Method for predicting toxicological properties of cosmetic raw materials

Abstract

The invention discloses a method for predicting toxicological properties of cosmetic raw materials. The invention discloses an organic compound toxicity prediction method and model based on molecular structure characteristics and a machine learning algorithm, and aims to predict four key toxicological properties of cosmetic raw materials, namely mutagenicity, carcinogenicity, eye irritation and skin sensitization. The method comprises the following steps: firstly, collecting structural data and four toxicological property data of a compound, and constructing a toxicity data set; then, dividing the data set into a training set and an external test set, calculating molecular feature fingerprints of the data set, and screening the molecular fingerprints of the training set by using a feature selection method; a toxicity classification prediction model is established by using a machine learning algorithm and training set data, and the prediction accuracy and reliability of the model are verified through external test set data. According to the method, the safety risk of cosmetic raw material research and development can be effectively reduced, the cost of toxicology testing is reduced, and meanwhile guiding significance and practical value are provided for safety risk assessment of cosmetic raw materials.

Description

Technical Field

[0001] The present invention belongs to the field of chemical informatics, and in particular relates to a model and method for predicting the toxicity of cosmetic raw materials. Background Art

[0002] The safety of cosmetic raw materials has always been the focus of attention both inside and outside the industry. The evaluation of the toxicological properties of cosmetic raw materials is a key link in ensuring product safety, especially the prediction and evaluation of toxicity of organic compounds used in cosmetics. Traditional toxicological tests mainly rely on in vitro experiments and animal experiments. Although they can provide accurate safety data, they are costly, time-consuming, and face challenges in animal ethics. With the continuous improvement of regulations and the increasing attention of consumers to the safety of cosmetics, more and more countries and regions have begun to restrict or prohibit animal testing of cosmetic ingredients, which makes the search for alternative toxicity testing methods an urgent problem to be solved. Therefore, it is particularly important to develop a fast, accurate and cost-effective toxicity prediction method. In recent years, computational toxicology has received widespread attention as a tool to replace or supplement experimental toxicology. In particular, machine learning algorithms, with their powerful data processing and pattern recognition capabilities, have shown great potential in drug development, environmental risk assessment and other fields. The molecular structure-based toxicity prediction method relies on modern computational chemistry and molecular simulation technology. By constructing the molecular fingerprint characteristics of the compound and using the existing toxicity data set to establish a prediction model, it can achieve rapid prediction of the toxic properties of chemicals. This method can not only reduce the use of animal experiments, but also significantly reduce the time and cost of toxicology testing, providing important support for safety screening in the cosmetics research and development process.

[0003] In the field of cosmetic raw materials, mutagenicity, carcinogenicity, eye irritation and skin sensitization are four key toxicological properties that are directly related to the health and safety of consumers. The evaluation of these properties is crucial for the screening of cosmetic raw materials, formulation design and safety evaluation of final products. Therefore, the present invention aims to achieve a method for accurately predicting multiple toxicological properties by constructing a toxicity prediction model based on molecular structure characteristics and machine learning algorithms. This method can not only improve the efficiency of cosmetic raw material research and development and reduce safety risks, but also provide scientific basis and technical support for safety supervision in the cosmetics industry. Summary of the invention

[0004] In order to solve the above problems existing in the research and development of cosmetic raw materials, the present invention proposes a method for predicting the toxicological properties of cosmetic raw materials, which is used to evaluate four toxicological properties of cosmetic raw materials: mutagenicity, carcinogenicity, eye irritation and skin sensitization.

[0005] The specific technical process includes:

[0006] Step 1: Collect the molecular structure data of the compound and its corresponding four toxicological property data from public toxicity databases, literature reports, etc. Organize these data to form a toxicity data set for the compound. The data set size is shown in the following table:

[0007] Toxicological properties Number of compounds Mutagenicity 6782 Carcinogenicity 1296 Eye irritation 1137 Skin sensitization 709

[0008] Note: Positive compounds indicate that the compounds are mutagenic, carcinogenic, eye irritating, and skin sensitizing; negative compounds indicate that they are not mutagenic, carcinogenic, eye irritating, and skin sensitizing.

[0009] Step 2: Use random stratified sampling to divide the collated toxicity dataset into a training set and an external test set. The training set is used for model training and optimization, while the external test set is used to evaluate the prediction accuracy and reliability of the toxicity model to ensure that the model has good generalization ability. The random stratified method of splitting the dataset ensures that the ratio of positive compounds to hidden compounds in the training set and the external test set is consistent.

[0010] Step 3: Use the canonical SMILES of the compound to calculate the molecular fingerprint of the compound, specifically, including the following eight types of molecular fingerprints: CDK, CDK-Extend, MACCS, Substructure, Pubchem, Klekota-Roth calculated by Padel-descriptor software. ECFP and FCFP are calculated using the AllChem.GetMorganFingerprintAsBitVect module of the Rdkit package. As a further limitation of the present invention, the radius of ECFP can be selected from 1, 2, 3; the number of bits can be selected from 512, 1024, 2048, 4906, 8192, 16384.

[0011] The molecular fingerprint of the toxicity prediction model is determined by using a feature selection method on the training set, wherein the feature selection method is one or a combination of zero variance filtering, chi-square (χ2) test, and mutual information. It is generally believed that molecular fingerprints with zero variance are meaningless for modeling, that is, when a certain segment of the molecular fingerprint is all constant 0 or 1, the segment is filtered out; then the chi-square test or mutual information is continued to be used to further reduce the fingerprint features used for modeling.

[0012] Step 4: Toxicity prediction models for mutagenicity, carcinogenicity, eye irritation and skin sensitization were established on the training set data that had completed molecular fingerprint screening, and the models were constructed using the following machine learning classification algorithms: 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), and lightweight gradient boosting machine (LGBM).

[0013] When training the toxicity prediction model, a 5-fold stratified cross-validation grid search method was used for hyperparameter optimization, and the hyperparameter configuration machine learning classification algorithm with the highest average F1 score of the model on the cross-validation set was selected.

[0014] The hyperparameters that need to be optimized when modeling the above machine learning classification algorithms 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 tree depth max_depth, and the minimum number of samples of internal nodes in 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.

[0015] As a further limitation of the present invention, due to the imbalance in the number of samples of the two types of activity data in the data set, a weighted cost-sensitive learning strategy is adopted when constructing the toxicity prediction model. Specifically, the setting of adjusting the weight ratio refers to the following calculation formula:

[0016]

[0017] In the formula represents the total number of training set samples, 2 represents binary classification, Represents the training set category The number of samples.

[0018] Step 5: Use an external test set to verify the accuracy of the toxicity prediction model. The following five indicators are used to evaluate the performance of the toxicity prediction model: recall (REC), precision (PRE), specificity (SPE), balanced-accuracy (BA), and F1 score. When the classes are unbalanced, the precision-recall (PR) curve and the area under the PR curve (PR-AUC) can better reflect the true prediction ability of the classification model. The present invention uses the PR curve and PR-AUC to visually evaluate the model.

[0019]

[0020]

[0021]

[0022]

[0023]

[0024] In the formula, TP (True Positive) represents the number of compounds correctly predicted as positive, TN (True Negative) represents the number of compounds correctly predicted as negative, FP (False Positive) represents the number of compounds incorrectly predicted as positive, and FN (False Negative) represents the number of compounds incorrectly predicted as negative.

[0025] Step 6: Based on the results of the toxicity prediction model of the external test set, select toxicity prediction models with excellent prediction performance to form a toxicity prediction consensus model. Applying the verified toxicity prediction consensus model to the prediction of the toxicological properties of cosmetic raw materials can quickly screen the safety of new raw materials in the early stages of product development and reduce potential toxicity risks. The prediction model of the present invention is capable of toxicity classification at a high level of accuracy. This method can not only significantly reduce the safety risks of cosmetic raw material research and development, reduce the cost and time of toxicological testing, but also provide a reliable safety assessment tool for R&D personnel in the cosmetics industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 Flow chart of model construction in Example 1 of the present invention.

[0027] Figure 2 Schematic diagram of hyperparameter optimization using RF algorithm modeling in Example 1.

[0028] Figure 3 This is the precision-recall curve of prediction model 2 in Example 1. DETAILED DESCRIPTION

[0029] 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.

[0030] Embodiment 1:

[0031] A classification prediction model for mutagenicity is provided, and a method for constructing the classification model comprises the following steps:

[0032] Constructing a mutagenicity dataset: removing salt compounds with non-covalent structures from the dataset; removing compounds with elements other than C, H, O, N, P, S, Se, F, Cl, Br, I, and Si. After processing, a mutagenicity dataset containing 6782 compounds was constructed, of which 3715 compounds were mutagenic and 3067 were non-mutagenic. Then, the SMILES of the dataset compounds were normalized using the Ridkit package and all converted to standard SMILES representations.

[0033] Splitting the data set: According to the random stratified sampling method, the train_test_split module of the sklearn library was used to split the mutagenicity data set into a training set and an external test set. The training set contained 6103 compounds and the test set contained 679 compounds.

[0034] Calculate and select molecular fingerprints: Calculate the molecular fingerprint ECFP of the compounds in the data set. When calculating ECFP, select the molecular segmentation radius as 2, and select the number of characteristic bits of the molecular fingerprint as 512, 1024, 2048, 4906, 8192, and 16384. Then use the zero variance filtering and chi-square (χ^2) test method to select molecular fingerprints, with a p value less than 0.05 as the limit, and filter out fingerprint fragments with low χ^2 values.

[0035] Machine learning classification algorithm modeling: LR, RR, PA, SVM, LSVM, RF, ERT, XGB, and LGBM algorithms are used to build classification prediction models on the feature-screened training set data. The hyperparameter optimization process is explained by using the RF algorithm modeling as an example. By setting a wider parameter range, it is easy to see the impact of parameter value changes on the RF algorithm model training set and validation set. It is found that n_estimators, max_depth, and min_samples_split all have the highest average F1 score of the validation set in a certain range ( Figure 1 ). Then, the three hyperparameters with narrower intervals were combined into a hyperparameter space for a 5-fold stratified cross-validation grid search to determine the optimal hyperparameters for configuring the classification model. Then, according to the weight adjustment formula, the weight of the positive compound was calculated to be 0.913, and the weight of the negative compound was 1.106. The weight ratio of the two types of compounds was set to 0.913:1.106. Finally, the optimized hyperparameters and weights were configured to complete the construction of the compound mutagenicity classification model.

[0036] Evaluation of the predictive performance of the mutagenicity classification model: The constructed mutagenicity classification model was used to predict the mutagenicity of the external test set compounds. The performance of the mutagenicity prediction model was evaluated using the REC, PRE, SPE, BA, and F1 score indicators. Four models with good predictive performance were selected to form the mutagenicity consensus model. The indicators of the mutagenicity consensus model in the training set and external tests are shown in the following table.

[0037] PR curve and PR-AUC of mutagenic compounds predicted by mutagenicity classification model ( Figure 3 ), showing that the model has a high recognition rate and accuracy for mutagenic compounds. The mutagenicity consensus model established by the present invention helps to exclude compounds or raw materials with mutagenic risks during raw material development or cosmetic development.

[0038] Consensus Model REC PRE SPE BA F1 Prediction Model 1 0.879 0.847 0.808 0.843 0.863 Prediction Model 2 0.858 0.862 0.834 0.846 0.860 Prediction Model 3 0.863 0.854 0.821 0.842 0.858 Prediction Model 4 0.852 0.854 0.824 0.838 0.853

[0039] 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 method for predicting the toxicological properties of cosmetic raw materials, characterized in that: The steps include: Step 1: Organize and construct the mutagenicity, carcinogenicity, eye irritation and skin sensitization data sets; Step 2: The data set is divided into a modeling training set and an external test set; Step 3: Calculate and screen molecular fingerprint features; Step 4: The machine learning algorithm constructs a classification prediction model for the four toxicological properties; Step 5: Evaluate the performance of the prediction models and select models to form a consensus prediction model.

2. The method for predicting toxicological properties of cosmetic raw materials according to claim 1, characterized in that: The step 3 specifically includes: Step 301: using the canonical SMILES of the compound to calculate the molecular fingerprint of the compound, including the following eight molecular fingerprint algorithms: CDK, CDK-Extend, MACCS, Substructure, Pubchem, Klekota-Roth, ECFP and FCFP; Step 302: Apply a feature selection method on the training set to determine the best molecular fingerprint feature set for the toxicity prediction model.

3. The method for predicting toxicological properties of cosmetic raw materials according to claim 2, characterized in that: In step 301 , the radius of the ECFP molecular fingerprint algorithm can be 1, 2, or 3; the number of bits can be 512, 1024, 2048, 4906, or 8192.

4. The method for predicting toxicological properties of cosmetic raw materials according to claim 2, characterized in that: The feature selection method in step 302 is one or a combination of zero variance filtering, chi-square test, and mutual information.

5. The method for predicting toxicological properties of cosmetic raw materials according to claim 1, characterized in that: In step 4, a weighted cost-sensitive learning strategy is used to construct the toxicity prediction model. Specifically, the calculation formula for adjusting the weight ratio is: In the formula represents the total number of training set samples, 2 represents binary classification, Represents the training set category The number of samples.

Images