paradigm specification, and a search strategy; wherein the task description is used to explain that this task is to build a graph neural network with an optimal architecture based on the graph neural network parameter training in the search space; the response paradigm specification is used to specify the parameter form of the large language model response output; and the search strategy is used to provide a method for obtaining the optimal graph neural network configuration by inferring the large language model.

In this embodiment, preferably, the search strategy described in the prompt template design module 410 includes: in the exploration stage, i.e., the stage where experimental data is lacking, focusing on graph neural network configurations that have not appeared in the search space during the search; in the development stage, i.e., the stage where experimental data is available, providing common configurations of high-performance graph neural network architectures, corresponding performance indicators, performance indicator means, and deviations from high-performance graph neural network architecture performance indicators and means, retaining common points in high-performance graph neural network configurations during the search, and modifying other parts of the configuration to generate a new graph neural network architecture.

In this embodiment, preferably, the iterative repetition of the above-mentioned sampling, construction, training, and screening processes in the optimal model construction module 430 to obtain the final optimal graph neural network includes: adopting a training method that combines low-precision and high-precision, and dividing the training rounds into low-precision, high-precision, and training until convergence; in the training stage using the training set, first obtain a low-precision graph neural network model through a few rounds of training, obtain the performance of the low-precision graph neural network model on the test set, screen out half of the models with the best performance and then conduct most rounds of training, and then feed back the training results to the pre-trained large language model; the large language model samples and outputs the better graph neural network configuration parameters obtained by its reasoning; after multiple iterations, select the model with the best performance from all constructed graph neural networks, train until convergence, and obtain the optimal graph neural network model based on the final performance screening.

For the undescribed parts of a system for building an optimal architecture of a graph neural network according to an embodiment of the present invention, please refer to the above detailed description of the method embodiment.

Another embodiment of the present invention further provides a method for predicting drug molecular properties based on graph neural network, the method comprising:

Get the drug molecule graph dataset;

The drug molecule graph dataset is used as a training set and a test set, and the optimal graph neural network architecture construction method described in the above embodiment is used to obtain a molecular property prediction model based on the optimal graph neural network;

The molecular graph to be predicted is input into the molecular property prediction model based on the optimal graph neural network to obtain the predicted property results.

According to an embodiment of the present invention, a specific drug property prediction task and a corresponding data set for the execution link are selected. In this embodiment, the constructed optimal graph neural network supports all data forms containing two core fields: the drug molecule smiles string and the target label, and 6 sub-datasets in the MoleculeNet dataset have been provided in advance as examples.

As shown in Figure 5, the constructed optimal graph neural network is applied to the drug property prediction task, and various properties of the drug are predicted based on the smiles string structure of the drug. The self-designed prompt word template guides the large language model to infer the search space and previous experimental results to obtain a GNN with better performance.

On the six sub-datasets used, the present invention is compared with five classic GNN models: GCN, GAT, GAT-V2, GraphSAGE, and GIN, and the results are compared using random sampling in the same search space. The comparison results are shown in Table 2.

Table 2

Using the same partitioning on the same dataset, we can see that the performance of the present invention has obvious advantages on most datasets. It should be noted that the ESOL, Lipophilicity, and FreeSolv datasets are regression tasks, so the RMSE (root mean square error) indicator is used to measure the performance. The smaller the indicator, the better the performance. The Tox21, ToxCast, and BACE datasets are classification tasks, so the AUC (area under the ROC curve) indicator is used to measure the performance. The closer the indicator is to 1, the better the performance.

Another embodiment of the present invention proposes a drug molecular property prediction system based on graph neural network, which has a program module corresponding to the steps of the drug molecular property prediction method based on graph neural network described in the above embodiment, and executes the steps in the above drug molecular property prediction method during runtime.

The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some or all of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims (10)

1. A method for constructing an optimal architecture of a graph neural network, comprising: Pre-design a search space containing multiple graph neural network parameters; A prompt template is pre-designed for guiding the large language model to respond; the prompt template includes a description of the pre-designed search space; Obtaining the initial triggering prompt based on the prompt template; Input the initial triggering prompt into the pre-trained large language model for sampling to obtain the initial multiple sets of graph neural network configuration parameters; Construct corresponding multiple graph neural networks based on the initial multiple groups of graph neural network configuration parameters; Use the training set to train multiple graph neural networks separately, and use the test set to screen them; The above sampling, construction, training, and screening processes are repeated iteratively until the maximum number of iterations is reached to obtain the final optimal graph neural network. 2. According to claim 1, a method for constructing an optimal graph neural network architecture is characterized in that the multiple graph neural network parameters in the search space include: normalization operation, activation function, convolutional layer operation, readout layer operation, number of loops, learning rate, hidden layer dimension, batch number, ratio; each parameter contains multiple options. 3. According to claim 1, a method for constructing an optimal graph neural network architecture is characterized in that the prompt template also includes a task description, a response paradigm specification, and a search strategy; wherein the task description is used to explain that this task is to construct a graph neural network with an optimal architecture based on the graph neural network parameter training in the search space; the response paradigm specification is used to specify the parameter form of the large language model response output; and the search strategy is used to provide a method for obtaining the optimal graph neural network configuration by reasoning with a large language model. 4. According to claim 3, a method for constructing an optimal graph neural network architecture is characterized in that the search strategy includes: in the exploration stage, i.e., the stage where experimental data is lacking, focusing on graph neural network configurations that have not appeared in the search space during the search; in the development stage, i.e., the stage where experimental data is available, giving common configurations of high-performance graph neural network architectures, corresponding performance indicators, performance indicator means, and deviations from high-performance graph neural network architecture performance indicators and means, retaining common points in high-performance graph neural network configurations during the search, and modifying other parts of the configuration to generate a new graph neural network architecture. 5. According to claim 1, a method for constructing an optimal graph neural network architecture is characterized in that the iterative repetition of the above-mentioned sampling, construction, training, and screening processes to obtain the final optimal graph neural network includes: adopting a training method that combines low precision and high precision, and dividing the training rounds into low precision, high precision, and training until convergence; in the training stage using the training set, first obtain a low-precision graph neural network model through a few rounds of training, obtain the performance of the low-precision graph neural network model on the test set, screen out half of the models with the best performance and then conduct a majority of rounds of training, and then feed back the training results to the pre-trained large language model; the large language model samples and outputs the better graph neural network configuration parameters obtained by its reasoning; after multiple iterations, select the model with the best performance from all constructed graph neural networks, train until convergence, and obtain the optimal graph neural network model based on the final performance screening. 6. A graph neural network optimal architecture construction system, characterized by comprising: A prompt template design module, configured to pre-design a search space including a plurality of graph neural network parameters; pre-design a prompt template for guiding a large language model to respond; the prompt template including a description of the pre-designed search space; An initial prompt acquisition module, configured to acquire an initial triggered prompt based on a prompt template; The optimal model construction module is configured to input the initial triggered prompt into the pre-trained large language model for sampling to obtain the initial multiple sets of graph neural network configuration parameters; construct the corresponding multiple graph neural networks based on the initial multiple sets of graph neural network configuration parameters; use the training set to train the multiple graph neural networks separately, and use the test set to screen them; iteratively repeat the above sampling, construction, training, and screening processes until the maximum number of iterations is reached to obtain the final optimal graph neural network. 7. A graph neural network optimal architecture construction system according to claim 6, characterized in that the multiple graph neural network parameters in the prompt template design module include: normalization operation, activation function, convolution layer operation, readout layer operation, number of loops, learning rate, hidden layer dimension, batch number, ratio; each parameter contains multiple options; The prompt template also includes task description, response paradigm regulations, and search strategy; wherein the task description is used to explain that this task is to build a graph neural network with an optimal architecture based on the graph neural network parameter training in the search space; the response paradigm regulations are used to specify the parameter form of the large language model response output; the search strategy is used to provide a method for obtaining the optimal graph neural network configuration by inferring the large language model. 8. According to claim 7, a graph neural network optimal architecture construction system is characterized in that the search strategy includes: in the exploration stage, i.e., the stage where experimental data is lacking, focusing on graph neural network configurations that have not appeared in the search space during the search; in the development stage, i.e., the stage where experimental data is available, giving common configurations of high-performance graph neural network architectures, corresponding performance indicators, performance indicator means, and deviations from high-performance graph neural network architecture performance indicators and means, retaining common points in high-performance graph neural network configurations during the search, and modifying other parts of the configuration to generate a new graph neural network architecture. 9. A method for predicting drug molecular properties based on graph neural network, comprising: Get the drug molecule graph dataset; Using the drug molecule graph dataset as a training set and a test set, and using a graph neural network optimal architecture construction method according to any one of claims 1 to 5 to obtain a molecular property prediction model based on an optimal graph neural network; The molecular graph to be predicted is input into the molecular property prediction model based on the optimal graph neural network to obtain the predicted property results. 10. A drug molecular property prediction system based on graph neural network, characterized in that the system has a program module corresponding to the steps described in claim 9 above, and executes the steps in the above drug molecular property prediction method when running.