A method for constructing a use classification model based on 2D structure of a compound and application thereof

Abstract

This invention discloses a method and application for constructing a compound use classification model based on 2D compound structure, belonging to the field of high-throughput prediction of compound uses. This invention collects compound structures and uses from the CompTox Chemicals Dashboard database, characterizes compound structures using molecular descriptors and hybrid molecular fingerprints, then removes redundancy and inputs the data into a random forest regression model for feature selection, hyperparameter search, model training, and validation. By using the compound structure, it effectively predicts the compound's use, effectively filling the gap in current data where many chemicals have only structures but unknown uses, providing effective assistance and guidance for the control of chemicals in my country and globally.

Description

Technical Field

[0001] This invention belongs to the field of high-throughput prediction of compound uses, and more specifically, relates to a method for constructing and applying a use classification model based on the 2D structure of compounds. Background Technology

[0002] Driven by the demands of human society and the booming development of the chemical industry, a massive amount of chemicals are synthesized and produced. As of August 2023, over 274 million compounds were registered in the authoritative database CAS REGISTRY, but most of these compounds only have their structures and their uses are unknown. However, the use of a compound is usually one of the first pieces of information people want to know when searching for a compound. Currently, databases such as PubChem and CPDat only collect a small number of compound uses and related industries, which is insufficient to fill the gaps in the vast number of missing chemical uses. Therefore, there is an urgent need to establish a high-throughput, high-precision model that can effectively predict and classify uses or functions based solely on compound structures.

[0003] The uses of chemicals are typically related to their structure and functional groups. The CompTox Chemicals Dashboard database records multiple lists of chemicals organized by similar compound structures or uses. Its large number of compounds and high-quality use records provide a foundation for effectively constructing chemical structure-use training sets. Furthermore, random forests are a classic and effective algorithm for classification problems, often achieving good classification prediction results. Therefore, developing a chemical structure-use classification model based on random forests is crucial for effectively filling the gaps in understanding the unknown uses of a large number of chemicals. Summary of the Invention

[0004] 1. The problem to be solved

[0005] To address the problem of missing uses for many existing chemicals, this invention provides a method and application for constructing a use classification model based on the 2D structure of compounds. This invention collects chemical structures and uses from the CompTox Chemicals Dashboard database and establishes a random forest classification model. By using the compound's structure, it effectively predicts the compound's use, overcoming the deficiency of existing models that only provide structural information but lack application details. This contributes to the safe use of existing chemicals and the green synthesis of new chemicals, improving the understanding and regulation of chemicals.

[0006] 2. Technical Solution

[0007] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0008] The present invention provides a method for constructing an application classification model based on the 2D structure of compounds, comprising the following steps:

[0009] S10. Collect compound structure and use data: Collect a list of compounds with clear uses and supplement information on compounds with missing structures;

[0010] S20. Compilation of Compound Structure and Use List: Characterize the structure of the compounds in the compound list using molecular descriptors and mixed molecular fingerprints, and classify the corresponding compounds by their uses;

[0011] S30. Modeling data cleaning and feature selection: Remove features with low variability and strong collinearity, and use the recursive feature elimination algorithm to evaluate the importance of features and the impact of the number of features on the model's prediction accuracy.

[0012] S40. The dataset is divided into training and test sets: The compound structure-application data is randomly divided into training and test sets;

[0013] S50, Model Training: Put the training set data from step S40 into the random forest classification model for training, and use the grid search method to optimize the hyperparameters.

[0014] S60. Model performance verification: Put the test set data from step S40 into the trained random forest classification model and evaluate the model's prediction performance and effect.

[0015] Preferably, the specific process of step S10 is as follows: collect a list of compounds with clear uses from the CompTox Chemicals Dashboard database; for compounds in the compound list that are missing the SMILES structure, search for and supplement them in the Pubchem database by name and CAS number; and delete compounds from the compound list from which the SMILES structure cannot be found.

[0016] Preferably, the specific process of step S20 is as follows: the structural formulas of the compounds in the compound list are characterized using molecular descriptors and mixed molecular fingerprints, the uses of the corresponding compounds are organized into a series of clear use classifications, and the format is set to factors;

[0017] Preferably, the specific process of step S30 is as follows: for the molecular descriptor and mixed molecular fingerprint features obtained in step S20, further remove features with variance close to 0 and high collinearity, then standardize the remaining features, use a recursive feature elimination algorithm to select features, and retain the number of features corresponding to the maximum accuracy for subsequent model training.

[0018] Preferably, the specific process of step S40 is as follows: the compound structure-classification data are randomly divided into a training set and a test set in a ratio of 8:2.

[0019] Preferably, in step S50, a random forest classification model is constructed using 3 repetitions of 10x cross-validation. The training process is based on a grid search method to perform hyperparameter search for the minimum number of nodes for the important parameter. The number of nodes includes 1, 3, and 5. By performing hyperparameter search on the random forest model, the training effect of the model is evaluated for the minimum number of nodes for the important parameter being 1, 3, and 5, respectively.

[0020] Preferably, in both step S30 (recursive feature elimination algorithm) and step S50 (model training), the doParallel function is used to parallelize and accelerate model training, and the makeCluster and registerDoParallel functions are used to allocate the number of threads, wherein the number of threads is set to 50% of the maximum number of threads.

[0021] Preferably, in step S60, the test set data is classified and predicted using the predict function, and the specificity and sensitivity indices are further calculated using the confusionMatrix function to evaluate the model's predictive performance on the use of compounds in the test set.

[0022] Preferably, in step S20, the molecular descriptor and the mixed molecular fingerprint are calculated by PADEL software, and the maximum time for PADEL to calculate the molecular descriptor and molecular fingerprint of each compound is set to 30 seconds.

[0023] Preferably, in step S30, the nearZeroVar function in R language is used to remove features with variance close to 0, and the cor and findCorrelation functions are combined to remove features with high collinearity, where the collinearity is not less than 0.95; for the remaining features, the recursive feature elimination algorithm provided by the caret R package is used for feature selection, and the prediction effects of 25, 50, 100, 500, 1000 and all features are evaluated respectively, and the repeatedcv method is used to resample the samples.

[0024] The above-described method is used to construct an application model based on the 2D structure of a compound for predicting the use of the compound. The compound structure is input into the application classification model to predict the use of the input compound.

[0025] 3. Beneficial effects

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] (1) The present invention provides a method for constructing a classification model of the use of compounds based on 2D structure, which can effectively fill the gap that a large number of chemicals only have structures but unknown uses, which can help the safe use of existing chemicals and the green synthesis of new chemicals, and provide strong guidance and support for the subsequent control of new pollutants.

[0028] (2) The present invention provides a method for constructing a compound use classification model based on 2D compound structure. The list of compound structures and uses used in the model training and testing phases comes from the standard database CompTox Chemicals Dashboard. The amount of compound data is large, and the data on structure and use is more reliable. Moreover, these use classifications are usually related to human life, production and environmental pollutants. Predicting the use of these compounds will be more practical and will help management departments to trace and monitor compounds.

[0029] (3) The present invention provides a method for constructing a classification model for the use of compounds based on 2D structures. The PADEL software was used to calculate 1,444 2D molecular descriptors and 6,372 mixed molecular fingerprints for each compound. Using such mixed fingerprints can fully describe the structural features of the compounds and effectively avoid misclassification due to structural similarity. In the process of model training and application, it can effectively improve the classification effect of the model on the use of compounds.

[0030] (4) The present invention provides a method for constructing a classification model for the use of compounds based on 2D structures. By extracting key molecular descriptors and molecular fingerprints through feature selection, and by optimizing the hyperparameters of random forest classification through grid search, the training speed of the model and the prediction accuracy of compound uses can be greatly improved. This method can effectively provide a focus and foothold for the control of chemicals in my country and globally.

[0031] (5) The present invention provides a method for constructing a use classification model based on the 2D structure of compounds. The constructed use classification model is modular and can be supplemented and optimized at any time according to the use needs of government or enterprise units (such as compound use, modeling method, etc.). The model has a large training and prediction throughput, high prediction performance, can batch export compound use, and is friendly to people without machine learning background. It has good application prospects in compound use prediction. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating a method for constructing an application classification model based on the 2D structure of compounds according to the present invention.

[0033] Figure 2 This is a schematic diagram showing the number of compounds collected for various applications used in modeling and verification in this invention;

[0034] Figure 3 This is a schematic diagram of the molecular descriptor and the redundancy removal of each molecular fingerprint in this invention;

[0035] Figure 4 This is a schematic diagram illustrating the number of features and model accuracy of the present invention;

[0036] Figure 5 This is a schematic diagram of the hyperparameter optimization results of the random forest model of the present invention;

[0037] Figure 6 This is a schematic diagram illustrating the sensitivity and specificity of the various applications of this invention in the model validation stage;

[0038] Figure 7 This is a schematic diagram showing the distribution of compounds for various uses of the present invention as classified into other uses. Detailed Implementation

[0039] The present invention will be further described below with reference to specific embodiments.

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0041] The present invention provides a method for constructing an application classification model based on the 2D structure of compounds, comprising the following steps:

[0042] S10. Collect compound structure and use data: Search the chemical list recorded in the U.S. Environmental Protection Agency's CompTox ChemicalsDashboard database, retain compounds with clearly defined uses based on chemical names and detailed descriptions, collect a list of compounds with clearly defined uses, and for compounds in the compound list that are missing SMILES structures, search for and supplement them in the Pubchem database by name and CAS number. Compounds for which SMILES structures cannot be found are deleted from the compound list.

[0043] S20. Compilation of Compound Structure and Use List: The structural formulas of the compounds in the compound list are characterized using molecular descriptors and mixed molecular fingerprints. The corresponding compounds are classified according to their uses, and these are compiled into a series of clear categories. The format is set as a factor, and the structural formulas and use categories of the compounds are combined into a table as independent and dependent variables, respectively, for use in the subsequent random forest model building.

[0044] It should be noted that the structural formula of the compound is represented by a molecular descriptor calculated by PADEL software and a mixed molecular fingerprint. The mixed molecular fingerprint is generated by mixing molecular fingerprints calculated using at least two different methods, and for optimal results, it should contain at least 10 sets of molecular fingerprints. The molecular fingerprints include:

[0045] CDK fingerprint;

[0046] CDK extended fingerprint;

[0047] Estate fingerprint;

[0048] CDK graph only fingerprint;

[0049] MACCS fingerprint;

[0050] Pubchem fingerprint;

[0051] Substructure fingerprint;

[0052] Substructure fingerprint count;

[0053] 2D atom pairs;

[0054] 2D atom pairs count;

[0055] The maximum time PADEL takes to calculate the molecular descriptor and molecular fingerprint for each compound is set to 30 seconds, with other parameters left at their default values. Subsequently, compounds for which molecular descriptors cannot be calculated are removed, and compounds with nC greater than 1 in their molecular descriptors are retained (i.e., organic compounds are retained).

[0056] The uses of compounds are categorized into 14 major categories that are widely used in human society, including surfactants, antioxidants, pesticides, perfluorinated compounds, biological-related compounds, food, plastics, natural toxins, natural products, rubber, disinfection byproducts, pharmaceuticals, raw materials, and flame retardants. Among these, raw materials mainly refer to some industrial raw materials, while biological-related compounds mainly refer to some endogenous substances.

[0057] S30. Modeling Data Cleaning and Feature Selection: Based on the molecular descriptors and mixed molecular fingerprint features compiled in step S20, the `nearZeroVar` function in R is used to remove features with variance close to 0. The `cor` and `findCorrelation` functions are then used to remove highly collinear features with collinearity exceeding 0.95. The remaining features are then standardized, and a recursive feature elimination algorithm is used for feature selection, retaining the number of features corresponding to the maximum accuracy for model building. For the remaining features, the recursive feature elimination algorithm provided by the `caret` R package is used for feature selection. The number of features evaluated by the recursive feature elimination algorithm should be reasonably distributed between 0 and all features. For example, the prediction effects of 25, 50, 100, 500, 1000, and all features are evaluated respectively. The `repeatedcv` method is used to resample the samples. Furthermore, the number of retained features should generally be the maximum value of the feature number and model fitting accuracy curve. To improve training speed, the number of features corresponding to the plateau period of the curve can also be retained as appropriate.

[0058] S40. The dataset is divided into training and test sets: The compound structure-classification data is randomly divided into training and test sets in a ratio of 8:2.

[0059] S50, Model Training: The training set data from step S40 is put into the random forest model for training. The random forest classification model is constructed using 3 repetitions of 10x cross-validation. The training process is based on the grid search method to search for the minimum number of important parameters and the number of nodes. The number of nodes includes 1, 3 and 5. By searching for the hyperparameters of the random forest model, the training effect of the model is evaluated when the minimum number of important parameters is 1, 3 and 5 respectively.

[0060] S60. Model Performance Validation: Input the test set data from step S40 into the trained random forest use classification model to evaluate the model's predictive performance and effectiveness. The `predict` function is used to predict the classification of the test set data, and the `confusionMatrix` function is used to calculate specificity and sensitivity indices to evaluate the model's predictive effectiveness for the uses of compounds in the test set.

[0061] It should be further noted that in both step S30 (recursive feature elimination algorithm) and step S50 (model training), the doParallel function is used to parallelize and accelerate model training, and the makeCluster and registerDoParallel functions are used to allocate the number of threads, with the number of threads set to 50% of the maximum number of threads.

[0062] Example 1

[0063] like Figure 1 As shown in this embodiment, a method for constructing an application classification model based on the 2D structure of compounds specifically includes the following steps:

[0064] S10. Collect data on compound structure and applications:

[0065] A search of the U.S. Environmental Protection Agency's CompTox Chemicals Dashboard database for Chemical Lists yielded 424 chemical lists. Based on the names and detailed descriptions of the chemical lists, 56 lists with clearly defined uses were retained for modeling. The list of 56 compounds is shown in Table 1 below.

[0066] For compounds without SMILES structures, the Pubchem database is searched and supplemented by the compound name or CAS number. Compounds for which SMILES structures cannot be found will be removed from the compound list because these compounds are mainly mixtures and it is difficult to describe their structures.

[0067] S20, List of Compound Structures and Uses:

[0068] The compound list from step S10 was compiled into a table of compound SMILES structures and their uses. After removing duplicate values ​​from the table, 1444 2D molecular descriptors and 6372 mixed molecular fingerprints were further calculated using PADEL software based on the compound SMILES structures. The mixed molecular fingerprints included:

[0069] 1024 CDK fingerprints;

[0070] 1024 CDK extended fingerprint;

[0071] 79 Estate fingerprints;

[0072] 1024 CDK graph-only fingerprints;

[0073] 166 MACCS fingerprints;

[0074] 881 Pubchem fingerprints;

[0075] 307 Substructure fingerprints;

[0076] 307 Substructure fingerprint count;

[0077] 780 2D atom pairs;

[0078] 780 2D atom pairs count;

[0079] The maximum time for PADEL to calculate the molecular descriptor and molecular fingerprint for each compound was set to 30 seconds, while other calculation parameters remained at their default values. Subsequently, compounds for which molecular descriptors could not be calculated were removed, and compounds with nC greater than 1 in their molecular descriptors (i.e., organic compounds were retained) were preserved.

[0080] This embodiment collected a total of 36,940 compounds and 46,762 compound structure-use entries (there were cases where a compound was included in multiple categories). Perfluorinated compounds, pharmaceuticals, and plastics applications each had more than 8,000 entries, making them the three largest categories of compound applications. Figure 2 As shown.

[0081] The uses of compounds were categorized into 14 distinct classes, including surfactants, antioxidants, pesticides, perfluorinated compounds, bio-related compounds, food, plastics, natural toxins, natural products, rubber, disinfection byproducts, pharmaceuticals, raw materials, and flame retardants. Raw materials primarily refer to industrial raw materials, while bio-related compounds mainly consist of endogenous substances. These 14 categories play crucial roles in normal human production and daily life, involving a vast number of compounds; therefore, selecting these categories as the model's prediction endpoints is highly meaningful. Subsequently, the structure and uses of compounds were combined into a table as independent and dependent variables, respectively, and the format of the use categories was further set as factors for subsequent model building.

[0082] S30. Modeling data cleaning and feature selection:

[0083] For the molecular descriptors and mixed molecular fingerprints in the table compiled in step S20, the nearZeroVar function was used to remove low-variability molecular descriptors with variance close to 0. Subsequently, cor and findCorrelation were combined to remove molecular descriptors with collinearity exceeding 0.95. Each compound retained 612 molecular descriptors and 2855 molecular fingerprints, totaling 3467 features and at least one classification information. The retention rate of molecular descriptors was 58%, and the retention rate of molecular fingerprints was 55%. Among them, CDK fingerprint, CDK extended fingerprint, and CDK graph only fingerprint each retained more than 400 features, making them the top three molecular descriptors with the most retained features. Figure 3 As shown.

[0084] The 3467 features were further selected using the recursive feature elimination algorithm provided by the caret R package to evaluate the optimal number of features for model prediction performance. Therefore, the rfe function was used to evaluate the model prediction accuracy with 25, 50, 100, 500, 1000, and all features, and the repeatedcv method was used to resample the samples. During the feature selection process, we found that although the model approached saturation with 1000 variables, the best model prediction accuracy was still achieved with all variables. Therefore, in this embodiment, no features were deleted during feature selection; instead, all features were included in the model for training. Figure 4 As shown.

[0085] Data cleaning and feature selection remove redundant data with minimal information loss, significantly accelerating model training and effectively improving prediction performance. Finally, the molecular descriptors and mixed molecular fingerprints of compounds are standardized and used as independent variables, while the uses of the compounds are used as dependent variables for subsequent model training.

[0086] S40. The dataset is divided into a training set and a test set:

[0087] The 46,762 data points were randomly divided into a training set (37,410 data points) and a test set (9,352 data points) in an 8:2 ratio.

[0088] S50, Model Training:

[0089] Based on the training set data from step S40, a random forest classification model was constructed using three repeated 10x cross-validation iterations. The training process employed a grid search method to search for the minimum number of nodes (1, 3, and 5) as the key parameters of the random forest model. Through training, the model accuracy was better with a minimum number of nodes of 5 than with 1 and 3. Therefore, in the formal training of the model, the minimum number of nodes was set to 5. Figure 5 As shown.

[0090] In this embodiment, the doParallel function is used to parallelize and accelerate model training in both step S30 (recursive feature elimination algorithm) and step S50 (model training), and the makeCluster and registerDoParallel functions are used to allocate the number of threads.

[0091] S60, Model Performance Verification:

[0092] Based on the training set, the model's accuracy in predicting compounds was 0.7564 (95% confidence interval: 0.7476–0.7651), the model p-value was <2.2 × 10⁻¹⁶, and the Kappa value was 0.7088. This indicates that the random forest classification model successfully captured the features of most use classifications. Subsequently, the test set data from step S40 was fed into the trained random forest classification model, and the use classification of the test set data was predicted using the predict function. Furthermore, specificity, sensitivity, and other indicators were calculated using the confusionMatrix function.

[0093] The results show that the specificity for all use categories exceeds 0.9, indicating that the model effectively controls false positives. In terms of sensitivity for use categories, perfluorinated compounds, pharmaceuticals, raw materials, bio-related compounds, flame retardants, and plastics all exceed 0.7. The use prediction model in this embodiment demonstrates excellent predictive performance for these compounds, indicating that the model successfully identifies the characteristics of compounds for these uses. The predictive sensitivity for food, surfactants, disinfection byproducts, and natural toxins exceeds 0.3, indicating good predictive performance. However, the predictive performance for compounds for antioxidants, natural products, and rubber uses is relatively poor. This may be due to the high structural specificity and multi-classification of compounds in these uses. For example, many drugs used in human activities and production are derived from the purification and extraction of natural products, which to some extent hinders the model's judgment of the uses of these compounds. Figure 6 As shown.

[0094] This embodiment further investigates the misclassification of antioxidants, natural products, and rubber compounds with poorly predicted uses. Antioxidants were primarily misclassified as bio-related and pharmaceutical compounds, likely due to insufficient data on their applications. Natural products were mainly misclassified as pharmaceuticals, a result that can be reasonably explained since many active pharmaceutical ingredients originate from plant and microbial synthesis. Rubber was largely classified as plastic, primarily due to its intended use; rubber is often studied as a special type of plastic. Therefore, further adjustments and optimizations to the intended uses of these compounds are necessary. Figure 7 As shown.

[0095] In summary, the method for constructing a use classification model based on the 2D structure of compounds established in this embodiment can effectively predict the uses of most compounds. The lower prediction performance is mainly due to the overlap and intersection of compound uses, which does not affect the model's judgment on the uses of most compounds. Therefore, this model will become an important auxiliary tool for chemical safety supervision and use, greatly filling the gaps in the lack of usage information for many chemicals in my country and globally.

[0096] Table 1 List of Chemicals with Specific Uses

[0097]

[0098]

[0099]

[0100]

[0101] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the data used is only one embodiment of the present invention. The actual combination of data is not limited to this. Therefore, if those skilled in the art are inspired by this description and, without departing from the spirit of the present invention, devise similar embodiments and examples of the technical solution without creative design, all such embodiments and examples should fall within the protection scope of the present invention.

Claims

1. A method for constructing an application classification model based on the 2D structure of compounds, characterized in that: Includes the following steps: S10. Collect compound structure and application data: Collect a list of compounds with clear applications from the CompTox Chemicals Dashboard database. For compounds in the compound list that are missing SMILES structures, search for and supplement them in the Pubchem database by name and CAS number. Compounds for which SMILES structures cannot be found are deleted from the compound list. S20. Compilation of Compound Structure and Use List: The structures of the compounds in the compound list are characterized using molecular descriptors and mixed molecular fingerprints, and the corresponding compounds are classified by use. The uses of the compounds are organized into 14 major categories that are widely present in human society, including surfactants, antioxidants, pesticides, perfluorinated compounds, biological-related, food, plastics, natural toxins, natural products, rubber, disinfection by-products, drugs, raw materials, and flame retardants, and the format is set to factors. S30. Modeling data cleaning and feature selection: Remove features with low variability and high collinearity, and use the recursive feature elimination algorithm to evaluate the importance of features and the impact of the number of features on the model's prediction accuracy. S40. The dataset is divided into training and test sets: The compound structure-application data is randomly divided into training and test sets; S50, Model Training: Put the training set data from step S40 into the random forest classification model for training, and use the grid search method to optimize the hyperparameters. S60. Model performance verification: Input the test set data from step S40 into the trained random forest classification model and evaluate the model's prediction performance and effectiveness. In both step S30 (recursive feature elimination algorithm) and step S50 (model training), the `doParallel` function is used to parallelize and accelerate model training. The `makeCluster` and `registerDoParallel` functions are used to allocate the number of threads, with the number of threads set to 50% of the maximum number of threads. In step S30, the `nearZeroVar` function in R is used to remove features with variance close to 0, and the `cor` and `findCorrelation` functions are used to remove features with high collinearity, where the collinearity is not less than 0.

95. For the remaining features, the recursive feature elimination algorithm provided by the `caret` R package is used to select features. The prediction effects of 25, 50, 100, 500, 1000 and all features are evaluated respectively, and the `repeatedcv` method is used to resample the samples.

2. The method for constructing an application classification model based on the 2D structure of a compound according to claim 1, characterized in that: The specific process of step S40 is as follows: the compound structure-classification data are randomly divided into training set and test set according to an 8:2 ratio.

3. The method for constructing an application classification model based on the 2D structure of a compound according to claim 1, characterized in that: In step S50, a random forest classification model is constructed using 3 repetitions of 10x cross-validation. The training process is based on a grid search method to perform hyperparameter search for the minimum number of nodes for the important parameter. The number of nodes includes 1, 3, and 5. By performing hyperparameter search on the random forest model, the training effect of the model is evaluated under the conditions of 1, 3, and 5 nodes for the minimum number of nodes for the important parameter.

4. The method for constructing an application classification model based on the 2D structure of a compound according to claim 1, characterized in that: In step S60, the predict function is used to predict the classification of the test set data, and the specificity and sensitivity indices are further calculated using the confusionMatrix function to evaluate the model's predictive performance on the use of compounds in the test set.

5. The method for constructing an application classification model based on the 2D structure of a compound according to claim 1, characterized in that: In step S20, the molecular descriptor and the mixed molecular fingerprint are calculated by PADEL software, and the maximum time for PADEL to calculate the molecular descriptor and molecular fingerprint of each compound is set to 30 seconds.

6. The application of a compound use classification model based on the 2D structure of a compound obtained by the construction method according to any one of claims 1-5 in predicting the use of a compound, characterized in that: By inputting the structure of a compound into the application classification model, the application of the input compound is predicted.

Images