Method and system for data mining of energetic compounds

Abstract

The present invention discloses a method and system suitable for data mining of energetic compounds, belonging to the technical field of structure-activity relationship of energetic compounds in chemical informatics calculation, comprising the steps of: S1, data preprocessing; S2, extracting molecular topological structure features and features of quantum chemical calculation to generate fingerprint features; S3, calculating Pearson correlation and information entropy, screening out linearly correlated and non-correlated features, performing feature dimensionality reduction, and deleting redundant features whose linear correlation between features is higher than a set value; S4, knowledge mining and model verification, exploring the correlation and knowledge association between feature libraries. The present invention can classify data of energetic materials according to skeletons and groups, mine multi-dimensional linear features and non-linear features, embed more features based on chemical expertise into the graph structure model training process, and utilize the data mining engine of the present invention to screen out strongly correlated molecular features, which can make up for the shortcomings of existing models.

Description

Technical Field

[0001] The present invention relates to the technical field of structure-activity relationship of energetic compounds in chemical informatics calculations, and more specifically, to a method and system suitable for data mining of energetic compounds. Background Art

[0002] Energetic materials are compounds or mixtures containing explosive groups or oxidants and combustibles that can independently undergo chemical reactions and output energy. They are essential components of military explosives, propellants, and rocket propellant formulations, and are widely used in defense, aerospace, and civilian applications. Because the experimental synthesis of such materials is dangerous and costly, and practical applications place high demands on sensitivity and energy properties (low sensitivity: good thermal stability, low mechanical sensitivity; detonation performance: excellent detonation pressure, detonation velocity, and detonation heat), accelerating the research and development of energetic materials through theoretical calculations is of great research value.

[0003] Since the 1950s, the rise of numerical simulation using computers, and the ability to accurately and reliably simulate molecular conformations based on density functional theory, have greatly accelerated the research of energetic materials. However, the computational process of such molecular simulations is cumbersome and time-consuming, consuming significant computational costs. The need to develop faster and more accurate structure-activity relationship prediction models is pressing.

[0004] Since the beginning of this century, the performance of computers has been greatly enhanced, and corresponding to this is the rapid growth of data, which has promoted the development of big data and artificial intelligence technology. Existing machine learning methods have long been used in many important properties of energetic materials. The most cutting-edge research has achieved a huge improvement in model performance using molecular graphs as input graph neural networks. However, in the field of energetic materials, there are problems such as a shortage of data samples and uneven data distribution. Secondly, certain chemical properties such as decomposition and deflagration processes involve complex reaction mechanisms, which require more feature embedding graph structure model training based on chemical expertise. Therefore, screening out strongly correlated molecular features through data mining engines to make up for the shortcomings of existing models has become the key to research in this field. The technical solution of the present invention aims to provide a solution to the above technical problems. Summary of the Invention

[0005] The present invention aims to overcome the shortcomings of the existing technology by providing a method and system for data mining of energetic compounds. This method addresses the aforementioned issues: uniform data distribution and a more comprehensive feature embedding graph structure model training process based on chemical expertise. The data mining engine of the present invention can identify strongly correlated molecular features, addressing the shortcomings of existing models.

[0006] The object of the present invention is achieved through the following solutions:

[0007] A method for data mining of energetic compounds, comprising the following steps:

[0008] S1, Data Preprocessing: Obtain energetic molecular structures and their corresponding physicochemical property values. Based on the component and structure screening conditions, screen for compounds containing only the four elements CH, N, and O, as well as single-molecule electrically neutral compounds containing nitro groups in their structures, where the number of atoms in the molecule is less than the set range. De-duplicate the screening results, encode them, and store them as a dataset. The dataset is then classified according to the type of nitro-bonded chemical environment, constructing training and test sets with evenly distributed types.

[0009] S2, calculation features: extract the pre-processed molecular topological structure features and quantum chemical calculation features respectively, generate fingerprint features, build a feature library, and establish multiple molecular feature libraries for different material properties. One feature library contains all calculated features; including the following types: molecular 2D structural features extracted using the Rdkit tool, molecular topological structure features related to the field of energetic materials defined in this invention, and features calculated using quantum chemistry, to complete the construction of the feature library;

[0010] S3, feature screening: Pearson correlation and information entropy are calculated to screen out linearly correlated and uncorrelated features, and recursive feature elimination (RFECV) is used to reduce the feature dimension and delete redundant features whose linear correlation between features is higher than the set value.

[0011] S4, knowledge mining and model validation: Discover the correlation and knowledge association between feature libraries, verify the strong correlation between attributes and features from the perspective of expert knowledge, splice fingerprint features as input to the machine learning model, adjust parameters, and use cross-validation methods on different models to verify and evaluate features to obtain the optimal quantitative structure-property relationship (QSPR) model for data mining of energetic compounds.

[0012] Furthermore, in step S1, the molecular structure is stored in regular SMILES encoding, and the determination of whether the molecular formula contains elements other than CHON is implemented by Python's character processing molecular formula, and the generated regular SMILES set is used for deduplication; it is determined whether the molecule contains a nitro group and the connection type of the nitro group is recorded, and the situation where the molecular structure contains ionic compounds or hydrates is excluded, and the total number of atoms contained in the molecule is controlled within eighty; and, hydrates and ionic salts are implemented by string comparison, and the search and type of nitro groups are implemented by substructure search and molecular graph algorithm search; nitro groups are divided into seven categories according to the type of bonding chemical environment, including chain oxygen nitro groups, chain nitrogen nitro groups, chain carbon nitro groups, ordinary ring nitrogen nitro groups, ordinary ring carbon nitro groups, aromatic ring nitrogen nitro groups, and aromatic ring carbon nitro groups; the training set and the test set are constructed according to the set ratio to ensure that each type of molecule is divided according to the same ratio, and then the training set and the test set are added in sequence to ensure the rationality of the data division.

[0013] Furthermore, in step S2, the fingerprint features are generated, and multiple molecular feature libraries are established for different material properties. One feature library contains all the calculated features, including the following sub-steps:

[0014] Step (1): Using Rdkit software, calling the rdMolDescriptors module to calculate the 2D descriptor, the Morgan fingerprint method is used to calculate the fingerprint features, and used to describe the extended connectivity of the molecular structure;

[0015] Step (2): Design features to describe the topological structure of energetic materials;

[0016] Step (3): Extract the features of the DFT calculation file.

[0017] Furthermore, in step S3, the screening feature covers linear and nonlinear screening methods, specifically including the following sub-steps:

[0018] Step 1): Use Pearson correlation coefficient to screen features;

[0019] Step 2): Use the nonlinear evaluation method Copula negative entropy to calculate the correlation of nonlinear features;

[0020] Step 3): Select any machine learning model and use the recursive feature elimination (RFECV) method to expand the candidate feature library. The procedure is iterated n times to solve the accidental effects of the splitting of the training dataset and the random selection process of features. n is a positive integer.

[0021] Step 4): Calculate the feature correlation matrix, remove redundant features whose Pearson linear correlation between features is greater than the set value, and give priority to retaining features with strong correlation with attribute values.

[0022] Furthermore, in step S4, the different models include support vector regression SVR, ridge regression KRR, and random forest.

[0023] Furthermore, in step S4, the method of discovering the correlation and knowledge association between feature libraries also includes the steps of performing linear and nonlinear analysis on the features, analyzing the correlation between the attribute values ​​and those features from the perspective of chemical mechanism, thereby mining the chemical knowledge in big data.

[0024] Furthermore, in step (1), the 2D descriptors include descriptors characterizing TPSA, MolLogP, Charges, and hydrogen bond donor and acceptor properties.

[0025] Furthermore, in step 3), the step is included: putting the features supported by the model n times into a feature candidate library.

[0026] Furthermore, step S4 includes a sub-step of normalizing the features to make the data more separable.

[0027] A system suitable for energetic compound data mining includes a computer device; the computer device includes a processor and a memory, the memory stores a computer program, and when the computer program is loaded by the processor and executes any of the methods described above.

[0028] The beneficial effects of the present invention include:

[0029] The technical solution of the present invention mines multi-dimensional features from topological calculations and high-throughput calculation results based on energetic molecules, and provides a method that can be used to build an expert knowledge base of material properties and enhance the performance of QSPR models.

[0030] First, in step S1, the data preprocessing process can not only ensure that the energetic material molecules in the data set meet the requirements on composition and structure, but also cope with the current uneven data distribution in the field of energetic materials. The data preprocessing divides the data set into seven categories according to the type of chemical environment of the most important energetic group bonding, the nitro group, and the training set and test set with uniform distribution of structural types. The lack of experimental data on energetic materials is an important reason for the difficulty in accurately predicting the properties of high-performance materials. In particular, for a small number of energetic molecules with excellent performance, due to the complexity and particularity of their physical and chemical mechanisms, the existing technologies and methods are difficult to handle the prediction of properties similar to sensitivity properties or involving complex reaction mechanisms due to the lack of domain knowledge. Therefore, the present invention designs a feature library construction method covering topological features and DFT features in step S2. The results of the invention examples show that these features are novel and effective for the thermal stability properties of energetic molecules. The second major challenge facing the prediction of energetic material properties is model overfitting. With the development of deep learning, the complexity of the model increases. The linear screening methods (such as Pearson correlation) in previous studies no longer meet the needs and have great limitations in improving the generalization ability of the model. Therefore, the present invention proposes a feature screening scheme that combines linear features with nonlinear features, which solves the problems of nonlinear feature loss and model overfitting. Finally, the obtained high-correlation features are input into the machine learning model in step S4 for cross-validation. The fingerprint, like other molecular features in the feature library, can be used to construct a one-dimensional vector of the same format as the input for model training. The fingerprint represents the extended connectivity information of the entire molecule, and the other features represent multidimensional information related to the physical and chemical mechanisms, which can be used to enhance the model as domain knowledge embedding. In order to integrate the information represented by the two, the present invention creatively adopts the method of constructing a vector by splicing fingerprint features to further enhance the performance of the model.

[0031] In summary, the technical aspects of the present invention are based on the conceptual design of data mining. The engine for data mining of energetic compounds constructed by the present invention can extract, calculate, and screen out strongly correlated molecular features, so that more features based on chemical expertise can be embedded in the training process of the energetic molecule structure-activity relationship model, which can well make up for the shortcomings of the existing model. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0033] Figure 1 This is a flowchart of the method steps of an embodiment of the present invention;

[0034] Figure 2 This is a flow chart of the steps of the method for mining data on the decomposition temperature of energetic materials in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] All features disclosed in all embodiments in this specification, or steps in all methods or processes implicitly disclosed, except for mutually exclusive features and / or steps, can be combined and / or expanded or replaced in any manner.

[0036] The technical solution of the present invention belongs to the technical field of structure-activity relationship of energetic compounds in chemical informatics calculations, and in particular relates to molecular structure characterization and calculation of molecular characteristics (high-throughput computing characteristics and topological structure characteristics), as well as a method for evaluating the correlation (linear and nonlinear relationship) between these characteristics and the physicochemical properties of materials.

[0037] The technical solution of the present invention is as follows: Utilizing the technical solution of the present invention, a data mining engine for energetic material compounds is provided. Multidimensional features are mined from topological and high-throughput calculations based on energetic molecules, and used to construct an expert knowledge base of material properties and enhance the performance of QSPR models. The technical solution of the present invention includes a data mining method for correlating properties such as sensitivity and energy of energetic materials with their molecular structures. Generally speaking, the method comprises the following steps: 1) extracting molecular descriptors calculated using the Rdkit chemical information processing tool and custom-designed descriptors related to energetic molecules (structural descriptors related to nitro groups, nitrogen atoms, and cyclic structures); 2) extracting conformational information from DFT calculation files (.log and .chk), descriptors related to chemical surface charge analysis, nitro bond dissociation energy (BDE), nitro group charge, three-dimensional distance-related features, planarity, formation enthalpy, and performance descriptors calculated using the KJ equation; 3) building upon steps 1 and 2), constructing a feature library, inputting feature linear correlation filters and nonlinear feature filters to obtain a feature space with strong correlations among specific properties, and then performing feature redundancy removal. This process can be summarized as feature dimensionality reduction. These features are input into several machine learning models, and the models are optimized through parameter optimization and cross-validation. The feature inputs can be connected to form feature vectors by fingerprints, or input independently and freely combined according to the above classification.

[0038] In a further inventive concept of the present invention, a method for energetic compound data mining is provided, comprising the following steps:

[0039] Step 1: Preprocess the data to obtain the structure file of the energetic material elemental explosive molecule and its corresponding physical and chemical property value, screen compounds containing only four elements of CH, N, O, and Z, single molecular structures containing nitro groups, and neutral compounds, and store them in canonical smiles encoding.

[0040] Step 2: Calculate features. The calculation program based on molecular topology is implemented in a Python environment that integrates the Rdkit chemical information calculation tool. Quantum chemical calculations produce two result files with suffixes of .log and .wfn. The extraction of wfn is implemented by calling the MultiWFN program, and the extraction of feature information in the result files is uniformly implemented using Python scripts.

[0041] Step 3: Feature screening. Pearson correlation coefficient, information entropy, and recursive feature elimination (RFECV) methods are used to perform feature screening and dimensionality reduction to obtain linearly correlated and non-linearly correlated features.

[0042] Step 4: Knowledge mining and model verification, to explore the correlation and knowledge association between feature libraries. On the one hand, it can verify the strong correlation between a certain attribute and feature from the perspective of expert knowledge. On the other hand, cross-validation methods are used on different models to verify and evaluate features and obtain the optimal model of QSPR.

[0043] In the above technical solution, a further inventive concept lies in: in step 1, the data preprocessing method specifically includes the following steps: reading the molecular Mol file, determining whether the molecular formula contains elements other than CHON, generating a canonical SMILES set for deduplication, determining whether the molecule contains a nitro group and recording the nitro group's connection type, eliminating any unusual structures such as ionic compounds or hydrates in the molecular formula, and limiting the total number of molecular atoms to less than 80. Hydrates and ionic salts are compared using string comparison, and nitro groups are searched and recorded using substructure search and molecular graph algorithms. Nitro groups are classified into seven categories based on the type of chemical environment: chain oxygen nitro groups, chain nitrogen nitro groups, chain carbon nitro groups, ordinary ring nitrogen nitro groups, ordinary ring carbon nitro groups, aromatic ring nitrogen nitro groups, and aromatic ring carbon nitro groups. At this stage, the distribution of data for each category can be visualized. Taking the decomposition temperature dataset of energetic materials as an example, data with an initial decomposition temperature below 50°C are removed. Rdkit was used to perform two steps of initial spatial coordinate optimization (EmbedMolecule method) and universal force field conformation optimization (UFFOptimizeMolecule method) on the Mol object, and then it was saved again for subsequent DFT calculations.

[0044] In the above technical solution, a further inventive concept is that in step 2, the detailed method for calculating the feature includes the following steps:

[0045] (1) Using Rdkit software, the rdMolDescriptors module was called to calculate 2D descriptors, including but not limited to 200 descriptors characterizing TPSA, MolLogP, Charges, hydrogen bond donor and acceptor properties, etc. Fingerprint features are used to describe the extended connectivity of molecular structures. The Morgan fingerprint method was used for calculation, with the connectivity radius set to 2 and the bit length set to 2048, taking into account characteristic and chiral molecules.

[0046] (2) In addition to the classification information of the chemical environment of the nitro group, more features are designed to describe the topological structure of the following energetic materials, including the type of nitro group that can be determined in step 1, the type of azole and its number of occurrences, the minimum or maximum length of the bond of the nitro nitrogen atom in the conformation, the maximum number of adjacent appearances of three identical atoms C, N, and O in the molecular structure, the number of nitroamines, the ratio of aromatic atoms, the ratio of nitro atoms, the number of amino groups, the number of bridging oxygens such as NON or COC structures, whether it contains a fused ring, and the ratio of atoms on the ring.

[0047] Furthermore, the maximum number of consecutive occurrences of the same backbone atoms is obtained. The steps designed by the method of the present invention are described as follows:

[0048] Step 1, obtain the atomic list of a certain skeleton of the molecule;

[0049] Step 2: Get the traversed atom list, apply a recursive algorithm to each atom, and record the recursive return result;

[0050] Step 3: In the recursive algorithm, input the atom index and traverse the neighboring atoms of the atom. If the recursive condition is met, the neighboring atom is used as the next input. The maximum number of adjacent atoms is equal to the recursive depth plus 1. Atoms that have been traversed are recorded and not traversed again. The initial recursive return value is set to 1.

[0051] Step 4: Obtain the maximum value of the recursive algorithm result on the atom list as the result of the algorithm.

[0052] (3) Python is used to extract features from DFT calculation files. The features included are the following types: molecular surface electrostatic potential analysis results, spin density, chemical bond distances between non-hydrogen atoms, HOMO atomic aromaticity index, stacking index, molecular shape and radius, band gap, the strongest and weakest bond lengths of NO2-R nitro group, and nitro group charge. The extraction and calculation methods are as follows:

[0053] 1) On the one hand, directly intercept relevant paragraphs and extract feature information from log files and wfn files, including the following steps:

[0054] Step 1: Read the file to obtain the file IO stream;

[0055] Step 2: Compile a regular expression between the start character and the end character to match the feature text;

[0056] Step 3: Build a dictionary based on the property name as the return result.

[0057] 2) On the other hand, accurate conformational information is obtained from the log file to calculate the interatomic distance and nitro bond length, including the following steps:

[0058] Step 1, intercepting the bond level table segment containing atomic connection information;

[0059] Step 2, parse out the two-dimensional matrix of interatomic distances according to the bond level table;

[0060] Step 3, traverse the bond order table and two-bit matrix except hydrogen atoms, and save the spacing values ​​greater than 0.05;

[0061] Step 4: Construct an atomic distance type dictionary and update the maximum and minimum values ​​as the return result.

[0062] In the above technical solution, a further inventive concept is that in step three, the feature screening method includes the following steps:

[0063] 1) Use the Pearson correlation coefficient to screen features. The calculation principle can be expressed as the following formula, where n is the sample size. is the sample mean, S x ,S y Represents the standard deviation of x and y variables respectively:

[0064]

[0065] Items that satisfy the Pearson correlation coefficient > 0.3 and P-value < 0.05 are added to the candidate features.

[0066] 2) The nonlinear evaluation method Copula negentropy is used to calculate the correlation of nonlinear features. The calculation principle is shown in the formula. Copula negentropy can be widely used in various fields. Its definition is shown in the following formula, where x is a multi-random variable, u is a multi-random variable of x, and c is the Copula density function:

[0067] Hc(x)=-∫ u c(u)logc(u)du.

[0068] The Copula code library is called for calculation, and the negative entropy value greater than 0.1 is input into the candidate feature library.

[0069] 3) Select the recursive feature elimination (RFECV) method based on any machine learning model to screen features. Since the splitting of the training data set and the random selection of features are accidental, the program is iterated n times, and the features supported by the model n times are put into the feature candidate library. The execution steps are as follows:

[0070] Step 1: Set global parameters such as model, number of iterations, minimum number of features, number of cross-validations, feature elimination step, and scoring function. The number of cross-validations refers to the number of times the machine learning model is fitted and scored during each iteration; the feature elimination step refers to the number of features eliminated by the program each time; the scoring function refers to the evaluation index used for the final feature ranking, which has the degree of fit R. 2 And two methods: absolute error (MAE).

[0071] Step 2: loop iterates the number of times, initializes the recursive feature elimination object and reads in the features and labels, performs cross-validation fitting, and saves the results of each feature elimination to a temporary directory.

[0072] Step 3: Construct a feature set, traverse the feature elimination results, and add features with multiple iteration ranking values ​​Rank < 3 to the set.

[0073] 4) Calculate the feature correlation matrix and remove redundant features with Pearson linear correlation greater than 0.95. Prioritize features with strong correlation with attribute values. The attribute feature library is categorized into three types of features for subsequent input: fingerprint, topology calculation features, and high-throughput calculation features.

[0074] In the above technical solution, a further inventive concept is that: in step four, the features are combined and input into the machine learning model: the model calls the specific implementation in Scikit-learn, and the training process involves three models, namely support vector regression (SVR), ridge regression (KRR), and random forest. On the basis of constructing a feature library for a certain property, the fingerprint bit vector is spliced ​​with the high-throughput computing features and the topological structure computing features to form a combined vector and input it into the above machine learning model. The model first uses the grid parameter optimization method to obtain the optimal parameters, and finally uses the five-fold cross-validation method to evaluate the fitting effect of the QSPR model on the molecular properties. Certain features are processed by feature normalization to make the data more separable. The formula can be expressed as follows: X min is the minimum value of this feature in the generalized experimental sample data, X max The corresponding maximum value is:

[0075]

[0076] The model evaluation functions include mean error MAE and fit R 2 , the calculation formula is as follows:

[0077]

[0078] y and represent the true value and the predicted value respectively. Indicates the true value average, SS res Refers to the residual sum of squares, e i Pointer error, SS tot Refers to the sum of squares after removing the mean, R 2 The calculation results are in the range of [0,1], and the larger the value, the better the fitting effect.

[0079]

[0080]

[0081]

[0082] Furthermore, linear and nonlinear analysis of features can be performed to identify which features are associated with attribute values ​​from a chemical mechanism perspective, thereby mining chemical knowledge from big data. As shown in the feature table, the decomposition temperature is strongly correlated with the chemical principle corresponding to the feature. While nonlinear features have limited effectiveness in traditional models, they are closely related to the more precise, strongly correlated features of high-throughput computation. Therefore, the feature library extracted and constructed by the data mining engine can be input into a deep neural network-based correction model.

[0083] The last step is to use the attribute feature library, taking the data mining of energetic material decomposition temperature as an example. Figure 2 The following steps can be described in detail. The characteristics and property values ​​that are strongly correlated with the decomposition temperature are shown in Table 1:

[0084] Table 1

[0085]

[0086]

[0087] Table 2 shows the fitting results of different feature combinations on the random forest model. The data in the table shows that the best fitting effect is achieved by combining fingerprint and topological features (for non-labeled training sets, all are the results of five-fold cross-validation on the test set. FP stands for fingerprint, Topology refers to topological features, DFT refers to high-throughput computing features, and dashes _ represent feature combinations):

[0088] Table 2

[0089]

[0090] Example 1

[0091] A method for data mining of energetic compounds, comprising the following steps:

[0092] S1, data preprocessing: obtain energetic molecular structures and their corresponding physicochemical property values, and according to the screening conditions of components and structures, screen compounds containing only four elements of CHON in the components, and single-molecule electrically neutral compounds containing nitro groups in the structure, and the number of atoms in the molecules is less than the set range (80). The screening results are deduplicated and encoded and stored as a data set; the data set is then classified according to the type of nitro-bonded chemical environment, and a training set and a test set with uniform distribution of types are constructed;

[0093] S2, calculated features: extract the pre-processed molecular topological structure features and the features calculated by quantum chemistry respectively, generate fingerprint features, and establish multiple molecular feature libraries for different material properties. One feature library contains all the calculated features, and the construction of the feature library is completed; all the calculated features include the following types: molecular 2D structural features extracted using the Rdkit tool, molecular topological structure features related to the field of energetic materials defined in the present invention, and features calculated using quantum chemistry;

[0094] S3, feature screening: Pearson correlation and information entropy are calculated to screen out linearly correlated and uncorrelated features, and recursive feature elimination (RFECV) is used to reduce the feature dimension and delete redundant features whose linear correlation between features is higher than the set value.

[0095] S4, knowledge mining and model validation: Discover the correlation and knowledge association between feature libraries, verify the strong correlation between attributes and features from the perspective of expert knowledge, splice fingerprint features as input to the machine learning model, adjust parameters, and use cross-validation methods on different models to verify and evaluate features to obtain the optimal quantitative structure-property relationship (QSPR) model for data mining of energetic compounds.

[0096] Example 2

[0097] Based on Example 1, in step S1, the molecular structure is stored in a regular SMILES code, and whether the molecular formula contains elements other than CHON is determined by Python's character processing molecular formula, and the generated regular SMILES set is used for deduplication; whether the molecule contains a nitro group and the connection type of the nitro group is recorded, the situation where the molecular structure contains ionic compounds or hydrates is excluded, and the total number of atoms contained in the molecule is controlled within eighty; in addition, hydrates and ionic salts are implemented by string comparison, and the search and type of nitro groups are implemented by substructure search and molecular graph algorithm search; nitro groups are divided into seven categories according to the type of bonding chemical environment, including chain oxygen nitro groups, chain nitrogen nitro groups, chain carbon nitro groups, ordinary ring nitrogen nitro groups, ordinary ring carbon nitro groups, aromatic ring nitrogen nitro groups, and aromatic ring carbon nitro groups; the training set and the test set are constructed according to the set ratio to ensure that each type of molecule is divided according to the same ratio, and the training set and the test set are added in sequence to ensure the rationality of the data division. After the data preprocessing in step S1, the distribution of each type of data can be seen.

[0098] Example 3

[0099] On the basis of Example 1, in step S2, the fingerprint features are generated, and multiple molecular feature libraries are established for different material properties. One feature library contains all calculated features, including the following sub-steps:

[0100] Step (1): Using Rdkit software, calling the rdMolDescriptors module to calculate the 2D descriptor, the Morgan fingerprint method is used to calculate the fingerprint features, and used to describe the extended connectivity of the molecular structure;

[0101] Step (2): Design features to describe the topological structure of the energetic material; specifically, the type of nitro group that can be determined in step S1, the type of azole and its occurrence frequency, the minimum or maximum length of the bond of the nitro nitrogen atom in the conformation, the maximum number of adjacent occurrences of three identical atoms of C, N, and O in the molecular structure, the number of nitroamines, the ratio of aromatic atoms, the ratio of nitro atoms, the number of amino groups, the number of bridging oxygens, whether there is a condensed ring, and the ratio of atoms on the ring;

[0102] Step (3): extracting the features of the DFT calculation file, specifically comprising the following steps: step 1) constructing a Gaussian calculation file at a selected calculation level, calculating and generating a log file and a chk file, and further processing the chk file into a corresponding wfn file; step 2) intercepting relevant paragraphs and pseudo code for extracting feature information from the log file and the wfn file; and step 3) obtaining pseudo code for accurate conformational atomic spacing and nitro bond length from the log file. The features include molecular surface electrostatic potential analysis results, spin density, chemical bond distances between non-hydrogen atoms, HOMO atomic aromaticity index, stacking index, molecular shape and radius, band gap, the strongest and weakest bond lengths of the NO2-R nitro group, and nitro charge.

[0103] Example 4

[0104] Based on Example 1, in step S3, the screening feature covers linear and nonlinear screening methods, specifically including the following sub-steps:

[0105] Step 1): Use Pearson correlation coefficient to screen features;

[0106] Step 2): Use the nonlinear evaluation method Copula negative entropy to calculate the correlation of nonlinear features;

[0107] Step 3): Select any machine learning model and use the recursive feature elimination (RFECV) method to expand the candidate feature library. The procedure is iterated n times to solve the accidental effects of the splitting of the training dataset and the random selection process of features. n is a positive integer.

[0108] Step 4): Calculate the feature correlation matrix, remove redundant features whose Pearson linear correlation between features is greater than the set value, and give priority to retaining features with strong correlation with attribute values.

[0109] Example 5

[0110] Based on Example 1, in step S4, the different models include support vector regression (SVR), ridge regression (KRR), and random forest. After constructing a feature library for a certain property, the fingerprint bit vector is concatenated with high-throughput computational features and topological structure computational features to form a combined vector that is input into any of the three machine learning models. The model first uses grid parameter optimization to obtain the optimal parameters, and finally uses five-fold cross-validation to evaluate the fitting effect of the QSPR model on the molecular properties.

[0111] Example 6

[0112] Based on Example 1, in step S4, the method of discovering the correlation and knowledge association between feature libraries also includes the steps of: performing linear and nonlinear analysis on the features, analyzing the correlation between the attribute values ​​and those features from the perspective of chemical mechanism, thereby mining chemical knowledge in big data.

[0113] Example 7

[0114] Based on Example 3, in step (1), the 2D descriptor includes 200 descriptors characterizing TPSA, MolLogP, Charges, and hydrogen bond donor and acceptor properties.

[0115] Example 8

[0116] On the basis of Example 4, in step 3), the step is included: putting features supported by the model n times into a feature candidate library.

[0117] Example 9

[0118] On the basis of Example 5, step S4 includes a sub-step of normalizing the features to make the data more separable.

[0119] Example 10

[0120] A system suitable for data mining of energetic compounds, comprising a computer device; the computer device comprises a processor and a memory, wherein the memory stores a computer program, and when the computer program is loaded by the processor and executes the method described in any one of Examples 1 to 9.

[0121] The units involved in the embodiments of the present invention may be implemented in software or hardware, and the units described may also be provided in a processor. In some cases, the names of these units do not limit the units themselves.

[0122] According to one aspect of the present application, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described above.

[0123] As another aspect, the present application further provides a computer-readable medium, which may be included in the electronic device described in the above embodiments, or may exist independently without being incorporated into the electronic device. The computer-readable medium carries one or more programs, and when the one or more programs are executed by the electronic device, the electronic device implements the method described in the above embodiments.

[0124] The parts not involved in the present invention are the same as the existing technology or can be implemented by using the existing technology.

[0125] The above technical solution is only one embodiment of the present invention. For those skilled in the art, it is easy to make various types of improvements or modifications based on the application methods and principles disclosed in the present invention, and it is not limited to the method described in the above specific embodiment of the present invention. Therefore, the method described above is only preferred and does not have a restrictive meaning.

[0126] In addition to the above examples, those skilled in the art may obtain other embodiments based on the above disclosure or by utilizing knowledge or technology in related fields to make modifications. The features of each embodiment may be interchangeable or replaced. The modifications and changes made by those skilled in the art do not depart from the spirit and scope of the present invention and should be within the scope of protection of the claims attached to the present invention.

Claims

1. A method for data mining of energetic compounds, characterized in that: The following steps are involved: S1, data preprocessing: Obtain energetic molecular structures and their corresponding physical and chemical property values. Based on the component and structure screening conditions, screen compounds containing only four elements (CH, N, O), single-molecule electrically neutral compounds containing nitro groups, and the number of atoms in the molecule is less than the set range. De-duplicate the screening results, encode them, and store them as a data set. The dataset is then classified according to the type of nitro-bonded chemical environment, and training and test sets are constructed with evenly distributed types. S2, calculation features: extract the pre-processed molecular topological structure features and quantum chemical calculation features respectively, generate fingerprint features, and establish multiple molecular feature libraries for different material properties. One feature library contains all the calculated features, completing the construction of the feature library; S3, feature screening: Pearson correlation and information entropy are calculated to screen out linearly correlated and nonlinearly correlated features, and recursive feature elimination (RFECV) is used to reduce the feature dimension, removing redundant features whose linear correlation between features is higher than a set value. In step S3, the feature screening covers both linear and nonlinear screening methods, and specifically includes the following sub-steps: Step 1): Use Pearson correlation coefficient to screen features; Step 2): Use the nonlinear evaluation method Copula negative entropy to calculate the correlation of nonlinear features; Step 3): Select any machine learning model and use the recursive feature elimination (RFECV) method to expand the candidate feature library. The program is iterated n times to solve the accidental effects of the splitting of the training data set and the random selection process of features. n is a positive integer. Step 4): Calculate the feature correlation matrix, remove redundant features whose Pearson linear correlation between features is greater than the set value, and give priority to retaining features with strong correlation with attribute values; S4, knowledge mining and model validation: Discover the correlation and knowledge association between feature libraries, verify the strong correlation between attributes and features from the perspective of expert knowledge, splice fingerprint features as input to the machine learning model, adjust parameters, and use cross-validation methods on different models to verify and evaluate features to obtain the optimal quantitative structure-property relationship (QSPR) model for data mining of energetic compounds.

2. The method for energetic compound data mining according to claim 1, characterized in that: In step S1, in step S1, the molecular structure is stored in regular SMILES encoding, and the determination of whether the molecular formula contains elements other than CHON is implemented by Python's character processing molecular formula, and the generated regular SMILES set is used for deduplication; it is determined whether the molecule contains a nitro group and the connection type of the nitro group is recorded, and the situation where the molecular structure contains ionic compounds or hydrates is excluded, and the total number of atoms contained in the molecule is controlled within eighty; and hydrates and ionic salts are implemented by string comparison, and the search and type of nitro groups are implemented by substructure search and molecular graph algorithm search; nitro groups are divided into seven categories according to the type of bonding chemical environment, including chain oxygen nitro groups, chain nitrogen nitro groups, chain carbon nitro groups, ordinary ring nitrogen nitro groups, ordinary ring carbon nitro groups, aromatic ring nitrogen nitro groups, and aromatic ring carbon nitro groups; the training set and the test set are constructed according to the set ratio to ensure that each type of molecule is divided according to the same ratio, and then the training set and the test set are added in sequence to ensure the rationality of data division.

3. The method for energetic compound data mining according to claim 1, characterized in that: In step S2, the fingerprint features are generated, and multiple molecular feature libraries are established for different material properties. One feature library contains all the calculated features, including the following sub-steps: Step (1): Using Rdkit software, call the rdMolDescriptors module to calculate the 2D descriptor, use the Morgan fingerprint method to calculate the fingerprint features, and use it to describe the extended connectivity of the molecular structure; Step (2): Design features to describe the topological structure of energetic materials; Step (3): Extract the features of the DFT calculation file.

4. The method for energetic compound data mining according to claim 1, characterized in that: In step S4, the different models include support vector regression SVR, ridge regression KRR, and random forest.

5. The method for energetic compound data mining according to claim 1, characterized in that: In step S4, the process of discovering the correlation and knowledge association between feature libraries further includes the steps of performing linear and nonlinear analysis on the features, analyzing the correlation between the attribute values ​​and those features from the perspective of chemical mechanism, thereby mining chemical knowledge in big data.

6. The method for energetic compound data mining according to claim 3, characterized in that: In step (1), the 2D descriptors include descriptors characterizing TPSA, MolLogP, Charges, and hydrogen bond donor and acceptor properties.

7. The method for energetic compound data mining according to claim 1, characterized in that: In step 3), the steps include: putting the features supported by the model n times into the feature candidate library.

8. The method for energetic compound data mining according to claim 4, characterized in that: In step S4, the sub-step is included: normalizing the features to make the data more separable.

9. A system suitable for data mining of energetic compounds, characterized in that: The method comprises a computer device, wherein the computer device comprises a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is loaded by the processor and executes the method according to any one of claims 1 to 8.

Images