Virtual screening method of tyrosinase inhibitor and composition with whitening function

Abstract

The invention discloses a virtual screening method of a tyrosinase inhibitor and a composition with a whitening function. The method comprises the following steps: firstly, constructing and preprocessing a data set of a tyrosinase inhibitor, then respectively establishing a QSAR classification model and a QSAR regression model, and carrying out model training and parameter optimization by utilizing various machine learning algorithms. Virtual screening is performed on a large-scale compound library through the two models, candidate compounds with high inhibitory activity on tyrosinase are preliminarily screened out, and the activity of the candidate compounds is further verified through an in-vitro tyrosinase inhibition test. Experimental results show that the screened compound has a remarkable inhibition effect on tyrosinase and has potential application prospects in whitening cosmetics and freckle-removing medical products. The method is low in cost and short in time consumption, the novel tyrosinase inhibitor can be efficiently found, and a powerful tool is provided for research and development of raw materials with whitening effects.

Description

Technical Field

[0001] The invention belongs to the field of chemical informatics, and specifically relates to a virtual screening method for tyrosinase inhibitors, and a composition with whitening function. Background Art

[0002] Tyrosinase (TYR) is an enzyme that plays a key role in the synthesis of melanin, and its catalytic activity directly affects the production of melanin. Tyrosinase inhibitors can directly inhibit the catalytic activity of tyrosinase, thereby reducing the production of melanin and having a significant whitening effect. However, the tyrosinase inhibitors currently used in cosmetics on the market, such as arbutin, kojic acid, and glabridin, have certain safety risks, addition restrictions, or high prices, which has prompted the industry to urgently develop new, efficient, and safe tyrosinase inhibitors. The traditional method of developing cosmetic active ingredients relies on in vitro and in vivo experimental tests. This process is not only time-consuming and costly, but the expected results are often not ideal, which greatly increases the difficulty of developing innovative functional raw materials. Although high-throughput screening technology can complete the in vitro activity screening of a large number of compounds in a short period of time with the help of automated experimental methods, its high platform cost and screening costs, as well as the limitation that it is not applicable to all biological targets, make this technology not a panacea.

[0003] In recent years, the rapid development of machine learning technology has provided a powerful research tool for the fields of chemical informatics, pharmacy, and bioinformatics. With the help of machine learning technology, a quantitative structure-activity relationship (QSAR) model can be constructed to achieve virtual screening of compounds to predict their biological activity from their molecular structure characteristics. Compared with traditional experimental methods, virtual screening has significant advantages such as low cost, short time consumption, and high success rate. Therefore, virtual screening using machine learning QSAR models has become an important way to discover new tyrosinase inhibitors. Based on the above background, the present invention proposes a virtual screening method for tyrosinase inhibitors, which aims to achieve the goal of quickly and efficiently discovering new tyrosinase inhibitors through virtual screening and in vitro activity verification, and provide strong support for the research and development of whitening cosmetics and freckle removal medical products. Summary of the invention

[0004] To achieve the above object, the technical solution provided by the present invention is a virtual screening method for tyrosinase inhibitors, comprising the following steps:

[0005] S1. Building a data set: Building a data set based on the biological activity data of tyrosinase inhibitors. The method for obtaining the biological activity data includes the following steps:

[0006] (1) Collect tyrosinase inhibitors with known activity and their IC from the ChEMBL database and literature50 value;

[0007] (2) Remove non-covalent salt compounds and a small number of silicon- and halogen-containing compounds from the data set;

[0008] (3) Select compounds with molecular weights in the range of 100-500 and retain IC 50 The duplicate compound with the smallest value;

[0009] (4) Use the Rdkit Structure Normalizer node in KNIME to delete compounds with unclear structures;

[0010] S2. Construct QSAR classification model: divide the data set obtained in S1 into training set and test set by random stratified sampling; calculate molecular fingerprint for each inhibitor compound; optimize molecular fingerprint by feature selection method in training set; train QSAR classification model by machine learning classification algorithm; optimize hyperparameters by grid search method with 5-fold stratified cross validation; select the model corresponding to the machine learning regression algorithm with the highest average F1 score as the final QSAR classification model;

[0011] S3. Construct QSAR regression model: Use Butina clustering algorithm to divide the data set obtained in S1 into training set and external test set. Specifically, use Butina algorithm to cluster the data set compounds into 21 clusters. For small cluster compounds, randomly select one compound to be included in the external test set, and the remaining compounds are classified as training set; for large cluster compounds, randomly select 12% of the compounds to be included in the test set; calculate the molecular descriptors of the compounds, including Mordred (1D, 2D) descriptors and Padel (1D, 2D) descriptors; calculate three types of molecular fingerprints of the compounds (CDK, CDK-Extend and ECFP); train QSAR regression model by machine learning regression algorithm; use 5-fold stratified cross-validation grid search method for hyperparameter optimization; select average R 2 The corresponding model obtained by configuring the machine learning regression algorithm with the highest hyperparameters is the final QSAR regression model;

[0012] S4. Virtual screening of tyrosinase inhibitors: First, based on the molecular characteristics of the highly active compounds in the dataset, candidate compounds were screened from the ZINC20 database. Then, the QSAR classification model was used for preliminary screening. Finally, the QSAR regression model was used to predict the IC 50 value.

[0013] The calculated molecular fingerprint types described in S2 and S3 include: CDK, CDK-Extend, MACCS, Substructure, Pubchem, Klekota-Roth, ECFP and FCFP; the radius of the ECFP molecular fingerprint can be selected as 1, 2, or 3; the number of bits can be selected as 512, 1024, 2048, 4906, or 8192.

[0014] In step S3, the process of optimizing the molecular descriptors and / or molecular fingerprints in the training set by using a feature selection method is also included.

[0015] The feature selection method for optimizing molecular fingerprints is one or a combination of zero variance filtering, chi-square test, mutual information, Kendall rank correlation coefficient, and Point-biserial correlation coefficient; the feature selection method for optimizing molecular descriptors is one or a combination of low variance filtering, Spearman rank correlation coefficient, Pearson correlation coefficient, and embedding method.

[0016] The machine learning classification algorithms described in step S2 are: logistic regression (LR), ridge regression (RR), passive attack (PA), support vector machine (SVM), linear support vector machine (LSVM), random forest (RF), extreme random tree (ERT), extreme gradient boosting (XGB), lightweight gradient boosting machine (LGBM).

[0017] The hyperparameters of the hyperparameter optimization described in step S2 are: LR, PA, and LSVM optimize the regularization parameter C, RR optimizes the regularization parameter alpha, SVM selects RBF kernel coefficient gamma and regularization parameter C for optimization, RF and ERT optimize the number of estimators n_estimators, the maximum depth of the tree max_depth, and the minimum number of samples of internal nodes of the branch min_samples_split, XGB optimizes n_estimators, max_depth, min_child_weight, and learning_rate, and LGBM optimizes num_leaves, n_estimators, max_depth, min_child_samples, and learning_rate.

[0018] The machine learning regression algorithms described in step S3 are: support vector machine (SVM), random forest (RF), extreme random tree (ERT), extreme gradient boosting (XGB), lightweight gradient boosting machine (LGBM).

[0019] The hyperparameters of the hyperparameter optimization described in step S3 are: SVM selects RBF kernel, optimizes gamma, C, and epsilon, RF and ERT optimize the number of estimators n_estimators, the maximum depth of the tree max_depth, and the minimum number of samples of the internal nodes of the branch min_samples_split, XGB optimizes n_estimators, max_depth, min_child_weight, and learning_rate, and LGBM optimizes num_leaves, n_estimators, max_depth, min_child_samples, and learning_rate. All hyperparameter optimization processes follow the principle of rough search first and then fine search.

[0020] The molecular features described in step S4 are: hydrogen bond donor, hydrogen bond acceptor, LogP, and molecular weight value range.

[0021] The present invention also provides a composition with whitening function, wherein the composition comprises at least one of the following compounds of formula I, formula II and formula III, and the following compounds of formula I, formula II and formula III are obtained by virtual screening of any of the above tyrosinase inhibitors:

[0022]

[0023]

[0024] In order to further verify the actual inhibitory effect of some compounds screened out by virtual screening on tyrosinase, in vitro tyrosinase inhibition test was performed on some compounds. The specific steps include: selecting the compounds with excellent performance in virtual screening and dissolving and diluting them appropriately; using PBS buffer to prepare tyrosinase, substrate L-tyrosine, etc. to appropriate concentrations; constructing a reaction system containing tyrosinase, substrate and different concentrations of the test compound, and setting α-arbutin as a positive control group; using an ELISA instrument to detect the absorbance value of the reaction product to reflect the inhibitory activity of the sample on tyrosinase, and calculating the inhibition rate of each compound.

[0025] Compared with the prior art, the present invention has the following technical effects: (1) The present invention uses machine learning methods to carry out virtual screening of tyrosinase inhibitors, which is fast and efficient, greatly reducing research and development costs; (2) Through in vitro tyrosinase inhibition tests, it is found that the virtual screening hit compounds have obvious tyrosinase inhibitory activity. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 The figure is a flow chart of the construction of the QSAR model of the present invention;

[0027] Figure 2 The influence of the hyperparameters of the XGB algorithm modeling in Example 1 on the training set and the validation set F1;

[0028] Figure 3 PR curve of the QSAR classification model constructed by the XGB algorithm in Example 1;

[0029] Figure 4 For Example 2 SVM algorithm modeling hyperparameters for training set and validation set R 2 The impact of

[0030] Figure 5 The QSAR regression model constructed by the SVM algorithm in Example 2 predicts pIC values ​​of compounds 50 and experimental pIC 50 comparison;

[0031] Figure 6 The inhibition rate of tyrosinase by solutions of compound Ⅰ and α-arbutin at different concentrations. DETAILED DESCRIPTION

[0032] The embodiments described herein are specific embodiments of the present invention, which are used to illustrate the concept of the present invention, are illustrative and exemplary, and should not be construed as limiting the embodiments of the present invention and the scope of the present invention. In addition to the embodiments described herein, those skilled in the art can also adopt other obvious technical solutions based on the claims and the contents disclosed in the specification of this application, including technical solutions that adopt any obvious replacement and modification of the embodiments described herein.

[0033] Example 1

[0034] A machine learning QSAR classification model for tyrosinase inhibitors is provided. The method for constructing the classification model comprises the following steps:

[0035] Dataset construction: tyrosinase inhibitors with known activity and their IC values ​​were collected from the ChEMBL database and literature 50 value; remove non-covalent salt compounds and a small number of silicon-containing and halogen-containing compounds in the data set; considering that the molecular weight of commonly used whitening tyrosinase inhibitors is generally less than 500, retain compounds with a molecular weight range of 100-500; perform a standardization check on the two-dimensional structure data of the compounds, and use the Rdkit Structure Normalizer node in KNIME to delete compounds with unclear structures; all compounds are normalized and converted into standard SMILES representations.

[0036] Table 1 lists the detailed information of the five compounds in the dataset.

[0037] Molecular number Standardizing SMILES active CHEMBL1243253 CCCCNC(=O)c1ccc(-c2cc(OC)c(OC)c(OC)c2)cc1 0 CHEMBL4541503 OC[C@H]1O[C@@H](OCCCCc2ccc(O)cc2O)[C@H](O)[C@@H](O)[C@@H]1O 1 CHEMBL222541 O=c1cc(-c2ccc(O)c(O)c2)oc2c(O)c(O)ccc12 1 CHEMBL1650602 O=C(O)c1cccc(O)n1 0 CHEMBL2206927 COc1ccc(CCc2ccc(O)cc2O)cc1 0

[0038] Split the data set: According to the random stratified sampling method, use the train_test_split module of the sklearn library to randomly split the data set into training set and test set.

[0039] Calculate and select molecular fingerprints: Calculate the CDK molecular fingerprints of the compounds in the dataset, using zero-variance filtering combined with chi-square (χ 2 ) test method to select molecular fingerprints, with p value less than 0.05 as the limit, and χ 2 Specifically, the initial number of CDK molecular fingerprints was 1024, which was reduced to 1022 after zero-variance filtering and 370 after chi-square test.

[0040] Machine learning classification algorithm modeling: Use LR, RR, PA, SVM, LSVM, RF, ERT, XGB, and LGBM algorithms to build a classification prediction model on the data represented by 370 CDK molecular fingerprints. Take the XGB algorithm modeling as an example to illustrate the hyperparameter optimization process. By setting a wider parameter range, it is easy to see the impact of parameter value changes on the training set and validation set of the XGB algorithm model. It is found that n_estimators, max_depth, min_child_weight, and learning_rate all have the highest average F1 score of the validation set in a certain range ( Figure 2 ). Then, the four hyperparameters with narrower intervals are combined into a hyperparameter space for a 5-fold stratified cross-validation grid search to determine the optimal hyperparameters for configuring the classification model. 50 = 11000nM is the dividing line, and the inhibitors are marked as two categories: strong activity and weak activity. Among them, IC 50 Compounds with a concentration <11000 nM were highly active, and IC 50 Molecules with ≥11000nM are weakly active; the weight ratio of the two activity categories is adjusted to complete the construction of the QSAR classification model. Before adjustment, the weights of strong and weak active inhibitors are both 1, and after adjustment, the weight of strong active inhibitors is 1.35, and the weight of weak active inhibitors is 0.79.

[0041] Table 2-1 shows the hyperparameter space optimized by the XGB algorithm.

[0042] Hyperparameter Name Value range n_estimators {30～100} max_depth {3～7} min_child_weight {1～4} learning_rate {0.05～0.4}

[0043] Table 2-2 compares the prediction performance of the XGB algorithm model on the test set before and after adjusting the weights.

[0044] XGB Algorithm REC PRE SPE BA F1 Unadjusted weights 0.667 0.818 0.913 0.790 0.735 Adjust weights 0.741 0.851 0.924 0.832 0.792

[0045] Evaluation of the predictive performance of the QSAR classification model: The constructed QSAR classification model was used to predict the tyrosinase activity of the external test set compounds, and the predictive performance of the QSAR classification model was evaluated using REC, PRE, SPE, BA, and F1 score indicators.

[0046] Figure 3 is the PR curve of the LSVM classification model. The PR curve uses REC as the horizontal axis and PRE as the vertical axis. When the categories are unbalanced, it can better judge the prediction ability of the QSAR classification model.

[0047] Table 3 shows the prediction performance of the QSAR classification model constructed using CDK fingerprint and machine learning classification algorithm on the external test set. The F1 and REC indicators show that the model constructed by the LSVM algorithm has the best prediction performance.

[0048] Classification Algorithms REC PRE SPE BA F1 LR 0.796 0.827 0.902 0.849 0.811 RR 0.815 0.786 0.870 0.842 0.800 PA 0.796 0.796 0.880 0.838 0.796 LSVM 0.815 0.815 0.891 0.853 0.815 SVM 0.759 0.837 0.913 0.836 0.796 RF 0.667 0.857 0.935 0.801 0.750 ERT 0.704 0.864 0.935 0.819 0.776 XGB 0.741 0.851 0.924 0.832 0.792 LGBM 0.741 0.851 0.924 0.832 0.792

[0049] Example 2

[0050] A machine learning QSAR regression model for tyrosinase inhibitors is provided, and a method for constructing the regression model comprises the following steps:

[0051] Constructing the data set: Using the data set of the QSAR classification model, removing the data of qualitative judgment, and in addition, the IC 50 The value ranges from 10nM to 2.511×10 7 nM range, the IC 50 Convert to pIC 50 , the converted pIC 50 Values ​​range from 1.600 to 8.000.

[0052] Table 4 shows the detailed information of the five compounds in the regression model dataset.

[0053] Molecular number Standardizing SMILES <![CDATA[pIC 50 ]]> CHEMBL480477 CC1(C)C=Cc2c(ccc3c2OC[C@@H](c2ccc(O)cc2O)C3)O1 6.366532 CHEMBL207313 O=C(NCc1ccccc1)c1ccccc1 2.701147 CHEMBL205382 O=C(NCc1cc(O)c(O)c(O)c1)c1ccc(O)cc1O 3.259637 CHEMBL204473 O=C(NCc1ccc(O)cc1)c1ccc(O)cc1O 2.739929 CHEMBL207298 O=C(NCc1ccccc1)c1ccc(O)cc1O 2.779892

[0054] Splitting the data set: Use the Butina algorithm to cluster the compounds in the data set, set the distance threshold to 0.8, and randomly select compounds to be included in the training set and test set after clustering to complete the segmentation of the data set.

[0055] Calculate and select molecular fingerprints: Calculate the CDK molecular fingerprints of the compounds in the dataset, then calculate the Kendall rank correlation coefficient (KRCC) between the fingerprint fragments and pIC50, retain the fingerprint fragments with KRCC absolute values ​​greater than 0.08, and finally retain 381 CDK molecular fingerprint fragments.

[0056] Machine learning classification algorithm modeling: Use SVM, RF, ERT, XGB, and LGBM algorithms to build a classification prediction model. Take the SVM algorithm modeling as an example to illustrate the hyperparameter optimization process. By setting a wider parameter range, it is easy to see the impact of parameter value changes on the SVM algorithm model training set and validation set. It is found that gamma, C, and epsilon are all within a certain range. The average R of the validation set 2 The highest score ( Figure 4 ). Then, the four hyperparameters with narrower intervals were used to form a hyperparameter space for a 5-fold cross-validation grid search to determine the optimal hyperparameters for configuring the regression model and complete the construction of the QSAR regression model.

[0057] Table 5 shows the hyperparameter space optimized by the SVM algorithm.

[0058] Hyperparameter Name Value range gamma {1～8} C {0.03～0.08} Epsilon {0.1～0.2}

[0059] Evaluation of the prediction performance of the QSAR regression model: The constructed QSAR regression model was used to predict the tyrosinase activity of the external test set compounds. 2 , MAE, and RMSE indicators were used to evaluate the predictive performance of the QSAR regression model.

[0060] Figure 5 Experimental pIC for the dataset compound 50 Prediction of pIC with SVM regression model 50 The comparison chart intuitively reflects the accuracy of the regression model prediction.

[0061] Table 6 shows the prediction performance of the QSAR classification model constructed using CDK fingerprint and machine learning regression algorithm on the external test set, among which the model constructed by SVM algorithm has the best prediction performance

[0062]

[0063]

[0064] Example 3

[0065] A tyrosinase activity inhibition activity verification experiment of a virtual screening compound is provided, and the verification method includes the following steps:

[0066] Solution preparation: Use 0.1 mol / L PBS buffer solution (pH=6.5) to prepare a tyrosinase solution with a concentration of 100 U / mL; use 0.1 mol / L PBS buffer solution (pH=6.8) to prepare a 0.05% L-tyrosine solution; use 0.1 mol / L PBS buffer solution (pH=6.8) to prepare a 0.1% L-dopa solution; use 0.1 mol / L PBS buffer solution (pH=6.8) to prepare virtual screening compound I and positive control α-arbutin with appropriate mass concentration gradients.

[0067] Activity test: First, add solution (PBS buffer, L-tyrosine / L-dopa, test sample) to test tubes A, B, C, and D as shown in Table 7. After mixing, heat each tube in a water bath at 37°C for 10 minutes. Then add tyrosinase solution to tubes A and C, and heat in a 37°C water bath for 25 minutes. Finally, transfer the reaction solution of each tube into a 96-well plate, and use an ELISA reader to measure the absorbance value of the reaction solution at a wavelength of 492nm. Make three groups for each sample, and three parallels in each group.

[0068] Table 7 is the composition of the reaction solution

[0069]

[0070] Note: A—absorbance value of blank group with enzyme system; B—absorbance value of blank group without enzyme system; C—absorbance value of sample group with enzyme system; D—absorbance value of sample group without enzyme system.

[0071] Test samples (positive control solution and test sample solution).

[0072] The inhibition rate of the test sample on tyrosinase was calculated according to the following tyrosinase inhibition rate calculation formula.

[0073]

[0074] Figure 6 This is a virtual screening of the inhibition rate of compound I and the positive control α-arbutin solution on tyrosinase at different concentrations. Figure 6 The concentration of the test sample is the abscissa, and the corresponding tyrosinase inhibition rate is the ordinate. The following four-parameter logistic equation is used to fit the test sample inhibition curve.

[0075]

[0076] Where: min is the point with the smallest response; max is the point with the largest response;

[0077] HC is the slope of the curve at the inflection point; x is the sample concentration; y is the inhibition rate.

[0078] According to the fitted regression equation, the IC of virtual screening compound I and positive control α-arbutin for inhibiting tyrosinase was calculated. 50 The IC value of compound I for inhibiting tyrosinase was calculated. 50 The IC value of α-arbutin for inhibiting tyrosinase is 0.18 mg / L. 50 The value was 104.48 mg / L. Compared with the positive control α-arbutin, the virtually screened compound I showed a significant inhibitory effect in the in vitro tyrosinase inhibition test, and the IC 50 The value is much smaller than that of α-arbutin, and it has the potential to become a cosmetic whitening raw material.

[0079] The technical features disclosed above are not limited to the disclosed combinations with other features. Those skilled in the art may also make other combinations between the technical features according to the purpose of the invention to achieve the purpose of the invention.

Claims

1. A virtual screening method for tyrosinase inhibitors, characterized in that: The steps include: S1. Building a data set: Building a data set based on the biological activity data of tyrosinase inhibitors. The method for obtaining the biological activity data includes the following steps: (1) Collect tyrosinase inhibitors with known activity and their IC from the ChEMBL database and literature 50 value; (2) Remove non-covalent salt compounds and a small number of silicon- and halogen-containing compounds from the data set; (3) Select compounds with molecular weights in the range of 100-500 and retain IC 50 The duplicate compound with the smallest value; (4) Use the Rdkit Structure Normalizer node in KNIME to delete compounds with unclear structures; S2. Construct QSAR classification model: divide the data set obtained in S1 into training set and test set by random stratified sampling; calculate molecular fingerprint for each inhibitor compound; optimize molecular fingerprint by feature selection method in training set; train QSAR classification model by machine learning classification algorithm; optimize hyperparameters by grid search method with 5-fold stratified cross validation; select the corresponding model obtained by configuring machine learning regression algorithm with hyperparameters with the highest average F1 score as the final QSAR classification model; S3, build QSAR regression model: use Butina clustering algorithm to split the data set obtained in S1 into training set and external test set; calculate the molecular descriptors and / or molecular fingerprints of the compounds; train QSAR regression model by machine learning regression algorithm; use grid search method with 5-fold stratified cross validation for hyperparameter optimization; select average R 2 The corresponding model obtained by configuring the machine learning regression algorithm with the highest hyperparameters is the final QSAR regression model; S4. Virtual screening of tyrosinase inhibitors: First, based on the molecular characteristics of the highly active compounds in the dataset, candidate compounds were screened from the ZINC20 database. Then, the QSAR classification model was used for preliminary screening. Finally, the QSAR regression model was used to predict the IC 50 value.

2. The virtual screening method for tyrosinase inhibitors according to claim 1, characterized in that: The calculated molecular fingerprint types described in S2 and S3 include: CDK, CDK-Extend, MACCS, Substructure, Pubchem, Klekota-Roth, ECFP and FCFP; the radius of the ECFP molecular fingerprint can be selected as 1, 2, or 3; the number of bits can be selected as 512, 1024, 2048, 4906, or 8192.

3. The virtual screening method for tyrosinase inhibitors according to claim 1, characterized in that: In step S3, the process of optimizing the molecular descriptors and / or molecular fingerprints in the training set by using a feature selection method is also included.

4. The virtual screening method for tyrosinase inhibitors according to claim 3, characterized in that: The feature selection method for optimizing molecular fingerprints is one or a combination of zero variance filtering, chi-square test, mutual information, Kendall rank correlation coefficient, and Point-biserial correlation coefficient; the feature selection method for optimizing molecular descriptors is one or a combination of low variance filtering, Spearman rank correlation coefficient, Pearson correlation coefficient, and embedding method.

5. The virtual screening method for tyrosinase inhibitors according to claim 1, characterized in that: The machine learning classification algorithms described in S2 are: logistic regression (LR), ridge regression (RR), passive attack (PA), support vector machine (SVM), linear support vector machine (LSVM), random forest (RF), extreme random trees (ERT), extreme gradient boosting (XGB), lightweight gradient boosting machine (LGBM).

6. The virtual screening method for tyrosinase inhibitors according to claim 1, characterized in that the hyperparameters of the hyperparameter optimization described in S2 are: LR, PA, and LSVM optimize the regularization parameter C, RR optimizes the regularization parameter alpha, SVM selects RBF kernel coefficient gamma and regularization parameter C for optimization, RF and ERT optimize the number of estimators n_estimators, the maximum depth of the tree max_depth, and the minimum number of samples of internal nodes of the branch min_samples_split, XGB optimizes n_estimators, max_depth, min_child_weight, and learning_rate, and LGBM optimizes num_leaves, n_estimators, max_depth, min_child_samples, and learning_rate.

7. The virtual screening method for tyrosinase inhibitors according to claim 1, characterized in that: The machine learning regression algorithms described in S3 are: support vector machine (SVM), random forest (RF), extreme random trees (ERT), extreme gradient boosting (XGB), and lightweight gradient boosting machine (LGBM).

8. The virtual screening method for tyrosinase inhibitors according to claim 1, characterized in that: The hyperparameters of hyperparameter optimization described in S3 are: SVM selects RBF kernel, optimizes gamma, C, and epsilon, RF and ERT optimize the number of estimators n_estimators, the maximum depth of the tree max_depth, and the minimum number of samples of internal nodes of the branch min_samples_split, XGB optimizes n_estimators, max_depth, min_child_weight, and learning_rate, and LGBM optimizes num_leaves, n_estimators, max_depth, min_child_samples, and learning_rate; all hyperparameter optimization processes follow the principle of rough search first and then fine search.

9. The virtual screening method for tyrosinase inhibitors according to claim 1, characterized in that The molecular characteristics described in S4 are: hydrogen bond donor, hydrogen bond acceptor, LogP, and molecular weight numerical range.

10. A composition with whitening function, characterized in that: The composition comprises at least one of the following compounds of formula I, formula II and formula III, and the following compounds of formula I, formula II and formula III are obtained by virtual screening of any one of the tyrosinase inhibitors of claims 1-9:

Images