Integrated regression model and generation method and prediction method for molecular virtual screening

Abstract

The present invention relates to the technical field of computer-aided drug design, and specifically discloses an integrated regression model and a generation method and a prediction method for molecular virtual screening. The generation method comprises: randomly selecting K molecules from a super-large-scale compound library as a training set; training a regression model; using the trained regression model to predict the molecules not selected in the super-large-scale compound library to obtain a comprehensive prediction result; selecting K molecules with the highest prediction score and the largest standard deviation from the positive prediction pool of the comprehensive prediction result, and adding them to the training set; training the regression model again; when the number of training times reaches the number of iterations, selecting multiple regression models with the best prediction performance to obtain an integrated regression model. The present invention selects the compound with the largest standard deviation of the prediction value from the compounds with the highest score for annotation, so as to focus on the most challenging or uncertain samples, thereby improving the annotation efficiency and model performance.

Description

Technical Field

[0001] The present invention relates to the technical field of computer-aided drug design, and in particular to an integrated regression model for molecular virtual screening, a generation method and a prediction method. Background Art

[0002] Although virtual screening technology can cover millions of compounds, compared with the total number of known drug-like molecular structures (about 10 63 ), its coverage is still relatively narrow. Therefore, how to significantly expand the number of molecules that can be explored in the early stages of drug discovery has become an urgent problem to be solved in the current drug research and development field.

[0003] With the advent of on-demand compound libraries, the number of purchasable compounds for virtual screening has exploded in recent years, exceeding 1 billion compounds. Although many studies have successfully applied ultra-large-scale molecular libraries of more than 1 billion compounds for virtual screening, there are indeed some limitations. First, screening large-scale databases will inevitably require a lot of computing resources. For example, ligand-based virtual screening of a library containing 10 10 The cost of generating a conformational set for a database of molecules is about $20,000. A structure-based virtual screen for heat shock protein 90 (HSP90) took a total of 18 hours using 45,000 CPUs, or about 55 years on a single CPU. This cost was $15,000, and that fee was only for searching the database, not for its generation or storage.

[0004] On the other hand, for targets without a sufficient number of known active ligands, ligand-based virtual screening methods cannot be used to quickly filter large virtual libraries, and structure-based screening methods can only be used. Current studies have used structure-based drug design workflows to screen very large databases. These studies require a lot of computing resources, which are not available to many academic researchers. In addition, this high computational cost makes it impractical to apply this strategy to many different protein targets. As virtual libraries become larger and larger, new strategies must be developed to alleviate the computational cost of these exhaustive screening activities. In order to find the best scoring ligands in billions of molecular libraries with minimal loss of accuracy.

[0005] To address this problem, the active learning (AL) method was introduced. The active learning method not only broadens the scope of application of ultra-large-scale virtual screening, but also provides a feasible solution for the rapid implementation of ultra-large-scale virtual screening under laboratory conditions. As a method that can increase the screening speed without additional resources, active learning is centered on using statistical models to replace docking algorithms to avoid time-consuming brute-force searches. The application of AL algorithms in the field of drug discovery can be traced back to the pioneering work of Warmuth et al., whose research highlighted the potential of AL in handling labeling tasks in drug development. In the past two decades, AL has shown broad application prospects in multiple drug development projects, including focused library design, rational de novo design, drug combinations, tracking cells, molecular property prediction, and the search for active compounds.

[0006] Previous studies have focused on using AL algorithms to identify ideal samples and improve the capabilities of machine learning (ML) models through additional training data. According to different application requirements, the preferences of AL can be changed by flexibly adjusting the sampling method. For example, exploration-oriented sampling methods tend to select samples with novel structures to broaden the applicability of the model. Development-oriented sampling methods, on the other hand, focus more on selecting samples with high performance, such as compound binding affinity for specific targets, which can usually significantly improve the hit rate. The sampling method, that is, the query strategy, is the key to the success of active learning methods. Inappropriate sampling methods may lead to poor model prediction performance, which in turn requires more active learning iterations. Summary of the invention

[0007] The present invention aims to solve the problem of poor model prediction performance caused by inappropriate sampling methods. To this end, the present invention provides an integrated regression model, generation method, and prediction method for molecular virtual screening, the core technical principle of which is to improve the efficiency and accuracy of virtual screening by using active learning strategies and the integration of multiple regression models. The present invention selects compounds with the largest standard deviation of predicted values ​​from the highest-scoring compounds for annotation, so as to focus on the most challenging or most uncertain samples, thereby improving annotation efficiency and model performance. At the same time, the differences and diversity between different models are fully utilized, and the predicted values ​​between the models are combined as the final prediction results by weighted averaging to reduce the deviation of a single model and improve the performance and robustness of the overall model. The present invention is particularly suitable for ultra-large-scale compound libraries with more than 1 billion compounds.

[0008] The present invention provides an integrated regression model generation method for molecular virtual screening, and the technical scheme adopted is as follows: comprising the following steps:

[0009] Step 1: Randomly select K molecules from the ultra-large-scale compound library as the training set;

[0010] Step 2: Perform docking simulation on the molecules in the training set to obtain docking scores;

[0011] Step 3: using the docking score and the training set to train n regression models;

[0012] Step 4: Determine whether the number of training times reaches the number of iterations. If not, proceed to step 5. If it reaches the number of iterations, proceed to step 7.

[0013] Step 5: Use the trained n regression models to predict the molecules that have not been selected in the ultra-large-scale compound library to obtain a comprehensive prediction result;

[0014] Step 6: Select K molecules with the highest prediction scores and the largest standard deviations from the positive prediction pool of the comprehensive prediction results, perform docking simulation, obtain docking scores, and add them to the training set; proceed to step 3;

[0015] Step 7: Evaluate the prediction performance of the n regression models after training, select the p regression models with the best prediction performance, and obtain the integrated regression model, n>p.

[0016] Furthermore, the molecules selected from the ultra-large-scale compound library are processed, which includes determining the stereoisomer configuration and the correct ionization state of the molecules, and converting the molecules in SMILES format into ECFP4 fingerprints using RDKit.

[0017] Furthermore, in step 5, the weighted average of the prediction results of the n regression models is taken as the comprehensive prediction result.

[0018] Furthermore, K is 1% of the total number of molecules in the ultra-large-scale compound library.

[0019] Furthermore, the number of iterations is 8 times.

[0020] Furthermore, in step 6, the process of selecting K molecules with the highest prediction scores and the largest standard deviations is:

[0021] Sort all molecules according to their prediction scores and select the N molecules with the highest scores, N>K;

[0022] Calculate the standard deviation of the predicted value for each of the N molecules and select the K molecules with the largest standard deviations.

[0023] Furthermore, the regression model includes multiple ones of RF (random forest), SVM (support vector machine), Ridge (ridge regression), XGBoost (extreme gradient boosting), LGBM (lightweight gradient boosting), DNN (deep neural network), GCN (graph convolutional network) and GAT (graph attention network).

[0024] Furthermore, in step 7, RMSE is used as an evaluation index of the prediction performance of the regression model.

[0025] The present invention also provides an integrated regression model for molecular virtual screening, and the technical scheme adopted is as follows: four regression models are selected from eight regression models as the integrated regression model by using the above-mentioned integrated regression model generation method for molecular virtual screening;

[0026] Regression models include: RF, SVM, Ridge, XGBoost, LGBM, DNN, GCN, and GAT;

[0027] The integrated regression model consists of SVM, XGBoost, GCN and GAT.

[0028] The present invention also provides a prediction method for molecular virtual screening, and the technical solution adopted is as follows: using the above-mentioned integrated regression model for molecular virtual screening, comprising the following steps:

[0029] S1: Obtain the molecule to be docked;

[0030] S2: inputting the ECFP4 fingerprint of the molecule to be docked into SVM and XGBoost respectively, and calculating the SVM prediction result and the XGBoost prediction result; inputting the molecular representation of the graph of the molecule to be docked into GCN and GAT respectively, and calculating the GCN prediction result and the GAT prediction result;

[0031] S3: Calculate the weighted average of the SVM prediction results, XGBoost prediction results, GCN prediction results, and GAT prediction results to obtain the final prediction result.

[0032] The above one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects:

[0033] 1. Computational efficiency is significantly improved. In terms of computing time, the present invention uses an AMD EPYC7742 CPU (128 cores) with a main frequency of 2.25GHz and an NVIDIA GTX2080 GPU. The model training and screening of a 1 billion molecular library took a total of 950 hours. Compared with traditional molecular docking methods, such as Glide, the efficiency is improved by 180 times. The efficient computing power of the present invention significantly reduces the screening time and provides a practical solution for large-scale virtual screening.

[0034] 2. Optimize resource utilization. This invention effectively utilizes the differences and diversity between different models by integrating multiple regression algorithms and combining the sampling strategy with the highest prediction score and the largest standard deviation, thereby improving the performance and robustness of the overall model. By intelligently selecting the most challenging samples for annotation, unnecessary waste of computing resources is reduced, thereby significantly reducing computing costs and time investment while maintaining high accuracy.

[0035] 3. Excellent screening effect. By adopting the integrated regression model (ADEL model) of the present invention, 38 candidate compounds were screened from a virtual molecular library of more than 1 billion compounds, and then verified by wet experiments. Among the 38 screened compounds, 10 compounds showed strong biological activity, especially the IC 50 The value is less than 100nM. These compounds are expected to provide new candidate drugs for the treatment of β2AR-related diseases (such as asthma and chronic obstructive pulmonary disease), demonstrating the great potential of the ADEL model in the field of drug discovery. The ADEL model significantly improves the efficiency and accuracy of the drug discovery process. By reducing the computing time and resource consumption of virtual screening, researchers can more quickly identify and verify potential drug candidate compounds and accelerate the process of drug development.

[0036] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are 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.

[0038] Figure 1 It is a flow chart of the method provided by the present invention.

[0039] Figure 2It is a diagram of evaluation results of three sampling strategies provided by the present invention.

[0040] Figure 3 It is a result diagram of the ten-fold cross-validation RMSE of each round of iteration of the eight models provided by the present invention.

[0041] Figure 4 It is a result diagram of the hit rate of different quantity model combinations provided by the present invention.

[0042] Figure 5 It is a result diagram of the similarity of different number model combinations provided by the present invention.

[0043] Figure 6 It is a diagram of evaluation results of models with different iterative sample sizes provided by the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention. The following embodiments are used to illustrate the present invention, but cannot be used to limit the scope of the present invention.

[0045] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the embodiment of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0046] Combine the following Figures 1 to 6 The present invention is further described in detail, and an integrated regression model, generation method, and prediction method for molecular virtual screening of the present invention are described:

[0047] In this embodiment, Figure 1 As shown, a method for generating an integrated regression model for molecular virtual screening is provided, comprising the following steps:

[0048] Step 1: Randomly select K molecules from the ultra-large-scale compound library as the training set.

[0049] In this embodiment, the ultra-large-scale compound library includes the Enamine HTS data set containing 2.1 million molecules and the Enamine compound library containing 1 billion molecules.

[0050] Similar to traditional docking, the chemical library of active learning in this embodiment also needs to be processed. Therefore, the molecules selected from the ultra-large-scale compound library are processed, and the processing includes preprocessing and converting the molecules in SMILES format into ECFP4 fingerprints using RDKit. The preprocessing is to determine the stereoisomeric configuration and correct ionization state of the molecule. ECFP4 fingerprint is based on a fixed-length binary bit vector that encodes the presence or absence of a specific substructure. The bits of ECFP4 fingerprints are used as features of RF, SVM, Ridge, XGBoost, LGBM, and DNN models to learn which substructures are associated with high affinity (based on docking scores). In addition, the molecular representation of the graph of the molecule is used as input to GCN and GAT.

[0051] Step 2: Perform docking simulation on the molecules in the training set to obtain docking scores. The docking scores can be determined based on data such as binding affinity in the docking simulation.

[0052] Step 3: Using the docking scores, 8 regression models were trained using the training set.

[0053] The regression models include RF (Random Forest), SVM (Support Vector Machine), Ridge (Ridge Regression), XGBoost (Extreme Gradient Boosting), LGBM (Light Gradient Boosting), DNN (Deep Neural Network), GCN (Graph Convolutional Network) and GAT (Graph Attention Network).

[0054] Step 4: Determine whether the number of training times reaches the number of iterations. If not, proceed to step 5. If it reaches the number of iterations, proceed to step 7.

[0055] Step 5: Use the trained 8 regression models to predict the molecules that have not been selected in the ultra-large-scale compound library, and take the weighted average of the prediction results of the 8 regression models as the comprehensive prediction result.

[0056] Step 6: Select K molecules with the highest prediction scores and the largest standard deviations from the positive prediction pool of the comprehensive prediction results, perform docking simulation, obtain docking scores, and add them to the training set; proceed to step 3.

[0057] The process of selecting the K molecules with the highest prediction scores and the largest standard deviations is:

[0058] Sort all molecules according to their prediction scores and select the N molecules with the highest scores, N>K; schematically, N=1.5K;

[0059] Calculate the standard deviation of the predicted value for each of the N molecules and select the K molecules with the largest standard deviations.

[0060] The essence of this step is to expand the training set. After each round of iteration, the training set gradually increases, and the number of molecules in the training set increases by K molecules. The regression model is retrained using the expanded training set. The regression model can learn from new data, especially from samples with high scores but large prediction uncertainties in the previous round of screening, thereby obtaining new information to improve its prediction performance. The prediction accuracy and generalization ability of the regression model will also improve accordingly. It should be noted that the molecules are sampled from the positive prediction pool of the previous iteration, rather than from the original ultra-large-scale compound library.

[0061] Step 7: Evaluate the prediction performance of the eight regression models after training. The evaluation index of the prediction performance of the regression model is RMSE. Select the four regression models with the best prediction performance to obtain the integrated regression model.

[0062] In this embodiment, K is 1% of the total number of molecules in the ultra-large-scale compound library. For example, on the β2AR target, based on the Enamine HTS data set, 21,000 molecules are randomly selected from 2.1 million molecules as a training set.

[0063] On the β2AR target, based on the Enamine HTS dataset, three different sampling strategies were compared on the model performance through small batch iterations (each iteration increased by 0.1%, i.e. 2100 labeled samples). The three strategies are: Strategy 1, only select molecules with the highest scores; Strategy 2, only select molecules with the largest standard deviation (SD); Strategy 3, select molecules with the largest standard deviation among the highest scores.

[0064] The calculation formula of SD is as follows:

[0065]

[0066] in, represents the j-th regression model prediction value of the sample, n represents the number of regression models, and n is 8 in this embodiment. Represents the average predicted value of n regression models for the sample.

[0067] Evaluation results such as Figure 2 As shown, Figure 2 (a) is the evaluation result of the hit rate of three sampling strategies. Figure 2(b) shows the evaluation results of the similarity of the three sampling strategies. Strategy 3 has the best performance, indicating the feasibility and effectiveness of the annotation strategy of selecting compounds with the largest standard deviation of prediction values ​​among the compounds with the highest scores. Therefore, the sampling strategy of step 6 is to select K molecules with the highest prediction scores and the largest standard deviations.

[0068] This example also discusses the number of iterations and model ensembles under large-scale iterations (1% increase in each iteration, i.e., 21,000 labeled samples). After 8 rounds of model iterations, the average root mean square error (RMSE) of each 10-fold cross-validation round tends to converge, as shown in Figure 2. Figure 3 As shown in (a).

[0069] The calculation formula of RMSE is as follows:

[0070]

[0071] Where m represents the number of samples. represents the actual value of the i-th sample, Represents the model's predicted value for the i-th sample.

[0072] Using RMSE as the evaluation index, the present invention evaluates the prediction performance of a single regression model, such as Figure 3 As shown in (b), the models are arranged in descending order of performance, namely SVM, XGBoost, GCN, GAT, LGBM, Ridge, DNN and RF.

[0073] In order to explore the impact of model combination on performance, this embodiment forms 8 different numbers of model combinations, and the number of integrated models ranges from 1 to 8. The hit rate results are shown in Figure 4 As shown, Figure 4 (a) is a line chart of the hit rate results. Figure 4 (b) is a box plot of the hit rate results. The combination of SVM, XGBoost, GCN and GAT (4 models) has the best overall performance, with a hit rate of 63.8%, which is higher than the hit rate of the traditional molecular docking method Glide (43.4%). Figure 5 As shown, Figure 5 (a) is a line graph of similarity results. Figure 5 (b) is a box plot of the similarity results. The similarity of the combination of SVM, XGBoost, GCN and GAT models reaches 91.3%.

[0074] On the basis of determining the model combination, in order to further explore the impact of iterative sample size on the model, this example divides the sample size of each iteration into six different specifications of 0.1%, 0.2%, 0.4%, 0.6%, 0.8% and 1% of the data set (ultra-large-scale compound library). The results are as follows Figure 6 As shown, Figure 6 (a) is the evaluation result of hit rate. Figure 6 (b) is the evaluation result of similarity. As the sample size of each iteration increases, the hit rate and similarity both show an upward trend. In this embodiment, a model with a 1% iteration sample size is finally selected for subsequent research, that is, K is 1% of the total number of molecules in the large-scale compound library.

[0075] In terms of computing time, this example uses an AMD EPYC 7742 CPU (128 cores) with a main frequency of 2.25GHz and an NVIDIA GTX2080 GPU. It takes 950 hours to train the model and screen a library of 1 billion molecules. Compared with traditional molecular docking methods such as Glide, the efficiency is improved by 180 times. This efficient computing power significantly reduces the screening time and provides a practical solution for large-scale virtual screening.

[0076] In this embodiment, an integrated regression model for molecular virtual screening is also provided, and the technical solution adopted is as follows: 4 regression models are selected from 8 regression models as the integrated regression model by using the above-mentioned integrated regression model generation method for molecular virtual screening;

[0077] Regression models include: RF, SVM, Ridge, XGBoost, LGBM, DNN, GCN, and GAT;

[0078] The integrated regression model consists of SVM, XGBoost, GCN and GAT.

[0079] In this embodiment, a prediction method for molecular virtual screening is also provided, and the technical solution adopted is as follows: using the above-mentioned integrated regression model for molecular virtual screening, comprising the following steps:

[0080] S1: Obtain the molecule to be docked;

[0081] S2: inputting the ECFP4 fingerprint of the molecule to be docked into SVM and XGBoost respectively, and calculating the SVM prediction result and the XGBoost prediction result; inputting the molecular representation of the graph of the molecule to be docked into GCN and GAT respectively, and calculating the GCN prediction result and the GAT prediction result;

[0082] S3: Calculate the weighted average of the SVM prediction results, XGBoost prediction results, GCN prediction results, and GAT prediction results to obtain the final prediction result.

[0083] By using an integrated regression model (ADEL model) consisting of SVM, XGBoost, GCN and GAT, this example screened out 38 candidate compounds from a virtual molecular library of more than 1 billion compounds, and then verified them through wet experiments. The experimental results showed that among these 38 compounds, 10 had IC 50 The values ​​were lower than 5 μM, and the IC 50 The values ​​were 22.19nM and 97.51nM, respectively, showing extremely high biological activity. In addition, the K i The values ​​were lower than 40nM, and the specific values ​​were 17.53nM, 29.19nM, 30.29nM and 37.85nM respectively. This shows that the ADEL model has excellent performance and significant advantages in quickly and accurately screening highly active candidate molecules from ultra-large-scale compound libraries.

[0084] In terms of accuracy, the ADEL model has a screening hit rate of 26.3%, which is significantly higher than the median hit rate of traditional virtual screening methods: about 13% (data source: Journal of Medicinal Chemistry, 2013, 56(17): 6560-6572). In comparison, the hit rates of other similar methods are much lower than this. For example, Gryniukova et al. used an AI-driven virtual screening method to screen out 434 potential active molecules from 2.63 million compounds, and only 7 compounds had IC 50 The value is lower than 10 μM, and the hit rate is only 1.6% (data source: J. Med. Chem. 2023, 66, 15, 10241–10251). In contrast, this embodiment has significant technical advantages in hit rate and screening efficiency.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating an integrated regression model for molecular virtual screening, characterized in that: The following steps are involved: Step 1: Randomly select K molecules from the ultra-large-scale compound library as the training set; Processing of molecules selected from a very large compound library, including determination of the stereoisomer configuration and correct ionization state of the molecule, and conversion of the SMILES formatted molecules into ECFP4 fingerprints using RDKit; Step 2: Perform docking simulation on the molecules in the training set to obtain docking scores; Step 3: using the docking score and the training set to train n regression models; The regression model includes multiple ones of RF, SVM, Ridge, XGBoost, LGBM, DNN, GCN and GAT; Step 4: Determine whether the number of training times reaches the number of iterations. If not, proceed to step 5. If it reaches the number of iterations, proceed to step 7. Step 5: Use the trained n regression models to predict the molecules that have not been selected in the ultra-large-scale compound library to obtain a comprehensive prediction result; Step 6: Select K molecules with the highest prediction scores and the largest standard deviations from the positive prediction pool of the comprehensive prediction results, perform docking simulation, obtain docking scores, and add them to the training set; proceed to step 3; The process of selecting the K molecules with the highest prediction scores and the largest standard deviations is: Sort all molecules according to their prediction scores and select the N molecules with the highest scores, N>K; Calculate the standard deviation of the predicted value for each of the N molecules and select the K molecules with the largest standard deviation; Step 7: Evaluate the prediction performance of the n regression models after training, select the p regression models with the best prediction performance, and obtain the integrated regression model, n>p.

2. The method for generating an integrated regression model for molecular virtual screening according to claim 1, characterized in that: In step 5, the weighted average of the prediction results of the n regression models is taken as the comprehensive prediction result.

3. The method for generating an integrated regression model for molecular virtual screening according to claim 1, characterized in that: The K is 1% of the total number of molecules in the ultra-large-scale compound library.

4. The method for generating an integrated regression model for molecular virtual screening according to claim 1, characterized in that: The number of iterations is 8 times.

5. The method for generating an integrated regression model for molecular virtual screening according to claim 1, characterized in that: In step 7, RMSE is used as the evaluation index of the prediction performance of the regression model.

6. An integrated regression model for molecular virtual screening, characterized in that: Selecting 4 regression models from 8 regression models as integrated regression models by using an integrated regression model generation method for molecular virtual screening according to any one of claims 1 to 5; Regression models include: RF, SVM, Ridge, XGBoost, LGBM, DNN, GCN, and GAT; The integrated regression model consists of SVM, XGBoost, GCN and GAT.

7. A prediction method for molecular virtual screening, characterized in that: Using an integrated regression model for molecular virtual screening as claimed in claim 6 comprises the following steps: S1: Obtain the molecule to be docked; S2: inputting the ECFP4 fingerprint of the molecule to be docked into SVM and XGBoost respectively, and calculating the SVM prediction result and the XGBoost prediction result; inputting the molecular representation of the graph of the molecule to be docked into GCN and GAT respectively, and calculating the GCN prediction result and the GAT prediction result; S3: Calculate the weighted average of the SVM prediction results, XGBoost prediction results, GCN prediction results, and GAT prediction results to obtain the final prediction result.

Images