Molecular prediction method and system for drug compounds that inhibit biological activity of target proteins

Abstract

The present invention belongs to the field of data processing technology, and provides a method and system for predicting drug compound molecules that inhibit the biological activity of a target protein. The method includes obtaining molecular descriptors of candidate anticancer drugs; screening a set number of key molecular descriptors from the molecular descriptors of candidate anticancer drugs according to the correlation with the target biological activity; encoding the key molecular descriptors and converting them into corresponding vector representations; based on the pre-constructed relationship model between the key molecular descriptor features and the target biological activity and the key molecular descriptor vector representation, obtaining the candidate drug anticancer molecule that maximally antagonizes the target biological activity.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular relates to a method and system for predicting drug compound molecules that inhibit the biological activity of a target protein. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] In drug development, potential drug compounds are screened by establishing an effective model for inhibiting the activity of target proteins, and new compound molecules with better inhibitory target biological activity are predicted. For example, estrogen receptor ERα is one of the important targets for the treatment of breast cancer, and compounds that antagonize ERα activity may be candidate drugs for the treatment of breast cancer. However, due to the high data dimension of compounds and their molecular descriptors, limited data volume, and similarity of compound structures, it is very challenging to establish an accurate prediction model. Therefore, it is of great practical significance to the research and development of new anti-breast cancer drugs to be committed to the screening of anti-cancer molecules in breast cancer drugs or the construction of new highly active inhibitor compounds.

[0004] In terms of anti-cancer drug candidates, the traditional feature representation of anti-cancer drug candidates is high-dimensional and sparse, with weak feature expression capabilities. In addition, traditional machine learning algorithms require a lot of manpower and material resources in feature engineering. In addition, for anti-cancer drugs, due to their strong domain specificity, large amount of data and complex content, the existing feature extraction methods are not accurate. Summary of the invention

[0005] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides a method and system for predicting drug compound molecules that inhibit the biological activity of target proteins. The method predicts candidate drug anticancer molecules that maximally antagonize the target biological activity based on a pre-constructed relationship model between key molecular descriptor features and target biological activity and a key molecular descriptor vector representation, which can improve the accuracy of the prediction results.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] The first aspect of the present invention provides a method for predicting drug compound molecules that inhibit the biological activity of a target protein, comprising:

[0008] Obtain molecular descriptors of anticancer drug candidates;

[0009] Screening a set number of key molecular descriptors from the molecular descriptors of anticancer drug candidates according to their relevance to target biological activity;

[0010] Encode key molecular descriptors and convert them into corresponding vector representations;

[0011] Based on the pre-constructed relationship model between key molecular descriptor features and target biological activity and the key molecular descriptor vector representation, candidate anticancer molecules that maximally antagonize target biological activity are obtained.

[0012] The second aspect of the present invention provides a drug compound molecule prediction system for inhibiting the biological activity of a target protein, comprising:

[0013] A candidate drug information acquisition module, which is used to obtain molecular descriptors of anticancer candidate drugs;

[0014] A key molecular descriptor screening module, which is used to screen a set number of key molecular descriptors from the molecular descriptors of the anticancer candidate drugs according to their relevance to the target biological activity;

[0015] A key molecule vector representation module, which is used to encode key molecule descriptors and convert them into corresponding vector representations;

[0016] The maximum antagonistic molecule determination module is used to obtain candidate drug anticancer molecules that maximally antagonize target biological activity based on a pre-constructed relationship model between key molecular descriptor features and target biological activity and a key molecular descriptor vector representation.

[0017] The third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the method for predicting drug compound molecules that inhibit the biological activity of a target protein as described above.

[0018] The fourth aspect of the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for predicting drug compound molecules that inhibit the biological activity of a target protein as described above are implemented.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] The present invention screens out a set number of key molecular descriptors from the molecular descriptors of candidate anticancer drugs. It not only considers a plurality of different molecular descriptors of drug compounds, but also starts from the perspective of "the structure of a substance determines its properties" in biomolecular physics, increases compound structure similarity analysis variables by utilizing the correlation with target biological activity, increases the dimension of the variable matrix, and utilizes a pre-constructed relationship model between key molecular descriptor features and target biological activity, thereby improving the accuracy of the prediction results of candidate drug anticancer molecules that maximally antagonize the target biological activity.

[0021] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0023] Figure 1 This is a diagram showing the process of the drug exerting anticancer effect in an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of calculating the correlation coefficient of the molecular descriptors of drug compounds in an embodiment of the present invention;

[0025] Figure 3 This is a flow chart of a method for predicting drug compound molecules that inhibit the biological activity of a target protein in an embodiment of the present invention;

[0026] Figure 4 This is a schematic diagram of the training and testing process of the quantitative prediction model based on machine learning in an embodiment of the present invention;

[0027] Figure 5 A schematic diagram of comparing multiple prediction model indicators in an embodiment of the present invention;

[0028] Figure 6 This is a schematic diagram of the effect of the prediction model for inhibiting the biological activity of the target ERα in an embodiment of the present invention;

[0029] Figure 7 This is a schematic diagram of improving n_estimators evaluation in random forest based on machine learning in an embodiment of the present invention;

[0030] Figure 8 Schematic diagram of the structure of the drug compound molecule prediction system for inhibiting the biological activity of the target protein in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0032] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0033] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0034] Embodiment 1

[0035] like Figure 1 and Figure 3 As shown, this embodiment provides a method for predicting drug compound molecules that inhibit the biological activity of a target protein, which specifically includes the following steps:

[0036] S101: Obtain molecular descriptors of anticancer drug candidates.

[0037] S102: Screening out a set number of key molecular descriptors from the molecular descriptors of the anticancer drug candidates according to their relevance to the target biological activity.

[0038] In step S102, the correlation between the molecular descriptor of the candidate anticancer drug and the target biological activity is characterized by using the Pearson correlation coefficient.

[0039] Among them, the relationship between the molecular indicators of different samples is intricate. Such a relationship will lead to repeated scoring when calculating the correlation, making the molecular measurement unbalanced. Therefore, considering using the Pearson correlation coefficient, establishing a numerical independence calculation model, and adjusting the original weights, it can solve the problems of repeatability and stability and objectively reflect the true situation of the molecular descriptors and biological activity effects in this problem.

[0040] The Pearson Correlation Coefficient is applicable to continuous variables with normal distribution. When the number of data sets exceeds 500, it can be approximately considered that the data are normally distributed.

[0041]

[0042] Among them, x i ,y i is the actual data value, and x and y are the corresponding values ​​of the corresponding model. When comparing indicators, take the current indicator as the model and observe the changes of x and y. It can be seen that when the change trends of x and y are the same, the covariance is positive; when the change trends of x and y are opposite, the covariance is negative. The correlation coefficient has a large value range, so it is necessary to map the value range to [-1, 1] and divide the covariance by the standard deviation.

[0043]

[0044] The correlation coefficient reflects the degree of correlation well. Its value range is [-1, 1]. Values ​​close to 1 indicate good correlation, 0 indicates no correlation, and values ​​close to -1 indicate negative correlation.

[0045] The overall correlation coefficient is denoted by ρ:

[0046]

[0047] in, are the overall variances of X and Y, respectively, and COV(X, Y) is the covariance of variables x and y mentioned in the above work. The correlation coefficient can well achieve decentralized processing. In order to clearly show the coefficient relationship between indicators, the correlation coefficient matrix is ​​used to show the correlation between every two indicators.

[0048] The covariance is the premise for calculating the Pearson correlation coefficient, and represents the overall error of the two variables. If the change trends of the two variables are consistent, that is, if one is greater than its expected value and the other is also greater than its expected value, then the covariance between the two variables is positive. If the change trends of the two variables are opposite, that is, one is greater than its expected value and the other is less than its expected value, then the covariance between the two variables is negative. That is, it represents the numerical characteristics of the relationship between the two variables (x, y).

[0049] Specifically, the process of screening out a set number of key molecular descriptors is as follows:

[0050] S1021: Pearson correlation coefficient was used to calculate the correlation matrix, and linear correlation analysis was performed on each molecular descriptor to obtain the correlation matrix;

[0051] S1022: Select molecular descriptors, and obtain a set number of molecular descriptors with higher correlations according to the variable correlations of the correlation matrix.

[0052] The correlation coefficient is a measure of the degree of correlation between variables, and its value range is between -1 and 1. A positive value indicates positive correlation, that is, the direction of change of the variables is consistent, for example, y increases as x increases; when the correlation coefficient is 1 or -1, there is a strict linear relationship between the two; when the correlation coefficient is 0, x and y are said to be uncorrelated. Uncorrelated means that there is no linear relationship between x and y, but there may be other functional relationships between x and y.

[0053] The closer the absolute value is to 1, the closer the relationship between the two variables is; the closer it is to 0, the less close the relationship between the two variables is.

[0054] S103: Encode the key molecular descriptors and convert them into corresponding vector representations.

[0055] S104: Based on the pre-constructed relationship model between key molecular descriptor features and target biological activity and the key molecular descriptor vector representation, the candidate drug anticancer molecule that maximally antagonizes the target biological activity is obtained.

[0056] In step S104, a random forest model supervised by the drug compound structure is constructed based on the pre-constructed relationship model between the key molecular descriptor features and the target biological activity, and then combined with the key molecular descriptor vector representation, the candidate drug anticancer molecule that maximally antagonizes the target biological activity is predicted.

[0057] The random forest model based on drug compound structure supervision contains multiple decision trees, and the prediction results of the multiple decision trees are averaged to obtain the final result.

[0058] In the random forest model based on drug compound structure supervision, a binary tree is used to represent a class of variables.

[0059] The specific implementation process of this embodiment is described in detail below by taking the inhibition of the biological activity of the breast cancer target ERα and screening the main anti-cancer molecules in the drug, that is, screening the drug compound molecular descriptors that have the most significant impact on the target biological activity as an example:

[0060] Specifically, in the process of building a relationship model between key molecular descriptor features and target biological activities, data to be screened with existing labels are obtained. These data are provided by the official website of the competition. There are 1,974 compounds and their 729 molecular descriptor sample data. The specific attribute columns of the data include compound structure, nAcid, ALogP, ALogp2, IC50, pIC50 and other information, which correspond to the 729 molecular descriptors and the strength of the drug compounds in antagonizing the biological activity of the target ERα.

[0061] This embodiment defines this process as a typical feature selection problem, and analyzes the correlation between each molecular descriptor and the target biological activity. First, the variable screening model is modeled according to the different types of molecular variables. Secondly, the direct Pearson correlation coefficient is selected for correlation analysis. The algorithm can adapt to linear and nonlinear situations and accurately find the correlation between each variable and the active factor. Then, the model is solved, and the top 20 variables are selected from all variables to become the main variables of the model. Finally, the independence and rationality of the main variables are verified and discussed using the correlation analysis method.

[0062] Take the 729 molecular descriptors of 1974 compounds as an example:

[0063] In order to further observe the relationship between each molecular descriptor of the compound and the biological activity of the target ERα, the correlation strength corresponding to the correlation coefficient in the Pearson correlation analysis was established based on the correlation analysis between the features as shown in Table 1:

[0064] Table 1 Correlation degree and coefficient relationship table

[0065] Correlation coefficient Relevance 0.8-1.0 Very strong correlation 0.6-0.8 Strong correlation 0.4-0.6 Moderately related 0.2-0.4 Weak correlation 0.0-0.2 Very weak or no correlation -1.0-0.0 Negative correlation

[0066] Using the established correlation calculation model, a quantitative prediction model for the biological activity of drug compounds on the target ERα was constructed. Correlation analysis showed that the molecular descriptors describing the top 20 most effective biological activities included "MDEC-23", "MLogP", "LipoaffinityIndex", "maxsOH", "minsOH", "nC", "nT6Ring", "n6Ring", "minsssN", "BCUTp-1h", "C2SP2", "hmin", AMR", "SwHBa", "maxsssN", "MDEC-22", "SP-5", "SaaCH", "CrippenLogP", and "maxHsOH". The experimental results of this embodiment are shown in Table 2:

[0067] Table 2 Top 20 molecular descriptors and their correlation coefficients

[0068] Molecular descriptors Correlation coefficient Molecular descriptors Correlation coefficient MDEC-23 0.538048 C2SP2 0.426632 MLogP 0.529321 hmin 0.426365 LipoaffinityIndex 0.491855 AMR 0.425149 maxsOH 0.466621 oeLh 0.423176 minsOH 0.466127 maxsssN 0.420674 nC 0.459549 MDEC-22 0.420095 nT6Ring 0.440443 SP-5 0.419318 n6Ring 0.432108 SaaCH 0.418828 minsssN 0.430690 CrippenLogP 0.412300 BCUTp-1h 0.428526 maxHsOH 0.408761

[0069] In order to more intuitively display the 20 molecular descriptors most relevant to the target bioactive factors, this example draws a related bar chart, such as Figure 2 As shown in Table 2 and Figure 2 It can be seen that the correlation coefficient between the molecular descriptors with significant influence and the biological activity of the compounds ranges from 0.4 to 0.6, which shows that there is a certain positive correlation between the two. Among them, the correlation coefficients of the MDEC-23 molecular descriptor and the MLogP molecular descriptor are relatively high, both above 0.5, and the correlation coefficients of the remaining 18 molecular descriptors (such as LipoaffinityIndex, maxsOH, minsOH, etc.) are not much different, and are evenly distributed between 0.4 and 0.5.

[0070] Comparison with Table 2 shows that the influence of a single molecular descriptor on the biological activity of the target ERα is moderately correlated. Therefore, by appropriately adding this type of molecular descriptor to the compound and participating in the further development of drugs, it is possible to effectively inhibit the overexpression of ERα and thus inhibit the further development of breast cancer.

[0071] In this embodiment, by comparing the output data of the model, it is found that the influence degree of most molecular descriptors of compounds on the biological activity of the compounds, that is, the numerical value range of the correlation coefficient is between 0.2 and 0.6. From the perspective of the strength of the correlation, the influence degree of most molecular descriptors on the biological activity of the compounds is moderately correlated or weakly correlated. However, the correlation coefficient of some molecular descriptor data is very small, almost close to 0, as shown in Table 3, and the correlation coefficient of some data is even negative, as shown in Table 4.

[0072] Table 3 Weak correlation table of the top 20 molecular descriptors

[0073] Molecular descriptors Correlation coefficient Molecular descriptors Correlation coefficient SHBint5 0.005773 SaaaC 0.004102 SXWh 0.004559 nHBint9 0.004099 maxHt 0.004559 nHBint8 0.003545 HkDJ 0.004559 SHBint8 0.003483 nH 0.004372 maxaaaC 0.002753 NT 0.004372 nF10Ring 0.002626 nHBint5 0.004299 nT10Ring 0.002626 maxtCH 0.00428 nT12Ring 0.001438 StCH 0.00428 minaaaC 0.001011 mintCH 0.00428 nF12Ring 0.000713

[0074] Through the correlation analysis and Table 4, we can see that the correlation coefficient values ​​of some molecular descriptors are negative. From the perspective of correlation strength, there is a certain negative correlation between molecular descriptors and the biological activity of compounds. Theoretically, as the number of molecular descriptors of this type increases, the biological activity of the compound decreases, that is, the expression of the intensity of the activity of the compound inhibited by this type of molecular descriptor. Therefore, in the process of drug development, attention should be paid to the effect of molecular descriptors of this type of compound on the development of breast cancer. They have less effect in inhibiting the growth of cancer cells and may even have side effects. This type of chemical molecule should be used reasonably in drug development.

[0075] Table 4 Negative correlation table of the top 20 molecular descriptors

[0076] Molecular descriptors Correlation coefficient Molecular descriptors Correlation coefficient C2SP1 -0.006544 VC-4 -0.074853 ATSc1 -0.028127 ATSc5 -0.075515 nP -0.035941 nBondsT -0.077666 nI -0.044223 C1SP1 -0.078458 VC-6 -0.047296 C4SP3 -0.078581 SC-5 -0.047753 nBr -0.092515 SC-4 -0.051509 SCH-5 -0.136838 BCUTw-1h -0.056019 BCUTp-1l -0.13885 SC-6 -0.056331 BCUTw-1l -0.163833 nCl -0.061131 n -0.186221

[0077] The above content realizes the analysis and screening of drug molecular descriptor variables. By calculating the similarity, a total of 20 representative, independent molecular descriptors with the most significant impact on biological activity (including "MDEC-23", "MLogP", "LipoaffinityIndex" and "maxsOH", etc.) were screened out, and the variables were displayed to illustrate the accuracy and superiority of their screening.

[0078] Using the given sample data of target ERα bioactivity values ​​and the screened molecular descriptors with the most significant impact on bioactivity, a series of drug compound molecular descriptors are used as independent variables, and the target bioactivity values ​​are used as dependent variables. According to the characteristics of sample data and screening variables, the relationship between the two is analyzed, and a quantitative structural property / bioactivity relationship model is constructed to predict the IC50 values ​​and pIC50 values ​​of multiple compounds.

[0079] From chemical knowledge, we know that cancer drugs are derived from compounds with the same medicinal properties, and the structural similarity between them is high. There is a specific correlation between the molecular structure and biological activity of the compound. The study of the material basis of anticancer drug properties from the relationship of "compound structure-molecular descriptor-biological activity" has a certain scientific basis. The molecular descriptors corresponding to similar compounds must also be similar, and the final biological activity effects are also similar. Therefore, in this embodiment, in addition to paying attention to the biological activity value of the target ERα and the information of multiple molecular descriptors, the compound structure similarity information will be creatively used as prior knowledge to assist in the prediction of biological activity.

[0080] In this example, a random forest model supervised by compound structure is established, and the chemical structure similarity index is calculated based on the Hellinger distance between atomic environments (AE). The similarity of chemical structures is used as additional supervision information for the model. This model has more obvious advantages in the prediction process.

[0081] The training samples and data sets of 1974 compounds provided in the sample data were processed. Considering that the relevant features of different compounds are different and the training data are limited, different compounds and their features are used as a training sample to generate 1974 training data. In order to effectively evaluate the performance of different models, for the data provided in the training set, the training set and validation set were randomly divided into 6:1 in stage 2, and 1700 were selected as training data and 294 as validation set data.

[0082] Because the data set is very small, if these data are directly used for model classification training, it is difficult for the model to truly and accurately learn the mapping relationship between input and output. Therefore, the cross-validation method is used in this embodiment to avoid overfitting. First, multiple model selections (hyperparameter selections) are verified on the training set and the validation set, and the model (hyperparameter) with the smallest average error is selected. After selecting the appropriate model (hyperparameter), the training set and the validation set can be combined, and the model can be retrained on it to obtain the final model, and then the test set can be used to test its generalization ability. The detailed model architecture diagram is shown in the figure. Figure 4 shown.

[0083] In the process of constructing a simple regression model, a suitable regression fitting function expression is manually designed, while the random forest algorithm model breaks through the limitations of the function form setting and is more suitable for the prediction of multivariate variables such as multivariate molecular descriptors in this embodiment. As a commonly used and effective algorithm in the field of machine learning and data analysis, regression random forest is very necessary for constructing a model for quantitative prediction of the biological activity of compounds on the target ERα.

[0084] In addition to using the molecular descriptor information and activity value information provided in the original data, the key to establishing this model is to calculate the similarity of compound structures as prior knowledge. Currently, the methods for calculating compound structural similarity coefficients include two methods: based on the molecular descriptors of drug compounds and based on compound structural information.

[0085] In the process of calculating compound similarity in this embodiment, toxmatch is mainly used as a calculation and visualization tool, and a structure-based similarity index calculation mode is applied to calculate the structural similarity between different compounds. This calculation mode includes the Hellinger distance between atomic environments (AEs), and the formula is as follows:

[0086]

[0087] The Hellinger distance between AE is: In the equation, N is the total number of all fragments in the molecular structures A and B, i refers to the probability of a fragment in compound A, and B i Refers to the probability of finding the first fragment in compound B. The Hellinger distance calculation method based on AE is used to calculate the similarity between chemical molecular formulas. The chemical component pairs with the structural similarity results between 0.6 and 0.7 based on atomic AE are the most common, and 0.7 to 0.8 are also common. The similarity between some components can even reach above 0.9.

[0088] In this embodiment, the drug compound correlation matrix is ​​used as supervisory information to guide the establishment of the random forest model. The specific establishment process is as follows:

[0089] Step 1: There are N samples in the original training set, and each sample has W-dimensional features. Randomly select x samples with replacement from the data set to form a training subset, and perform w sampling times in total, that is, generate w training subsets;

[0090] Step 2: Each training subset forms a decision tree, and a total of w decision trees are formed;

[0091] Step 3: For a single decision tree, randomly select m features from M features at each node of the tree and split them according to the principle of minimum node impurity. Each tree is split in this way until all training samples of the node belong to the same category. No pruning is required during the splitting process of the decision tree;

[0092] Step 4: Predict the data to be predicted based on the generated multiple decision tree classifiers. For regression trees, the final result is obtained by a simple average based on the voting results of each tree.

[0093] The RFR implemented in this embodiment is to combine multiple binary decision trees (CART). When training the binary decision tree model, it is necessary to consider how to select the splitting variables (features). For the selection of splitting variables and splitting points, this implementation adopts an exhaustive method, that is, traversing each feature and all the values ​​of each feature, and finally finding the best splitting variables and splitting points; for the quality of the splitting variables and splitting points, it is generally measured by the impurity of the node after the split, that is, the weighted sum of the impurities of each child node G (xi, vij), and its calculation formula is as follows:

[0094]

[0095] Among them, x i is a segmentation variable, v ij is a cutoff value of the cutoff variable, n left , n right , N s are the number of training samples of the left child node after segmentation, the number of training samples of the right child node, and the number of all training samples of the current node, respectively. left , X right Divided into a set of training samples of left and right child nodes, H(X) is a function to measure the node impurity (impurity function / criterion), and classification and regression tasks generally use different impurity functions. In the regression random forest implemented in this embodiment, H(X) uses MSE, that is, for a certain cut point:

[0096]

[0097] Among them, X left , X right Divided into a set of training samples for left and right child nodes, y i is one of the examples; is the mean of the left child node training sample set, is the mean of the combined training samples of the right child node; N s An overview of the collection fragments.

[0098] RFR extracts n samples with replacement from the original breast cancer data to form multiple sub-sample sets, and then uses the CART tree method to build decision trees for each of these sub-sample sets, ultimately forming a random forest. The prediction results of the new data are determined by voting of the classification trees. The prediction results of the new data can be shown as follows:

[0099]

[0100] Among them, h iis the basic classification model of a single classification tree, Y is the output variable (TNBC and non-TNBC), I(·) represents the indicative function, and n is the total number of samples.

[0101] This embodiment uses a random forest model supervised by compound structure to predict target biological activity, and the classification accuracy of the model is greatly improved.

[0102] The model solving process mainly includes the process of training, testing and result visualization. Python is used to solve the random forest-based bioactivity value prediction model. The training stage is as follows Figure 4 shown.

[0103] In the training phase, the regression random forest (RFR) uses bootstrap sampling to collect multiple different sub-training data sets from the input training data set to train multiple different decision trees in sequence, and adds the compound-related weight coefficients as list features to the model training. In the prediction phase, the random forest averages the prediction results of multiple internal decision trees to obtain the final result.

[0104] The trained breast cancer target bioactivity model was used for prediction and evaluation:

[0105] Selection of model evaluation indicators:

[0106] The mainstream indicators for evaluating regression problems include mean absolute error (MAE), mean squared error (MSE) also known as L2, root mean square error (RMSE), explained variation, goodness of fit (R-squared) and coefficient of determination.

[0107] For the random forest model under compound structure supervision proposed in this embodiment, when using evaluation indicators for evaluation, three indicators, MSE, MAE and R-squared, are selected to evaluate the model.

[0108] MSE is not only a measure of branch quality, but also a measure of the regression quality of the estimator and the estimated quantity. The mean square error formula in the random forest regression model is as follows:

[0109]

[0110] in, is the actual value, y i is the predicted value, and n is the total number of samples.

[0111] MAE takes the average of the absolute errors between the predicted value and the actual value. In the formula for mean square error, the square of the difference between the predicted value and the actual value is replaced by the absolute value of the difference. The formula is as follows:

[0112]

[0113] in, is the actual value, y i is the predicted value, and n is the total number of samples.

[0114] The R-squared is the proportion of the regression sum of squares to the total sum of squares. The R-squared formula can be obtained by the following calculation steps:

[0115]

[0116] So R squared:

[0117]

[0118] in, represents the mean of the observed values, is the actual value, y i is the predicted value. The closer the R-squared is to 1, the greater the proportion of the regression sum of squares in the total sum of squares.

[0119] Model evaluation results analysis:

[0120] In this embodiment, by calculating the established random forest regression model, the regression model's fit (R-squared), mean absolute error (MAE), and mean square error (MSE) are selected to evaluate the prediction model. The results of the above three indicators are shown in Table 5.

[0121] Table 5 Prediction model evaluation table

[0122]

[0123] The mean square error (MSE), the smaller the value, the worse the effect of the regression model;

[0124] The mean absolute error (MAE) is used to evaluate the closeness between the prediction result and the real data set. The smaller the value, the better the fitting effect.

[0125] The goodness of fit (R_squared) is the mean of the sum of squares of the errors between the fitted data and the sample points corresponding to the original data. The smaller the value, the better the fitting effect. As can be seen from the table, compared with multivariate linear regression, random forest performs better in MSE, MAE, and R_squared. Therefore, the fitting effect of random forest is better.

[0126] This embodiment uses cross validation and selects mean square error as the evaluation criterion. In the regression prediction model, the smaller the MSE, the better the trained model. R square can be positive or negative (if the residual sum of squares of the model is much larger than the total sum of squares of the model, the model is very bad and R square will be negative), while the mean square error is always positive. It can describe the level of fit of the multivariate regression equation and reflects the degree of explanation when the dependent variable y deteriorates due to the regression relationship alone.

[0127] As can be seen from Table 5, compared with Lasso, random forest has better performance in the three regression model indicators of MSE, MAE, and R_squared, which means that random forest has a smaller generalization error. Through the results, this embodiment can guess that the data is not strictly linearly distributed, but presents a random nonlinear distribution. In order to more intuitively display the model and test results, we use Figure 5 Displayed in the form of a bar graph.

[0128] Therefore, from the histogram Figure 5 It can be seen that the random forest prediction model has the best fitting effect, and the worst effect is multivariate linear regression. According to the number of samples, 10 rounds of cross-validation method are used to verify the prediction model. From the verification results, it can be seen that each prediction model shows good generalization and high prediction stability. The above prediction models are verified by cross-validation using multiple data in the test set. Then the image reflects the characteristics that the closer the regression line is to each observation point, which means that the proportion of the total variation of the y value that can be explained by the change of x is greater, so the regression fitting effect is better.

[0129] These three indicators have consistent results, and there will not be a situation where one model has a larger MSE but a smaller MAE than another model. The better the performance of the model, the smaller the mean square error and mean absolute error, and the closer the R-squared value is to 1.

[0130] In this embodiment, the regression prediction fitting result is as follows: Figure 6 As shown, the abscissa represents the compound number, the ordinate represents the pIC50 value, the dot curve represents the predicted value and the inverted triangle curve represents the actual value.

[0131] It can be seen from the figure that the use of linear regression such as random forest can achieve a good fit, and the trend of the predicted value is the same as the real data. At the same time, 10 rounds of cross-validation are used during the training process to ensure the integrity and accuracy of the model.

[0132] In this embodiment, the main parameter affecting the random forest algorithm is n_estimators: the number of decision trees. Theoretically, the more decision trees there are, the better the model will perform. However, as the number of decision trees increases, the marginal utility of improving model performance gradually approaches 0, which will lead to overfitting of computing resources.

[0133] Figure 7 It shows the change of MSE function value with the change of parameter n_estimators. The parameter adjustment process is as follows:

[0134] In the first step, for the parameter adjustment data, the data is randomly divided into two parts, one as training data, and a loss function is calculated under a certain n_estimators parameter value. The same test data uses the trained model to calculate a loss function value, and then the two data are swapped and the above process is repeated. The loss function value of the training data and the loss function value of the test data under the n_estimators parameter value are the average of the two calculations.

[0135] In the second step, for each value of n_estimators, repeat the process of the first step to obtain the training data loss function value and the test data loss function value under each value of n_estimators. Figure 7 The horizontal axis is the number of n_estimators, and the vertical axis is the value of the loss function. The loss function in this embodiment is the mean square error (mean_squared_error). As the number of decision trees increases, the loss function of the training data set changes. As the number of decision trees increases, the gap between the predicted value of the model and the true value increases.

[0136] This embodiment sets the value range of n_estimators from 3 to 14. Overall, as the number of decision trees increases, the error tends to increase. When the number of decision trees is 4, 7, and 8, the generalization error of the model is the smallest. After 8, the generalization error of the model increases. This embodiment predicts that the generalization error may be larger after 14, so there is no need to continue verification. Finally, the parameter value of n_estimators is selected to be 8.

[0137] In this embodiment, the data set is the data provided by the official website, wherein the appendix ERα_activity contains 1974 sample data. The target biological activity prediction task is completed using the random forest model supervised by the compound structure using the given appendix data set. This embodiment aims to maximize the inhibition of target ERα biological activity expression by drug compounds.

[0138] The method of the present embodiment is to collect a series of compounds and their biological activity data acting on the target, and then use a series of molecular structure descriptors as independent variables and the biological activity value of the target protein as the dependent variable to construct a quantitative structure property / activity relationship (QSAR). The QSAR refers to a theoretical research method for predicting the properties of substances by calculation methods, which is widely used in the fields of drug design, material chemistry, environmental toxicology, etc. The QSAR method is to model the correlation between the descriptors of the substance and its physical and chemical properties with statistical methods after theoretical calculation, and then reveal the influence of a certain feature of the substance on the properties.

[0139] Regarding the quantitative prediction model of the biological activity of anticancer molecules in drug compounds on the breast cancer target ERα, an effective model is established based on the QSAR method, and the biological activity values ​​of the target protein ERα related to breast cancer, the structural information of drug compounds and their corresponding molecular descriptor data information are mined and analyzed to study its molecular mechanism. The model is used to predict new compound molecules with better biological activity, or to guide the structural optimization of existing active compounds, to accelerate the research and development process of candidate breast cancer drugs, and to provide a theoretical reference for clinical treatment research of breast cancer.

[0140] Embodiment 2

[0141] like Figure 8 As shown, this embodiment provides a drug compound molecule prediction system for inhibiting the biological activity of a target protein, which specifically includes the following modules:

[0142] A candidate drug information acquisition module, which is used to obtain molecular descriptors of anticancer candidate drugs;

[0143] A key molecular descriptor screening module, which is used to screen a set number of key molecular descriptors from the molecular descriptors of the anticancer candidate drugs according to their relevance to the target biological activity;

[0144] A key molecule vector representation module, which is used to encode key molecule descriptors and convert them into corresponding vector representations;

[0145] The maximum antagonistic molecule determination module is used to obtain candidate drug anticancer molecules that maximally antagonize target biological activity based on a pre-constructed relationship model between key molecular descriptor features and target biological activity and a key molecular descriptor vector representation.

[0146] Specifically, in the maximum antagonist molecule determination module, a random forest model supervised by the drug compound structure is constructed based on the pre-constructed relationship model between the key molecular descriptor features and the target biological activity, and then combined with the key molecular descriptor vector representation, the candidate drug anticancer molecule that maximally antagonizes the target biological activity is predicted.

[0147] It should be noted here that each module in this embodiment corresponds to each step in Example 1 one by one, and the specific implementation process is the same, which will not be repeated here.

[0148] Embodiment 3

[0149] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-mentioned method for predicting drug compound molecules that inhibit the biological activity of a target protein.

[0150] Embodiment 4

[0151] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for predicting drug compound molecules that inhibit the biological activity of a target protein as described above are implemented.

[0152] The present invention is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products of the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram and the combination of the processes and / or blocks in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0153] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting drug compound molecules that inhibit the biological activity of a target protein, characterized in that: include: Obtain molecular descriptors of anticancer drug candidates; Screening a set number of key molecular descriptors from the molecular descriptors of anticancer drug candidates according to their relevance to target biological activity; The correlation between the molecular descriptors of anticancer drug candidates and the target biological activity was characterized by the Pearson correlation coefficient; The process of screening out a set number of key molecular descriptors is as follows: The Pearson correlation coefficient was used to calculate the correlation matrix, and linear correlation analysis was performed on each molecular descriptor to obtain the correlation matrix; Select molecular descriptors, and obtain a set number of molecular descriptors with higher correlation according to the correlation size of the variables in the correlation matrix; Encode key molecular descriptors and convert them into corresponding vector representations; Based on the pre-built relationship model between key molecular descriptor features and target biological activity and the key molecular descriptor vector representation, the candidate anticancer molecule that maximally antagonizes the target biological activity is obtained; Specifically, based on the pre-constructed relationship model between key molecular descriptor features and target biological activity, a random forest model supervised by the drug compound structure is constructed, and then combined with the key molecular descriptor vector representation, the candidate drug anticancer molecule that maximally antagonizes the target biological activity is predicted; The random forest model based on drug compound structure supervision contains multiple decision trees, wherein the prediction results of the multiple decision trees are averaged to obtain the final result.

2. The method for predicting drug compound molecules that inhibit the biological activity of a target protein according to claim 1, characterized in that: In the random forest model based on drug compound structure supervision, a binary tree is used to represent a class of variables.

3. A molecular prediction system for drug compounds that inhibit the biological activity of a target protein, characterized in that: include: A candidate drug information acquisition module, which is used to obtain molecular descriptors of anticancer candidate drugs; A key molecular descriptor screening module, which is used to screen a set number of key molecular descriptors from the molecular descriptors of the anticancer candidate drugs according to their relevance to the target biological activity; A key molecule vector representation module, which is used to encode key molecule descriptors and convert them into corresponding vector representations; A maximum antagonistic molecule determination module, which is used to obtain a candidate drug anticancer molecule that maximally antagonizes the target biological activity based on a pre-built relationship model between key molecular descriptor features and target biological activity and a key molecular descriptor vector representation; The correlation between the molecular descriptors of anticancer drug candidates and the target biological activity was characterized by the Pearson correlation coefficient; The process of screening out a set number of key molecular descriptors is as follows: The Pearson correlation coefficient was used to calculate the correlation matrix, and linear correlation analysis was performed on each molecular descriptor to obtain the correlation matrix; Select molecular descriptors, and obtain a set number of molecular descriptors with higher correlation according to the correlation size of the variables in the correlation matrix; In the maximum antagonist molecule determination module, a random forest model supervised by the drug compound structure is constructed based on the pre-constructed relationship model between the key molecular descriptor features and the target biological activity, and then combined with the key molecular descriptor vector representation, the candidate drug anticancer molecule with the maximum antagonism to the target biological activity is predicted; The random forest model based on drug compound structure supervision contains multiple decision trees, wherein the prediction results of the multiple decision trees are averaged to obtain the final result.

4. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the method for predicting drug compound molecules that inhibit the biological activity of a target protein as described in any one of claims 1 to 2 are implemented.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the method for predicting drug compound molecules that inhibit the biological activity of a target protein as described in any one of claims 1 to 2 are implemented.

Images