A compound performance prediction method and system based on adaptive classification

Abstract

This invention provides a method and system for predicting compound performance based on adaptive classification. The method first collects compounds with clearly defined performance and molecular structure data from literature and databases to form a dataset, and generates a Morgan fingerprint for each compound. Then, it calculates the structural similarity among all compounds and selects highly similar training samples to construct a personalized training set. Next, it evaluates feature importance using a random forest regressor, selects important features as input variables, and trains a regression model to predict the performance of the target compound. This invention significantly improves training efficiency and prediction accuracy, and is applicable to fields such as drug development and materials science.

Description

Technical Field

[0001] This invention presents a method and system for predicting compound performance based on molecular fingerprint similarity calculation combined with an adaptive classification mechanism. This method can be widely applied in scenarios requiring efficient prediction of compound performance, such as drug development, materials science, and chemistry. Background Technology

[0002] In recent years, with the rapid development of computer science, artificial intelligence, and cheminformatics, data-driven machine learning methods have gradually become an important means of predicting compound performance. Among them, molecular fingerprinting, as a technique that transforms molecular structures into computable binary or numerical vectors, is widely used in molecular similarity analysis, virtual screening, and model building due to its intuitiveness, efficiency, and ease of processing. Although existing studies have attempted to introduce molecular similarity as a basis for compound performance prediction, a systematic methodological framework is still lacking, which affects the accuracy of predictions. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a compound performance prediction method based on adaptive classification.

[0004] Therefore, the compound performance prediction method based on adaptive classification proposed in this invention is used to predict the target performance of a target compound. The method includes the following steps:

[0005] Step 1: Construct an initial dataset, which includes more than 500 known compounds and their target performance values; and the target performance values ​​of each compound are measured using the same target performance detection method.

[0006] Step 2: Calculate the structural similarity S between the target compound and each compound in the initial dataset using Equation (1); select compounds in the initial dataset whose structural similarity is greater than a threshold to form the training dataset; the threshold is the maximum similarity multiplied by 0.9;

[0007]

[0008] In formula (1):

[0009] A is the Morgan fingerprint vector of the target compound;

[0010] B is the Morgan fingerprint vector of any compound in the initial dataset;

[0011] A*B represents the dot product of vectors A and B;

[0012] ||A|| 2 This represents the number of 1s in Morgan's fingerprint vector A, which is the number of active bits in the molecular fingerprint.

[0013] ||B|| 2 This represents the number of 1s in Morgan's fingerprint vector B, which is the number of active bits in the molecular fingerprint.

[0014] Step 3: Use the RDKit program to calculate the feature value set of each compound in the training dataset; then use the feature value set of each compound as input to evaluate the contribution of each feature using a random forest regressor, and select 10-20 features with high contribution as effective features.

[0015] Step 4: Establish a regression model using the feature values ​​of the effective features of each compound in the training dataset as independent variables and the target performance of each compound in the training dataset as dependent variables.

[0016] Step 5: Use the RDKit program to calculate the feature values ​​of the effective features of the target compound. Use the feature values ​​of the effective features of the target compound as the input of the regression model constructed in Step 4, and output the target performance value of the target compound.

[0017] An alternative approach is to use the GetMorganFingerprint module in the RDKit software to generate the Morgan fingerprint vector of the compound.

[0018] Alternatively, both the target compound and the known compounds can be organic compounds.

[0019] Alternatively, both the target compound and the known compounds may be energetic compounds.

[0020] The present invention also provides a compound performance prediction system based on adaptive classification, the system comprising multiple initial databases, an adaptive training dataset screening module, an effective feature screening module, and a regression model construction and prediction module;

[0021] Each initial database includes more than 500 known compounds and their target properties; the target properties differ among the initial databases.

[0022] The adaptive training dataset filtering module uses the method described in step 2 of claim 1 to filter training datasets from the database corresponding to the performance to be predicted.

[0023] The effective feature filtering module uses the training dataset filtered by the adaptive training dataset filtering module to filter effective features using the method described in step 3 of claim 1.

[0024] The regression model building and prediction module uses the predicted performance of the target compound described in steps 4 and 5 of claim 1, based on the selected training dataset and effective features.

[0025] This invention utilizes molecular fingerprinting to rapidly calculate similarity, improving training efficiency. The training set, selected based on structural similarity, more closely resembles the target compound, enhancing prediction accuracy. Simultaneously, the system possesses adaptive classification capabilities, automating the entire process from data selection to modeling and prediction. Each target compound receives its own dedicated training set and model, enhancing the specificity of predictions and making it suitable for performance prediction tasks involving multiple types of compounds. It is particularly applicable to fields such as drug development and materials science. Attached Figure Description

[0026] Figure 1 The TNT molecular structure in Example 1 and the structure of the compound with the highest structural similarity to it are shown below.

[0027] Figure 2 The structures shown are the TATB molecular structure in Example 2 and the structure of the compound with the highest structural similarity to it. Detailed Implementation

[0028] Unless otherwise specified, the scientific and technical terms used in this article are intended for understanding by those skilled in the art.

[0029] This invention designs a training set extraction method based on molecular fingerprint similarity calculation. According to the molecular structural characteristics of the target compound, it automatically identifies training samples that are highly similar to it and constructs a dedicated training set accordingly, forming a highly targeted subset of data. This allows for the training of a more representative and predictive model, thereby reducing the prediction error of the target performance.

[0030] The predictable properties of the method of the present invention include, but are not limited to, crystal density, thermal decomposition temperature, molecular volume, dielectric constant, and transition energy.

[0031] The initial dataset of this invention is constructed to predict the target performance. Compounds with clearly defined performance values ​​and molecular structure data are collected from literature and databases to form the dataset, and the number of compounds must exceed 500. There are no special requirements for the types of compounds applicable to this invention; however, the performance to be predicted must be obtained using standardized calculation or testing methods.

[0032] As is well known in the art, the feature set of a specific compound is calculated using the MolecularDescriptorCalculator module in RDKit, which includes 218 feature values ​​such as 'BCUT2D_MWHI', 'BCUT2D_MRLOW', 'MaxPartialCharge', 'SMR_VSA4', 'VSA_EState4', and 'PEOE_VSA7'.

[0033] The following are specific embodiments of the present invention. It should be noted that the following embodiments are preferred examples of the present invention. Those skilled in the art can optimize technical factors such as compound type, feature type, descriptors in special data of similarity calculation method, number of cracking energies, fitting method, etc. within the scope of the present invention to obtain other implementation schemes that fall within the protection scope of the present invention.

[0034] Example 1:

[0035] This embodiment describes the construction of a prediction model using the method of the present invention, with the target compound being TNT molecules. Figure 1 The structure shown on the left is predicted to be the crystal density of the compound TNT.

[0036] Step 1: The initial dataset used in this embodiment is the crystal density of the target performance to be predicted. Compounds with clear crystal density data (crystal density measured by XRD experiments) and molecular structures are collected from literature and databases. The number of compounds collected is 1524.

[0037] Step 2: Based on the chemical information in the dataset, the Morgan fingerprint of each compound is generated using the GetMorganFingerprint module in the RDKit software. The structural similarity S between the target compound and each compound in the initial dataset is calculated using Equation (1). Compounds with structural similarity greater than the threshold in the initial dataset are selected to form the training dataset.

[0038] In this embodiment, the maximum similarity is 0.877, and the calculation process is as follows:

[0039]

[0040] The number of active bits in TNT's fingerprint is 60, which is the highest similarity to its molecule ( Figure 1 The number of activated bits in the structure shown on the right is 62, and the dot product (intersection) of the two is 55; that is, A*B = 57, ||A|| 2 =60, ||B|| 2 =62;

[0041] Using 0.877 multiplied by 90% to equal 0.789 as the threshold in this embodiment, compounds with a similarity higher than 0.789 are selected from the initial dataset to form the training dataset;

[0042] Step 3: Use the MolecularDescriptorCalculator module in the RDKit software to calculate the feature value set of each compound in the training dataset; then use the feature value set of each compound as input to evaluate the contribution of each feature using a random forest regressor, and select the 20 features with high contribution as effective features.

[0043] The 20 valid features selected in this embodiment are: ['BCUTd-1l','ETA_epsilon_4','ETA_psi_1','AMID_h','ETA_epsilon_5','AXp-1d','RingCount','SMR_VSA3','VSA_EState1','SMR_VSA6','PEOE_VSA8','PEOE_VSA3','SlogP_VSA6','MinEStateIndex','EState_VSA9','BCUT2D_LOGPLOW','MaxPartialCharge','SMR_VSA4','VSA_EState4','PEOE_VSA7'];

[0044] Step 4: Establish a regression model using the feature values ​​of the effective features of each compound in the training dataset as independent variables and the target performance of each compound in the training dataset as dependent variables.

[0045] Step 5: Use the RDKit program to calculate the eigenvalues ​​of the effective features of TNT ('BCUTd-1l','ETA_epsilon_4','ETA_psi_1','AMID_h','ETA_epsilon_5','AXp-1d','RingCount','SMR_VSA3','VSA_EState1','SMR_VSA6','PEOE_VSA8','PEOE_VSA3','SlogP_VSA6','MinEStateIndex','EState_VSA9','BCUT2D_LOGPLOW','MaxPartialCharge','SMR_VSA4','VSA_EState'). 4','PEOE_VSA7. The eigenvalues ​​are 10.53175926, 0.207777778, 0.570198697, 227.132, 222.092, 227.0178349, 84, 0, 0.285563787, -0.258302312, 0.285563787, 0.258302312. The values ​​of TNT's effective characteristics (0.875, 1.25, 1.5, 16.63840264, 10.12515738, 2.265871092, 0.357142857, 1079.291701, 0.50127803) are used as the input to the regression model constructed in step 4. The output value of TNT's crystal density is 1.672 g / cm³. 3 The known density of TNT crystals is 1.654 g / cm³.3 (Measured by XRD experiments) close.

[0046] Example 2:

[0047] This embodiment uses the method of the present invention to predict that the target compound is a TATB molecule. Figure 2 (The structure shown on the left) thermal decomposition temperature.

[0048] Step 1, the initial dataset construction is to collect compounds with clear thermal decomposition temperature data (measured by DSC method) and molecular structure from literature and databases to form a dataset, which consists of 1302 compounds;

[0049] Step 2, the maximum similarity is 0.744. The calculation process is as follows: the number of activated bits in the TATB fingerprint is 68, and it has the highest similarity with the molecule ( Figure 2 In the structure shown on the right, the number of activated bits is 75, and the dot product (intersection) of the two is 61.

[0050]

[0051] Using 0.744 multiplied by 90% to obtain a result of 0.669 as a threshold, compounds with a similarity higher than 0.669 were selected from the initial dataset to form the training set;

[0052] Step 3: Use the MolecularDescriptorCalculator module in the RDKit software to calculate the feature value set of each compound in the training dataset; then use the feature value set of each compound as input to evaluate the contribution of each feature using a random forest regressor, and select the 20 features with high contribution as effective features.

[0053] The 20 valid features selected in this embodiment are: ['BCUT2D_MWHI','C_contents','BCUT2D_MRLOW','Oxygen_Balance','BCUT2D_LOGPLOW','MinEStateIndex','fr_ArN','MaxPartialCharge','SMR_VSA4','fr_azide','SMR_VSA10','BCUT2D_CHGHI','O_contents','MaxAbsPartialCharge','MinPartialCharge','PEOE_VSA7','fr_NH2','MolLogP','BCUT2D_LOGPHI','SlogP_VSA10'];

[0054] Step 4: Establish a regression model using the feature values ​​of the effective features of each compound in the training dataset as independent variables and the target performance of each compound in the training dataset as dependent variables.

[0055] Step 5: Use the RDKit program to calculate the eigenvalues ​​of the valid TATB features, which are 'BCUT2D_MWHI', 'C_contents', 'BCUT2D_MRLOW', 'Oxygen_Balance', 'BCUT2D_LOGPLOW', 'MinEStateIndex', 'fr_ArN', 'MaxPartialCharge', 'SMR_VSA4', 'fr_azide', 'SMR_VSA10', 'BCUT2D_CHGHI', 'O_contents', 'MaxAbsPartialCharge', 'MinPartialCharge', 'PEOE_VSA7', 'fr_NH2', 'MolLogP', 'BCUT2D_LOGPHI', 'SlogP_VSA10'. The eigenvalues ​​are: 16.63915787, 10.11362724, 2.382551081, -2.056236883, 2.4769002, -2.103259107, 5.964916594, -0.398700305, 3.938405305, 480.3939375, 14.19615242, 8.523181337, 8.523181337, 8.196152423, 4.132290912, 4.132290912, 3.058175152, 3.058175152, 2.078443791, 2.078443791. Using the eigenvalues ​​of the effective features of TATB as input to the regression model constructed in step 4, the predicted output thermal decomposition temperature is 331℃, which is close to the known thermal decomposition temperature of 320℃ (measured by DSC method).

Claims

1. A method for predicting a target property of a target compound based on adaptive classification, the method for predicting a target property of a target compound, characterized in that, The method includes the following steps: Step 1: Construct an initial dataset, which includes more than 500 known compounds and their target performance values; and the target performance values ​​of each compound are measured using the same target performance detection method. Step 2: Calculate the structural similarity S between the target compound and each compound in the initial dataset using Equation (1); The training dataset is composed of compounds in the initial dataset whose structural similarity is greater than a threshold; the threshold is the maximum similarity multiplied by 0.

9. In formula (1): A is the Morgan fingerprint vector of the target compound; B is the Morgan fingerprint vector of any compound in the initial dataset; A*B represents the dot product of vectors A and B; ||A|| 2 This represents the number of 1s in Morgan's fingerprint vector A, which is the number of active bits in the molecular fingerprint. ||B|| 2 represents the number of bits in the Morgan fingerprint vector B that are 1, i.e. the number of bits activated in the molecular fingerprint; Step 3: Use the RDKit program to calculate the feature value set of each compound in the training dataset; then use the feature value set of each compound as input to evaluate the contribution of each feature using a random forest regressor, and select 10-20 features with high contribution as effective features. Step 4: Establish a regression model using the feature values ​​of the effective features of each compound in the training dataset as independent variables and the target performance of each compound in the training dataset as dependent variables. Step 5: Use the RDKit program to calculate the feature values ​​of the effective features of the target compound. Use the feature values ​​of the effective features of the target compound as the input of the regression model constructed in Step 4, and output the target performance value of the target compound.

2. The adaptive classification based compound performance prediction method of claim 1, wherein, The Morgan fingerprint vector of the compound was generated using the GetMorganFingerprint module in the RDKit software.

3. The adaptive classification based compound performance prediction method of claim 1, wherein, Both the target compound and the known compounds are organic compounds.

4. The adaptive classification based compound performance prediction method of claim 1, wherein, Both the target compound and the known compounds are energetic compounds.

5. A compound performance prediction system based on adaptive classification, characterized in that, It includes multiple initial databases, an adaptive training dataset selection module, an effective feature selection module, and a regression model building and prediction module; Each initial database includes more than 500 known compounds and their target properties; the target properties differ among the initial databases. The adaptive training dataset filtering module uses the method described in step 2 of claim 1 to filter training datasets from the database corresponding to the performance to be predicted. The effective feature filtering module uses the training dataset filtered by the adaptive training dataset filtering module to filter effective features using the method described in step 3 of claim 1. The regression model building and prediction module uses the predicted performance of the target compound described in steps 4 and 5 of claim 1, based on the selected training dataset and effective features.

Images