Compound physical property prediction method, system, equipment and medium
Through automated feature extraction and multimodal information integration, combining clustering and genetic algorithms to optimize feature combinations, the inaccuracy problem of traditional prediction models in small sample sets and strong polar compounds is solved, and the accuracy of compound physical properties prediction is significantly improved. and model generalization capabilities.
Patent Information
- Application Number
- CN202411408132.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-10
- Publication Date
- 2024-12-31
AI Technical Summary
Traditional material attribute prediction models perform poorly when dealing with small sample sets and strong polar compounds, especially when predicting compounds with high dielectric constants, resulting in inaccurate prediction results and prone to overfitting problems.
Through automated feature extraction and integration of multimodal information, the clustering module is used to obtain the correlation distance of compound molecular features, perform clustering grouping and linear fitting, and optimize feature combinations using genetic algorithms, combining linear and nonlinear coding models , find the optimal feature combination to improve prediction accuracy.
The correlation between the prediction results and the target variables is significantly improved, and the generalization ability and prediction accuracy of the model is enhanced, especially in the prediction of small sample sets and high dielectric constant compounds, which outperforms traditional methods.
Smart Images
Figure CN119230009A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data-driven technology, and in particular to a method, system, device and medium for predicting physical properties of a compound. Background Art
[0002] Traditional material property prediction models mainly rely on empirical formulas and physical theories, such as the Clausius-Mosotti equation, the Onsager and Kirkwood equations, etc. Although these theories can provide useful predictions in some cases, they are usually only applicable to specific groups of substances, such as rare gases or liquids with limited polarity, and the reliability of these theoretical models is significantly reduced when dealing with highly polar compounds. Later, researchers used multiple physical descriptors of molecules to predict the properties of objects by constructing quantitative structure-property relationship (QSPR) models. For example, the QSPR model constructed by Liu and Rowley predicts the dielectric constant of compounds based on descriptors such as molecular dipole-dipole moment, solubility parameters, van der Waals area and refractive index.
[0003] With the development of computational chemistry and machine learning technology, researchers have begun to try machine learning (ML) methods, applying traditional machine learning algorithms (such as linear regression, support vector machines, decision trees, random forests, etc.) to learn from molecular features and predict the properties of substances, such as dielectric constant, solubility, conductivity, etc. However, these individual machine learning methods have limitations in feature selection and model training, especially when dealing with high-dimensional nonlinear data relationships. In addition, ML methods are prone to overfitting problems during feature combination and model training.
[0004] Traditional material property prediction models predict the dielectric constant value through the physical properties of molecules, but they are usually based on limited molecular descriptors and perform poorly when dealing with compounds with dielectric constant values greater than 50. They are not only time-consuming and labor-intensive, but may also affect the performance of the model by ignoring other important features, resulting in inaccurate prediction results of the compound's physical properties. Summary of the invention
[0005] In view of the limitations of the prior art in processing small sample sets and highly polar compounds, especially the poor performance in predicting compounds with high dielectric constants, resulting in inaccurate prediction results of compound properties, the present invention proposes a compound property prediction method, system, equipment and medium, which realizes multi-scale description from micro to macro through automated feature extraction and integration of multimodal information, improves the correlation between the prediction results and the target variables, and thus solves the problems existing in the prior art.
[0006] A method for predicting physical properties of a compound comprises the following steps:
[0007] Obtain molecular characteristics of compounds;
[0008] The predicted value of the physical property of the compound is obtained by inputting the molecular characteristics of the compound into the physical property prediction model; the method specifically includes the following steps:
[0009] The clustering module is used to obtain the correlation distance between the molecular features of the compounds, the molecular features are clustered and grouped according to the correlation distance, and the molecular features in the same group are linearly fitted;
[0010] The molecular features after linear fitting are divided into multiple subspaces using the feature subspace combination module. Feature combinations are performed in each subspace, and global search and directional optimization are performed on the feature combinations to find the optimal combination of molecular features.
[0011] The coding module is used to perform linear and nonlinear changes on all molecular features of the optimal combination to find features suitable for nonlinear fitting and linear fitting; the predicted value of the physical property of the compound is obtained based on the features suitable for nonlinear fitting and linear fitting.
[0012] Furthermore, the molecular features of the compound are obtained by using the RDKit library and the Mordred tool, wherein the molecular features include Mordred descriptors, atomic level descriptors, molecular level descriptors, Rdkit ML descriptors, CATS2D descriptors, and MOPAC descriptors.
[0013] Furthermore, before obtaining the correlation distance between the molecular features of the compound through the clustering module, the feature processing module is used to transform and reduce the dimension of the molecular features. The transformation process transforms the molecular features in sequence through Fourier transform and cosine transform, and then uses sparse transform to reduce the dimension of the transformed features. The reduced-dimensional features are embedded and encoded based on the Transformer model; at the same time, special transformations are performed on the unstructured data features in the molecular features, and all discrete feature data are converted into unified independent coded variables.
[0014] Furthermore, the correlation distance between the molecular features of the compound is obtained through the clustering module, and the features are clustered and grouped according to the correlation distance; specifically, the Spearman or Pearson correlation coefficient is used to calculate the correlation distance between the molecular features, and according to the correlation distance, the features are clustered and grouped using DBSCAN or the nearest neighbor clustering algorithm.
[0015] Furthermore, a genetic algorithm is used to perform global search and directional optimization on the feature combination, thereby finding the optimal combination of molecular features; specifically, the following steps are included:
[0016] Set the genetic algorithm parameters, including feature dimension, population size, crossover rate, mutation rate, number of iterations, and target relevance threshold;
[0017] The feature combination in each subspace is iteratively calculated. When the specified number of iterations is reached and the correlation distance with the target is greater than the specified threshold, the iteration is stopped and the optimal combination of features is output.
[0018] The present invention also includes a compound property prediction system, comprising:
[0019] An acquisition module, used to obtain the molecular characteristics of the compound;
[0020] The prediction module is used to obtain the predicted value of the physical property of the compound by inputting the molecular characteristics of the compound into the physical property prediction model; it specifically includes:
[0021] A fitting unit, used for obtaining the correlation distance between the molecular features of the compound through a clustering module, clustering and grouping the molecular features according to the correlation distance, and performing linear fitting on the molecular features of the same group;
[0022] The optimal combination search unit is used to divide the molecular features after linear fitting into multiple subspaces using the feature subspace combination module, perform feature combination in each subspace, and perform global search and directional optimization on the feature combination, so as to find the optimal combination between the molecular features;
[0023] The prediction unit is used to use the coding module to perform linear and nonlinear changes on all the features of the optimal combination to obtain features suitable for nonlinear fitting and linear fitting; and obtain the predicted value of the physical property of the compound according to the features suitable for nonlinear fitting and linear fitting.
[0024] The present invention also includes a computer device for predicting compound properties, including: a memory, a processor, and a computer program stored in the memory, and the processor implements the steps of the compound property prediction method when executing the computer program.
[0025] The present invention also includes a readable storage medium, wherein the readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, they are used to execute the steps of the compound property prediction method.
[0026] The present invention provides a compound property prediction method, system, device and medium, which have the following beneficial effects:
[0027] The present invention constructs a physical property prediction model, combines the correlation distance and the genetic algorithm, linearly fits the same group of features, divides the fitted features and combines the features again, and merges the linearly related features by fitting the same group of features, reduces feature redundancy and improves the generalization ability of the model; then searches for the feature combination that is most relevant to the predicted target variable, and through this multi-level iterative method, it is possible to more effectively extract macroscopic properties from microscopic structural information, thereby improving the accuracy and interpretability of the prediction; the method realizes multi-scale description from micro to macro through automated feature extraction and integration of multimodal information, significantly improves the correlation between the prediction results and the target variables compared to traditional methods, improves the correlation between the prediction results and the target variables, and solves the problem of inaccurate prediction of the physical properties of compounds. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 Schematic diagram of the process of predicting the physical properties of compounds in the embodiments of the present invention. DETAILED DESCRIPTION
[0029] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.
[0030] The present invention proposes a method for predicting the physical properties of a compound, such as Figure 1 As shown, it includes a clustering module based on correlation distance, a feature subspace combination module based on genetic algorithm optimization, and a module suitable for linear and nonlinear coding. The density clustering module based on correlation distance uses a feature clustering algorithm based on correlation distance to group features and reduce linear redundant variables for linear fitting; the feature subspace combination module based on genetic algorithm optimization uses genetic algorithm to perform global search and directional feature combination optimization to find the optimal combination between features and evaluate; the module suitable for linear and nonlinear coding is used to extract original features and physical and chemical descriptors from the molecular structure to find features suitable for nonlinear fitting and linear fitting; the method specifically includes the following steps:
[0031] S1. Data acquisition; obtain the experimentally determined physical property values of relevant compounds and generate the SMILES and PDB formats of molecules using the CAS registration number and compound name. Then use the RDKIT library for electrical processing to ensure that all molecules are electrically neutral; then use the MOPAC program for quantum chemical optimization and property calculation; use the RDKit library and Mordred tools to perform molecular and atomic-level feature calculations and molecular fingerprint calculations and physical property values of relevant compounds, mainly including Mordred descriptors, atomic-level descriptors, molecular-level descriptors, Rdkit ML descriptors, CATS2D descriptors, MOPAC descriptors, Coulomb matrices, etc.; various properties directly queried from Pubchem, such as polarizable surface, logP, and various fingerprint technologies, such as MACCS fingerprint, Morgan fingerprint, Avalon fingerprint, topological fingerprint, etc.
[0032] S2. Feature processing, including feature combination, feature transformation, independent continuous feature processing, and feature normalization. First, high-dimensional features are transformed and reduced, including Fourier transform and cosine transform of molecular fingerprints, Coulomb matrices, and polarity distribution features, and then further compressed by sparse transformation, and embedded coding is performed based on the Transformer model. Secondly, special transformations are performed on unstructured data, including sequence data, two-dimensional matrices, and multi-dimensional payload data. All discrete features are converted into unified independent coding variables to achieve linear and nonlinear feature combination of subsequent modules. Finally, all the obtained features are normalized to ensure that the data is on the same scale and thus comparable; finally, all the features of this module (including all processed and unprocessed features) are used as input to the next module.
[0033] S3. Clustering based on correlation distance: After the feature engineering module is processed, the correlation distance between all features is calculated, and feature clustering is performed to reduce redundancy and improve the generalization ability of the model. This includes calculating feature correlation distances and using clustering algorithms to cluster and group features. Specifically, the Spearman or Pea rson correlation coefficient is used to calculate the correlation distance between features, DBSCAN or the nearest neighbor clustering algorithm is used to cluster and group features, and a linear model is used to fit the same group of features, thereby merging linearly related features and reducing feature redundancy. The specific processing process is as follows Figure 1 .
[0034] S4. Feature subspace combination based on genetic algorithm optimization: After clustering and linear fitting, the remaining features are combined using genetic algorithms to derive second-order nonlinear features with stronger correlation with the prediction target, that is, to select the most informative feature subset from the original features. The execution process of the genetic algorithm starts from iteration 0, iteratively optimizes the feature combination, and uses nonlinear models to transform the features. If the new features generated have stronger correlation, the probability of random selection of these features is increased, so that they are retained in the directed optimization. Otherwise, this feature combination is discarded, and iterative calculations are performed in each layer, combining local greedy strategies with global directed optimization. Finally, whether to continue iterating or stop is determined based on whether the specified number of iterations of 20 is reached and whether the distance of correlation with the target is greater than the specified threshold.
[0035] Use genetic algorithm to search feature space and find the feature combination that is most relevant to the target variable. Including: feature subspace combination, genetic algorithm parameter setting, nonlinear transformation, and optimization process. Divide the obtained features into multiple subspaces and combine them within these subspaces. Set the parameters of the genetic algorithm, such as DNA_SIZE (feature dimension), POP_SIZE (population size), CROSSOVER_RATE (crossover rate), MUTATION_RATE (mutation rate), N_GENERATIONS (number of iterations), and final_target (target correlation threshold). Once the feature combination is evaluated, nonlinear models such as KernelRidge, BaggingRegressor, ExtraTreesRegressor, etc. will be applied to transform the feature subspace and generate new variables. Finally, through the iterative process of the genetic algorithm, the new variables with stronger correlation with the target value will be retained. After this module, the remaining features are used as the input of the next module.
[0036] S5, global linear and nonlinear encoding; perform linear and nonlinear transformations on all features to extract deeper feature combinations. Including linear encoding, nonlinear encoding, model training, performance evaluation and optimization. For all feature combinations, use linear models (such as LinearRegression, RidgeCV, etc.) to perform linear transformations on all features, and use nonlinear models (such as ExtraTreesRegressor, XGBRegres sor, etc.) to perform nonlinear transformations on all features. Encode the results of linear and nonlinear transformations into a new feature set. Use evaluation indicators (such as R2 scores) to evaluate model performance and compare it with the correlation with the target variable. After this module, the remaining features are used as input to the next module.
[0037] S6. Finally, the model is optimized, combining the local greedy strategy and the global directed optimization, and the cascade iterative calculation is performed again. If the correlation between the newly generated features and the target value exceeds the preset threshold (for example, the correlation coefficient reaches 0.999), these new features will be retained. Through this multi-level iterative method, the system can more effectively extract macroscopic properties from microstructural information and improve the accuracy and interpretability of predictions. Ensure the effectiveness and efficiency of the model training process to respond to changes in system dynamics and disturbances in the external environment.
[0038] The present invention significantly improves the efficiency and accuracy of physical property prediction on a small sample set. Correlation distance and genetic algorithm are used to optimize feature selection, avoiding the problem of overfitting, and by simulating the layer-by-layer extraction of abstract features in deep learning, a rapid search for the optimal combination of features is achieved. Compared with the previous traditional machine learning model, the prediction accuracy of the model of the present invention is improved from 0.877 to 0.956 when predicting the target dielectric constant ε value, and the prediction result is significantly better than the traditional method when the predicted dielectric constant is greater than 50.
[0039] Based on the same inventive concept, the present invention also proposes a compound property prediction system, comprising:
[0040] The acquisition module is used to obtain the molecular characteristics of the compound.
[0041] The prediction module is used to obtain the predicted value of the physical property of the compound by inputting the molecular characteristics of the compound into the physical property prediction model; it specifically includes:
[0042] The fitting unit is used to obtain the correlation distance between the molecular features of the compound through the clustering module, cluster and group the molecular features according to the correlation distance, and perform linear fitting on the molecular features of the same group.
[0043] The optimal combination search unit is used to divide the molecular features after linear fitting into multiple subspaces using the feature subspace combination module, perform feature combination in each subspace, and perform global search and directional optimization on the feature combination, so as to find the optimal combination between molecular features.
[0044] The prediction unit is used to use the coding module to perform linear and nonlinear changes on all the features of the optimal combination to obtain features suitable for nonlinear fitting and linear fitting; and obtain the predicted value of the physical property of the compound according to the features suitable for nonlinear fitting and linear fitting.
[0045] The present invention also provides a computer device for predicting compound properties, comprising: a memory, a processor, and a computer program stored in the memory, and the processor implements the steps of the compound property prediction method when executing the computer program.
[0046] The present invention also provides a readable storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the compound property prediction method.
[0047] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.
Claims
1. A method for predicting physical properties of a compound, characterized in that: The following steps are involved: Obtain molecular characteristics of compounds; By inputting the molecular characteristics of the compound into the physical property prediction model, the predicted value of the physical property of the compound is obtained; It specifically includes the following steps: The clustering module is used to obtain the correlation distance between the molecular features of the compounds, the molecular features are clustered and grouped according to the correlation distance, and the molecular features in the same group are linearly fitted; The molecular features after linear fitting are divided into multiple subspaces using the feature subspace combination module. Feature combinations are performed in each subspace, and global search and directional optimization are performed on the feature combinations to find the optimal combination of molecular features. The coding module is used to perform linear and nonlinear changes on all molecular features of the optimal combination to find features suitable for nonlinear fitting and linear fitting; the predicted value of the physical property of the compound is obtained based on the features suitable for nonlinear fitting and linear fitting.
2. A method for predicting physical properties of a compound according to claim 1, characterized in that: The molecular features of the compound are obtained by using the RDKit library and the Mordred tool, wherein the molecular features include Mordred descriptors, atomic level descriptors, molecular level descriptors, RdkitML descriptors, CATS2D descriptors, and MOPAC descriptors.
3. A method for predicting physical properties of a compound according to claim 1, characterized in that: Before obtaining the correlation distance between the molecular features of the compound through the clustering module, the molecular features are transformed and reduced in dimension by the feature processing module. The transformation process transforms the molecular features in sequence through Fourier transform and cosine transform, and then uses sparse transform to reduce the dimension of the transformed features. The reduced-dimensional features are embedded and encoded based on the Transformer model; at the same time, the unstructured data features in the molecular features are specially transformed to convert all discrete feature data into unified independent coding variables.
4. A method for predicting physical properties of a compound according to claim 1, characterized in that: The clustering module is used to obtain the correlation distance between the molecular features of the compound, and the features are clustered and grouped according to the correlation distance; specifically, the Spearman or Pearson correlation coefficient is used to calculate the correlation distance between the molecular features, and according to the correlation distance, the DBSCAN or nearest neighbor clustering algorithm is used to cluster and group the features.
5. A method for predicting physical properties of a compound according to claim 1, characterized in that: Genetic algorithms are used to perform global search and directional optimization of feature combinations to find the optimal combination of molecular features; The specific steps include: Set the genetic algorithm parameters, including feature dimension, population size, crossover rate, mutation rate, number of iterations, and target relevance threshold; The feature combination in each subspace is iteratively calculated. When the specified number of iterations is reached and the correlation distance with the target is greater than the specified threshold, the iteration is stopped and the optimal combination of features is output.
6. A compound property prediction system, characterized in that: include: An acquisition module, used to obtain the molecular characteristics of the compound; A prediction module, used to obtain a predicted value of the physical property of the compound by inputting the molecular characteristics of the compound into the physical property prediction model; Specifically include: A fitting unit, used for obtaining the correlation distance between the molecular features of the compound through a clustering module, clustering and grouping the molecular features according to the correlation distance, and performing linear fitting on the molecular features of the same group; The optimal combination search unit is used to divide the molecular features after linear fitting into multiple subspaces using the feature subspace combination module, perform feature combination in each subspace, and perform global search and directional optimization on the feature combination, so as to find the optimal combination between the molecular features; The prediction unit is used to use the coding module to perform linear and nonlinear changes on all the features of the optimal combination to obtain features suitable for nonlinear fitting and linear fitting; and obtain the predicted value of the physical property of the compound according to the features suitable for nonlinear fitting and linear fitting.
7. A computer device for predicting physical properties of a compound, characterized in that: include: A memory, a processor, and a computer program stored in the memory, wherein when the processor executes the computer program, the steps of the method for predicting the physical properties of a compound according to any one of claims 1 to 5 are implemented.
8. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the method for predicting the physical properties of compounds according to any one of claims 1 to 5.