A prediction method and prediction model for the carcinogenicity of PAHs based on 2D molecular descriptors

Abstract

The present invention discloses a method and a prediction model for the carcinogenicity of PAHs based on 2D molecular descriptors. According to the OECD principles, the present invention performs chemometric QSAR modeling based on simple 2D molecular descriptors to predict the carcinogenicity of PAHs to rodents. Six QSAR models for the carcinogenicity of PAHs for female rats, male rats, female mice, male mice, and rats and mice were established and verified using multiple statistical validation criteria. Mechanism analysis clearly shows the correlation between descriptors and carcinogenicity, and some descriptors that coexist in different models indicate that similar mechanisms of action may exist. In addition, the present invention also establishes two iQCCR models to fill in interspecies data gaps. The present invention uses QSAR models and iQCCR models to predict the carcinogenicity of hundreds of untested PAHs, evaluates the prediction quality of each molecule, and can screen out priority compounds based on the carcinogenic potency of PAHs, so it is of great significance for risk assessment under the regulatory framework.

Description

Technical Field

[0001] The invention relates to a PAHs carcinogenicity prediction method and prediction model based on 2D molecular descriptors, belonging to the technical field of compound carcinogenicity prediction. Background Art

[0002] Cancer is one of the leading causes of death in countries around the world and a key factor in extending life expectancy. According to the World Health Organization's "2020 Global Health Assessment", among 185 countries, cancer is the first or second leading cause of death among people under 70 years old in 112 countries, and the third or fourth leading cause of death among people under 70 years old in 23 countries. In fact, many cancer deaths can be attributed to potential modifiable risk factors. If carcinogens and carcinogenic mechanisms can be fully understood and effective prevention strategies can be made, half of the tumors can be avoided. Among these risk factors, exposure to exogenous chemicals is one of the most important factors. IARC (International Agency for Research on Cancer) lists many environmental factors that are carcinogenic to humans, including many organic compounds that are considered carcinogens. Chemical carcinogens refer to compounds that can cause tumors in animals and humans and increase their morbidity or mortality.

[0003] Polycyclic aromatic hydrocarbons (PAHs) are important environmental organic pollutants that usually contain two or more benzene rings (fused and non-fused) and have received widespread attention because they are the largest category of chemical carcinogens. For example, IARC has identified several PAHs as Class 1 (e.g., benzo[a]pyrene, benzidine) or Class 2A (e.g., 1-nitropyrene, tetrabromobisphenol A) carcinogens. Some mixtures containing PAHs, such as particulate matter PM2.5 and diesel engine exhaust, are also classified as Class 1 carcinogens by IARC. On the other hand, PAHs are difficult to decompose, so they are widely distributed in the environment (air, water, soil, sediment, and food). In addition to carcinogenicity, other adverse reactions of PAHs have been observed, such as endocrine disruption, acute toxicity, allergic reactions, and skin or eye irritation. In this context, PAHs pose potential hazards to the environment and public health by contaminating air, water, and food.

[0004] Typically, the identification of carcinogens comes from human epidemiological surveys, long-term animal biological analyses, mutagenicity tests, toxicokinetics, and related cancer mechanism studies. Long-term animal bioassays are a common method for determining the carcinogenicity of chemicals, with the ultimate goal of assessing their carcinogenic risk to humans. Affected by a variety of factors, there are few human carcinogenicity data available for chemicals. At this stage, our understanding of rodent carcinogenicity relies solely on data from carcinogenicity studies. However, animal experimental methods used to study carcinogenicity have the following limitations: time-consuming (usually 18 to 36 months), and single standard tests are extremely expensive; animal welfare and ethical issues, which must involve a large number of animals to ensure statistically significant results.

[0005] Given the increasing social and economic pressure to reduce animal testing and support the 3R principle (reduce, replace, optimize), large-scale animal testing is not suitable for the identification of the carcinogenicity of a large number of chemicals. Therefore, computer simulation methods such as quantitative structure-activity relationship (QSAR) are regarded as potential alternative methods for toxicity prediction. QSAR prediction can be used for screening purposes to prioritize chemicals in the early stages of drug development or even before chemical synthesis. Especially in the pharmaceutical industry, QSAR can be used to preliminarily screen out candidate drugs with carcinogenicity, which can avoid incalculable cost losses in the later stage. At present, some international legislation, institutions and organizations support, recognize and recommend the use of QSAR methods for toxicity prediction under the regulatory framework, such as the EU REACH (Registration, Evaluation, Authorization and Restriction of Chemicals) legislation, ICH (International Council for Harmonization of Technical Requirements for Registration of Pharmaceuticals for Human Use), M7 guidelines, the US Food and Drug Administration (FDA), the US Environmental Protection Agency (EPA), the European Center for the Validation of Alternative Methods (ECVAM), the Agency for Toxic Substances and Disease Registry (ATSDR), the European Agency for the Evaluation of Medicines (EU-EMEA), the European Chemicals Agency (ECHA), the Pharmaceuticals and Medical Devices Agency (Japan-PMDA), the Australian Environment Agency (AEA), etc. In order to increase the acceptance of QSAR models, the OECD (Organization for Economic Cooperation and Development) has specifically formulated five principles for QSAR validation for regulatory use, including: well-defined endpoints; clear algorithms; clear scope of application; appropriate fit, robustness and predictive ability; and, if possible, an explanation of the toxicity mechanism.

[0006] To date, reports on predictive models for the carcinogenicity of chemicals to rodents can be divided into three categories: qualitative structure-activity relationship (SAR) models, which are usually binary models (predicting chemicals as carcinogens or non-carcinogens); QSAR models (predicting as continuous values); and expert systems. Several QSAR models have been established for non-similar chemicals, but their predictive capabilities are generally weaker than those of QSAR models for similar chemicals and lack mechanistic explanations. In addition, although some QSAR prediction models for similar chemicals have been established, such as nitro, aromatic amines, N-nitrosamines, and aromatic nitrogen compounds, no QSAR models for the carcinogenicity of PAHs to rodents have been reported, and the few available models are classification models. Summary of the invention

[0007] The purpose of the present invention is to provide a method and a prediction model for the carcinogenicity of PAHs based on 2D molecular descriptors. The present invention is based on simple 2D molecular descriptors and performs chemometric QSAR modeling to predict the carcinogenicity of PAHs to rodents. It can be used to predict the carcinogenicity of new or untested PAHs within its application domain.

[0008] Based on rodents, the present invention provides six different carcinogenicity QSAR models. In each model, pTD 50 Indicates carcinogenic potency, which is TD 50 The values ​​were converted to molar units (mol / kg / d) and then transformed into negative logarithms, which were positively correlated with the carcinogenic potency; the remaining parameters represent the 2D molecular descriptors of PAHs.

[0009] The first model is based on the carcinogenicity QSAR model (RF) obtained in female rats, as shown in formula (1):

[0010] TD 50 =4.8314+5.1819MATS5m+7.2761MATS6v-0.7063minsCH3-0.0866nAtomP+3.3684nFG12Ring (1)

[0011] The second model is based on the carcinogenicity QSAR model (RM) obtained in male rats, as shown in formula (2):

[0012] TD 50 =9.7694–3.3927J_D+0.0194ATSC3s-0.5947GATS5s-4.4007ATSc4-0.8934SHaaCH+1.6534minaaCH (2)

[0013] The third model is the carcinogenicity QSAR model (MF) based on female mice, as shown in formula (3):

[0014] TD 50 =58.8706-3.4972J_D-9.3578SpMaxA_B(s)+0.3697GATS8s-12.5216SpMax2_Bh(v)+0.9240Cl-089-1.7211minaaaC (3)

[0015] The fourth model is the carcinogenicity QSAR model (MM) based on male mice, as shown in formula (4):

[0016] TD 50 =–2.7777+0.7292CIC4-5.3604MATS4s+1.5330GATS7e+0.0085P_VSA_MR_7-1.2361B10[CC]+17.2964VCH-6+1.1910maxwHBa (4)

[0017] The fifth model is based on the rat-derived carcinogenicity QSAR model (R), as shown in formula (5):

[0018] TD 50 =-17.7427-0.6262SHaaCH+4.4089SIC1-4.6565J_D+3.7901MATS5m+25.0407SpMAD_EA-1.8307SM03_EA(bo)-0.5179nArOH (5)

[0019] The sixth model is a carcinogenicity QSAR model (M) obtained based on mice, as shown in formula (6):

[0020] TD 50 =3.9044+1.587MATS6e-1.1674SpMax8_Bh(e)+0.0192P_VSA_LogP_5+0.0092P_VSA_M R_7-8.423SpMaxA_EA(dm)+0.3554CATS2D_07_PL+1.2319minaaCH–0.1702ETA_dBeta (6)

[0021] The QSAR model provided by the present invention has high prediction reliability, such as R 2 ext =0.7102~0.9561, Q 2 F1 =0.6270~0.8972, Q 2F2 =0.6255~0.8955, Q 2 F3 =0.8062~0.7263、CCC test =0.8062~0.9547, etc.

[0022] In order to fill the gaps in interspecies data, the present invention also established two interspecies quantitative carcinogenicity iQCCR models (interspecies quantitative carcinogenicity-carcinogenicity relationship models), as shown in formula (7) and formula (8), respectively:

[0023] TD 50 (Rat) = 0.3553 + 0.8987 pTD 50 (Mouse) (7)

[0024] TD 50 (Mouse)=0.3426+0.8451pTD 50 (Rat) (8)

[0025] Where pTD 50 (Mouse) indicates the carcinogenicity of PAHs to mice (experimental value), pTD 50 (Rat) represents the carcinogenicity of PAHs to rats (experimental value).

[0026] The iQCCR model provided by the present invention also has high prediction reliability, such as R 2 ext =0.7446~0.7990, Q 2 Fn =0.7119~0.9103, and CCC test =0.8305~0.8815.

[0027] The present invention determines the detailed relationship between structural information (descriptors) and carcinogenicity through mechanistic interpretation.

[0028] The model provided by the present invention has the characteristics of high goodness of fit, robustness and strong predictability, and can be used to predict the carcinogenicity of new or untested PAHs within its application domain.

[0029] Based on the prediction model, the present invention also provides a method for predicting the carcinogenicity of PAHs, comprising the following steps:

[0030] Obtain 2D molecular descriptors of PAHs or their carcinogenicity data pTD for either rat or mouse 50 According to the carcinogenicity QSAR model or the carcinogenicity iQCCR model, the carcinogenicity of PAHs can be obtained, that is, the carcinogenicity prediction of PAHs can be achieved.

[0031] DRAGON and PaDEL-Descriptor software were used to obtain the 2D molecular descriptors of PAHs. It was verified that the model combining Dragon and PaDEL descriptors had better prediction performance.

[0032] The present invention also provides a PAHs carcinogenicity prediction device using the carcinogenicity QSAR model or the carcinogenicity iQCCR model.

[0033] The present invention further provides a computer device, comprising a processor and a memory storing a computer program; the processor is configured to execute the computer program to implement the method for predicting the carcinogenicity of PAHs according to the present invention.

[0034] The present invention further provides a computer storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the method for predicting the carcinogenicity of PAHs of the present invention is implemented.

[0035] The present invention performs chemometric QSAR modeling based on simple 2D molecular descriptors to predict the carcinogenicity of PAHs to rodents according to OECD principles. Six QSAR models for the carcinogenicity of PAHs in female rats, male rats, female mice, male mice, and rats and mice were established and validated using multiple statistical validation criteria. Mechanistic analysis clearly shows the correlation between descriptors and carcinogenicity, and some descriptors that coexist in different models indicate that similar mechanisms of action may exist. In addition, the present invention also establishes two iQCCR models to fill in interspecies data gaps.

[0036] The present invention utilizes QSAR model and iQCCR model to predict the carcinogenicity of hundreds of untested PAHs, wherein the prediction quality of each molecule is evaluated, and priority compounds can be screened out according to the carcinogenic potency of PAHs, so it is of great significance for risk assessment under the regulatory framework. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Flowchart for building and validating the model for the present invention.

[0038] Figure 2 is the chemical space distribution of compounds in the training set, test set, and real external set in each QSAR model, where Figure 2 (A) represents the female rat QSAR model; Figure 2 (B) represents the male rat QSAR model; Figure 2 (C) represents the female mouse QSAR model; Figure 2 (D) represents the male mouse QSAR model; Figure 2 (E) represents the rat QSAR model; Figure 2 (F) represents the mouse QSAR model; Figure 2 (G) represents the rat-mouse iQCCR model; Figure 2 (H) represents the mouse-rat iQCCR model.

[0039] Figure 3 is the variable importance diagram of six QSAR models, where Figure 3 (A) represents a female rat model, Figure 3 (B) represents the male rat model, Figure 3 (C) represents a female mouse model; Figure 3 (D) represents the male mouse model; Figure 3 (E) represents the rat model; Figure 3 (F) represents the mouse model.

[0040] Figure 4 are the loading diagrams of the six QSAR models, where: Figure 4 (A) represents the female rat model; Figure 4 (B) represents the male rat model; Figure 4 (C) represents a female mouse model; Figure 4 (D) represents the male mouse model; Figure 4 (E) represents the rat model; Figure 4 (F) represents the mouse model.

[0041] Figure 5 Descriptors and mechanistic explanations for the carcinogenicity of PAHs in female rats.

[0042] Figure 6 For the RF model ( Figure 6 (A) and Figure 6 (B))、RM model( Figure 6 (C) and Figure 6 (D)) and R model ( Figure 6 (E) and Figure 6 (F) Linear correlation diagram of predicted values ​​and experimental values ​​and application domain (Williams) diagram.

[0043] Figure 7 Descriptors and possible mechanistic explanations for the carcinogenicity of PAHs in male rats.

[0044] Figure 8 Descriptors and possible mechanistic explanations for the carcinogenicity of PAHs in female mice were provided.

[0045] Fig. 9 For the MF model ( Fig. 9 (A) and Fig. 9 (B)), MM model ( Fig. 9 (C) and Fig. 9(D)) and M model ( Fig. 9 (E) and Fig. 9 (F) Linear correlation plot of predicted values ​​and experimental values ​​and application domain (Williams) plot.

[0046] Fig.10 Descriptors and mechanistic explanations for the carcinogenicity of PAHs in male mice

[0047] Fig.11 Descriptors and mechanistic explanations for the carcinogenicity of PAHs in rats.

[0048] Fig.12 Descriptors and mechanistic explanations for the carcinogenicity of PAHs in mice.

[0049] Fig.13 The linear correlation diagram between the experimental value and the predicted value of the rat-mouse iQCCR model and the mouse-rat iQCCR model ( Fig.13 (A) and (C)), application domains of rat-mouse iQCCR model and mouse-rat iQCCR model (Williams) Fig.13 (B) and (D)).

[0050] Fig.14 Application domain (Insubria) diagrams of six QSAR models (A to F) and two iQCCR models (G and H) to real external set compounds. DETAILED DESCRIPTION

[0051] Unless otherwise specified, the experimental methods used in the following examples are conventional methods.

[0052] Unless otherwise specified, the materials and reagents used in the following examples can be obtained from commercial sources.

[0053] 1. Materials and Methods

[0054] 1. Data collection and organization

[0055] According to OECD Principle 1, “QSAR should have a well-defined endpoint”, and the endpoint of the study is the carcinogenic dose (TD 50 ) indicates the carcinogenicity. 50It represents the drug dose at which half of the test animals develop tumors at the end of the standard lifespan of the test species, in units of mg / kg body weight / day (mg / kg / d). Carcinogenicity data are collected from the Carcinogenic Potency Database (CPDB), which contains the results of 6540 chronic, long-term animal carcinogenicity tests on 1547 chemicals since 1950. The carcinogenicity tests in the CPDB involved rats, mice, hamsters, dogs and non-human primates, covering a total of 35 target organs (e.g., liver, lungs, skin). The present invention uses the liver as the target site, and its tumor type is mixed (more than one tumor type). Rats and mice are used as research targets, and the data are subdivided by gender. After excluding all organic salts, mixtures, polymers and metal complexes, all TD 50 The value (mg / kg / d) was converted to molar units (mol / kg / d) and then negative logarithmic conversion (-logTD 50 or pTD 50 ). 50 Compared with the pTD 50 The value is positively correlated with the carcinogenic potency.

[0056] Regarding the interspecies quantitative carcinogenicity-carcinogenicity relationship (iQCCR) model, 37 rat and mouse TD 50 Finally, 6 QSAR models and 2 iQCCR models were obtained as shown in Table 1. Figure 1 All the detailed information of the PAHs used in the modeling are listed in Table 2.

[0057] Table 1 Summary of PAHs dataset

[0058]

[0059] a N total represents the total number of PAHs in each data set; N tr 、N test and N true Represents the number of PAHs in the training set, test set and true external set in each dataset, respectively.

[0060] b RF, RM, MF, MM, R, and M represent the female rat, male rat, female mouse, male mouse, rat, and mouse datasets, respectively;

[0061] RM and MR represent rat-mouse and mouse-rat iQCCR models, respectively.

[0062] Table 2 Serial numbers, CAS numbers, names and TD values ​​of PAHs in rats and mice 50value

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069] 2. Descriptor calculation and dataset partitioning

[0070] The chemical structures of PAHs were downloaded from the PubChem website (https: / / pubchem.ncbi.nlm.nih.gov / ) and cross-checked through online databases such as Chemical Book (https: / / www.chemicalbook.com) and ChemIDplus (https: / / chem.nlm.nih.gov / chemidplus / ). Before calculating the descriptors, the structure was optimized by the 3D module of ChemBioDrawUltra14.0 software to put the molecules in the lowest energy state. In view of the complexity of the carcinogenic endpoint, the present invention combines the Dragon and PaDEL descriptors calculated by DRAGON and PaDEL-Descriptor software, respectively, to characterize the structural features of the compounds, so as to more fully extract the important structural information in the compounds. In order to facilitate the mechanism interpretation, only 2D descriptors with clear physicochemical significance were calculated for model establishment, and constant or near-constant descriptors and highly correlated descriptors (correlation coefficient>0.98) were deleted to reduce redundant variable information.

[0071] In order to obtain a QSAR model with higher prediction quality, the present invention uses three partitioning techniques to divide each data set into a training set and a test set at a ratio of approximately 3:1, namely (a) sorting by effect value (ORes), (b) sorting by structural similarity (OStr), and (c) random (Rnd). The test set that did not participate in the model establishment was used for external validation of the model to avoid possible bias. When partitioning by the ORes method, the compounds were sorted according to the carcinogenic potency pTD 50 The compounds are sorted in ascending order of pTD. For every 4 compounds, 3 are placed in the training set and 1 is placed in the test set. 50 Value and minimum pTD 50Compounds with ORes values ​​were always placed in the training set to cover the entire range of effects; for OStr partitioning, compounds were sorted by the first axis principal component (PC1) scores of all descriptors, and 1 of every 4 compounds was placed in the test set, with the compounds with the largest and smallest PC1 scores placed in the training set. For iQCCR modeling, only the ORes partitioning method was used, and about 25% of the compounds were set as the test set. Data set partitioning and model building were performed in QSARINSv2.2.4 software to ensure that the final model was robust and externally predictive.

[0072] 3. Descriptor selection and model building

[0073] According to OECD Principle 2, “QSAR models should have clear algorithms”. The genetic algorithm-variable subset selection (GA-VSS) module included in the QSARINS software was used to screen variables, and the ordinary least squares (OLS) method combined with multivariate linear regression (MLR) was used to establish the QSAR model. In the descriptor selection process, the leave-one-out cross-validation correlation coefficient (Q 2 LOO ) is set as the fitness function as a measure of model robustness. The QUIK (Q Under Influence of K) rule is applied to avoid multicollinearity with a threshold of 0.05. This rule is based on the K multivariate correlation index, assuming that the descriptor X plus the effect Y (K XY ) should always be higher than the correlation between descriptors (K XX ). The ratio of the number of training set compounds to the number of selected descriptors should be at least greater than 5, in accordance with the rule of thumb (also known as the Topliss ratio) to avoid the possibility of accidental correlation. The values ​​of gene generation, population size, and mutation rate in each model building process were set to 2000, 200, and 50, respectively. For the iQCCR model, the carcinogenicity of one species was used as the only variable to predict the carcinogenicity of the other species, and a rat (dependent variable)-mouse (independent variable) or mouse (dependent variable)-rat (independent variable) iQCCR model was constructed using a simple linear regression method.

[0074] 4. Model Performance Evaluation

[0075] According to OECD Principle 4, “QSAR models should have appropriate goodness of fit, robustness and predictive power”. The present invention uses multiple state-of-the-art internal and external validation indicators to evaluate the predictive performance of the established model. Model robustness and goodness of fit are measured by Q 2 LOO and the coefficient of determination R 2 (including R 2 adj) were evaluated, and the Y randomization test was repeated for 2000 times (where pTD 50 The values ​​are randomly arranged, while the descriptor matrix remains unchanged) is used to test the reliability of the model.

[0076] By calculating different statistical parameters, such as Q 2 F1 , Q 2 F2 , Q 2 F3 , R 2 ext 、r 2 m(test) , Δr 2 m(test) And the consistency correlation coefficient (CCC) of the test set test ) for external validation. In addition, the root mean square error (RMSE) was used to evaluate the prediction accuracy of the model. The prediction quality of the model was evaluated as "good", "medium" and "poor" based on the mean absolute error (MAE). This step uses the "Xternal Validation Plus" tool to classify the quality of the model (https: / / dtclab.webs.com / software-tools). All validation parameters used in the present invention are listed in Table 3.

[0077] Table 3 Statistical parameters for internal and external validation

[0078]

[0079]

[0080] 5. Application domain analysis

[0081] According to OECD Principle 3, "QSAR models should have a well-defined domain of application". Model-based predictions are reliable only when the compound is within the model's domain of application (AD). The present invention uses a leverage method combined with a standardized method to define AD. If the leverage value h of a compound is greater than the threshold h*, it will be identified as a structural outlier. The threshold h* is defined as 3(p+1) / n, where p is the number of descriptors in the model and n is the number of compounds in the training set. Data that is higher than 3 times the standardized residual is considered an effect outlier. Structural outliers and effect outliers can be visualized by constructing a Williams plot (h value vs standardized residual for each compound). However, for some compounds, even if h>h*, it is not considered an outlier because its standardized residual may be very small.

[0082] Insubria diagram (prediction value of training set, test set or real external set compound vs its h value) is used to visualize the degree of extrapolation of prediction, i.e., prediction reliability, and plays an important role in evaluating the prediction quality of real external set compounds (compounds without specific experimental values). In the present invention, real external set compounds are collected for each QSAR model and iQCCR model. If the h of the real external set compound is > h*, its prediction result has low confidence and the prediction is unreliable.

[0083] 6. Select the best QSAR model using MCDM

[0084] Based on internal and external validation, model performance was ranked using a multi-criteria decision making (MCDM) procedure in the QSARINS software with a score ranging from 0 (worst) to 1 (best). fit Based on R 2 , R 2 adj 、CCC tr The maximum value of R 2 -R 2 adj The minimum value of MCDM is obtained ext Based on Q 2 F1 , Q 2 F2 , Q 2 F3 and CCC test The maximum value of fit and MCDM ext The best QSAR models were selected and these models complied with OECD principles and various validation criteria.

[0085] 7. Screening of untested compounds using QSAR and iQCCR models

[0086] For the RF, RM, MF, MM, R, M, RM and MR models, the number of compounds in the true external set was 142, 132, 132, 129, 124, 120, 18 and 14, respectively (Table 1). The “Prediction Reliability Index” tool (http: / / dtclab.webs.com / software-tools) was used to check the prediction confidence of each true external set compound and evaluate its prediction quality as “good”, “moderate” and “bad”, with a comprehensive score of 3, 2 and 1, respectively.

[0087] 2. Results and Discussion

[0088] 1. Chemical diversity analysis

[0089] Generally speaking, the higher the structural diversity of compounds in the data set, the wider the application range of the model. As shown in Table 1, the data distribution of the best performing QSAR models is as follows: RF model (N tr =25, N test =8,N true =142), RM model (N tr =33, N test =10, N true =132), MF model (N tr =33, N test =10, N true =132), MM model (N tr =36, N test =10, N true =129), R model (N tr =40, N test =11,N true =124) and M model (N tr =44), N test =11,N true =120). For the iQCCR model, the training set, test set, and true external compounds in the rat-mouse and mouse-rat iQCCR models were 28, 9, 18 and 29, 8, 4, respectively. The molecular weight (MW) vs octanol-water partition coefficient ((A)logP) of the compounds was used to characterize the chemical space distribution of each data set. Figure 2 As shown, all test set and true external set compounds share similar chemical space with the training set compounds, indicating that the model established with the corresponding training set we used has a wide range of AD.

[0090] Figure 2 The chemical space distribution of compounds in the training set, test set, and true external set in each QSAR model: (A) female rat QSAR model; (B) male rat QSAR model; (C) female mouse QSAR model; (D) male mouse QSAR model; (E) rat QSAR model; (F) mouse QSAR model; (G) rat-mouse iQCCR model; (H) mouse-rat iQCCR model.

[0091] 2. QSAR model of PAHs carcinogenicity

[0092] According to the OECD QSAR validation guidelines, QSAR modeling was performed on PAHs for six carcinogenic endpoints. Initially, three types of QSAR models were established based on different descriptors, namely, a model based on the Dragon descriptor, a model based on the PaDEL descriptor, and a model based on the combination of Dragon and PaDEL descriptors. As expected, the model combining the Dragon and PaDEL descriptors has better prediction performance. From the statistical parameters, the PAHs carcinogenicity QSAR model established in the present invention has the characteristics of high goodness of fit, strong robustness and predictiveness, and can be used to fill the gaps in carcinogenicity data. The variance of the training set that the model can explain (R 2 ) ranges from 0.6917 to 0.8170 (Q 2 LOO 0.5 to 0.7155), the test set variance (R 2 ext ) ranged from 0.7102 to 0.9561, indicating that the model has strong predictive power. The Y randomization test gave a lower Q 2 Yscr (–0.3254 to –0.4705) and R 2 Yscr The values ​​of (0.1783-0.2047) indicate that the model is not generated by chance. In addition, various external validation indicators of the test set are satisfactory (e.g., Q 2 F1 =0.6270~0.8972, Q 2 F2 =0.6255~0.8955, Q 2 F3 =0.8062~0.9263,CCC test =0.7754~0.902, RSME test =0.4002~0.7736, MAE test =0.3488~0.7123). The detailed statistical results are listed in Table 4.

[0093] Table 4 Statistical parameters of the QSAR model for carcinogenicity of PAHs

[0094]

[0095] 2.1. Female rat (RF) QSAR model

[0096] The RF model obtained based on Rnd partition (N tr =25, N test =8) has the best effect. As shown in equation (1), the model can explain 81.7% of the variance of the training set (Q 2LOO 69.3%), and 95.61% of the test set (R 2 ext ) variance. Based on the MAE evaluation criterion, the prediction quality of the model was evaluated as “good”. A total of five descriptors were selected in the model, namely MATS5m, MATS6v, minsCH3, nAtomP, and nFG12Ring.

[0097] TD 50 =4.8314+5.1819MATS5m+7.2761MATS6v–0.7063minsCH3–0.0866nAtomP+3.3684nFG12Ring (1)

[0098] N tr =25, N test =8, R 2 =0.8170, Q 2 LOO =0.6930, R 2 ext =0.9561, Q 2 F1 =0.8972, Q 2 F2 =0.8955, Q 2 F3 =0.9263, CCC test =0.9547, RMSE test =0.4747, MAE test =0.3659, r 2 m(test) =0.9334, Δr 2 m(test) =0.0384, TSR=8.8027, MAE 95％test +3*SD 95％test =1.0781, Quality Test =Good

[0099] According to the variable importance plot (VIP) ( Figure 3 (A)), the most important descriptors in the model (VIP score ≥ 1.0) are nFG12Ring, MATS6v, and MATS5m. The secondary descriptors (VIP score < 1.0) are nAtomP and minsCH3. Similarly, from the loading graph ( Figure 4 (A)) It can be seen that nFG12Ring, MATS6v and MATS5m contribute the most to the carcinogenicity of female rats because they are close to the predicted effect value and far away from the coordinate origin, while nAtomP and minsCH3 are the opposite.

[0100] As stated in OECD Principle 5, “When possible, explain the mechanism of toxicity in the QSAR model”. The descriptors in the RF model are divided into two categories. The first category of descriptors are positively correlated with carcinogenicity, such as nFG12Ring, MATS5m and MATS6v. By definition, they obviously contribute to increasing the lipophilicity (hydrophobicity) of the molecule. Therefore, higher values ​​of these descriptors mean stronger carcinogenicity, and vice versa. The other category of descriptors are negatively correlated with carcinogenicity, such as minsCH3 and nAtomP, which impart polar characteristics. Figure 5 As shown in (A), compound 59 (2,3,7,8-tetrachlorodibenzo-p-dioxin) with a high nFG12Ring value has a high carcinogenic potency (pTD 50 =9.40), compound 6 with relatively high MATS6v and MATS5m values ​​(MATS6v = 0.145, pTD 50 =4.78) and compound 49 (MATS5m=0.258, pTD 50 =4.00) also had high carcinogenicity. In contrast, compound 54 (quercetin) with high nAtomP (22) and compound 63 (disperse yellow 3) with high minsCH3 (1.51) values ​​showed low carcinogenic potency (pTD 50 =0.60 and 0.76)( Figure 5 (B)).

[0101] From the linear correlation diagram between the predicted value and the experimental value ( Figure 6 (A)) and the Williams diagram used to characterize and define AD ( Figure 6 (B)) The prediction quality of the model can be observed more intuitively. The results show that there are no effect outliers in the RF model. Only one compound was found with an h value greater than the warning h*, which is a structural outlier, but its prediction residual is very small, indicating that the model's prediction is reliable.

[0102] 2.2. Male rat (RM) QSAR model

[0103] The RM model obtained based on ORes partitioning (N tr =33, N test =10) works best, as shown in equation (2), it can explain 62.56% of the variance of the training set (Q 2 LOO 50.48%), and 71.83% of the test set variance (R 2 ext). Based on the MAE criterion, the prediction quality of the RM model was evaluated as “good”. The model contains six descriptors (J_D, ATSC3s, GATS5s, ATSc4, SHAaCH and minaaCH).

[0104] TD 50 =9.7694–3.3927J_D+0.0194ATSC3s–0.5947GATS5s–4.4007ATSc4–0.8934SHaaCH+1.6534minaaCH (2)

[0105] N tr =33,N test =10, R 2 =0.6256, Q 2 LOO =0.5048, R 2 ext =0.7183, Q 2 F1 =0.6270, Q 2 F2 =0.6255, Q 2 F3 =0.8203, CCC test =0.7757, RMSE test =0.7736, MAE test =0.7123, r 2 m(test) =0.6079, Δr 2 m(test) =0.2056, TSR = 9.3548, MAE 95％test +3*SD 95％test =1.5475, Quality Test =Good

[0106] According to VIP chart analysis ( Figure 3 B), the most important descriptors in the model are SHaaCH and GATS5s, and the secondary descriptors are J_D, ATSc4, ATSC3s and minaaCH. According to the loading diagram ( Figure 4 (B) ), SHaaCH and GATS5s had the greatest impact on carcinogenicity in male rats, which was consistent with the results of VIP plot analysis.

[0107] Descriptors in the RM model can be divided into two categories. The first category of descriptors is positively correlated with carcinogenicity, including ATSC3s and minaaCH. The larger the value, the stronger the carcinogenicity of the compound, and vice versa. The other category of descriptors is negatively correlated with carcinogenicity, including SHAaCH, J_D, GATS5s and ATSc4. Figure 7 As shown in (A), for example, compound 2 (2-acetylaminofluorene) with a high minaaCH value (1.945) showed high carcinogenicity (pTD 50 =5.28), compound 47 (2,3,7,8-tetrachlorodibenzo-p-dioxin) with a high ATSC3s value (91.704) also has a high pTD 50 value (5.21). On the contrary, Figure 7 As shown in (B), compound 50 (phenolphthalein) (having a high SHaaCH value), compound 17 (chlorobenzoate) (having a high GATS5s value), compound 15 (butyl benzyl phthalate) (having a high J_D value) and compound 28 (benzoin) (having a high ATSc4 value) showed low carcinogenicity.

[0108] The linear correlation diagram between the predicted value and the experimental value and the Williams diagram used to characterize the definition of AD are respectively Figure 6 (C) and Figure 6 (D). Although compound 25 affected the regression to some extent, there were no effect outliers in the RM dataset; we observed that four training set compounds (23, 55, 58, and 66) and one test compound (60) had h values ​​larger than the alert h*, but their prediction residuals were within the acceptable range.

[0109] 2.3. Female mouse (MF) QSAR model

[0110] The MF model obtained based on Rnd partition (N tr =33,N test = 10) works best, as shown in equation (3), it can explain 81.32% of the variance of the training set (Q 2 LOO 71.55%), and 82.64% of the test set variance (R 2 ext ). Based on the MAE criterion, the prediction quality of the model was evaluated as “good”. The model contained six descriptors, namely J_D, SpMaxA_B(s), GATS8s, SpMax2_Bh(v), Cl-089, and minaaaC.

[0111] TD 50 =58.8706–3.4972J_D-9.3578SpMaxA_B(s)+0.3697GATS8s-12.5216SpMax2_Bh(v)+0.9240Cl-089-1.7211minaaaC (3)

[0112] N tr =33,N test=10, R 2 =0.8132, Q 2 LOO =0.7155, R 2 ext =0.8264, Q 2 F1 =0.8126, Q 2 F2 =0.8082, Q 2 F3 =0.8597, CCC test =0.8809, RMSE test =0.5882, MAE test =0.5038, r 2 m(test) =0.6230, Δr 2 m(test) =0.1844, TSR=9.2230, MAE 95％test +3*SD 95％test =1.1432, Quality Test =Good

[0113] like Figure 3 As shown in (C), the most important descriptors in the model are Cl-089, SpMax2_Bh(v) and GATS8s, and the secondary descriptors are SpMaxA_B(s), minaaaC and J_D. Figure 4 (C)) also showed the same result, that is, Cl-089, SpMax2_Bh(v) and GATS8s had the greatest impact on the model; the remaining three descriptors were relatively close to the origin and far away from the predicted effect endpoint, so their importance was second.

[0114] Similarly, the descriptors in this model were divided into two categories. The first category of descriptors was positively correlated with carcinogenicity, including Cl-089 and GATS8s ( Figure 8 (A)). For example, the Cl-089 value of compound 59 (2,3,7,8-tetrachlorodibenzo-p-dioxin) is 4.0, and the GATS8s value of compound 7 (4-aminobiphenyl) is 1.945. The larger the value, the stronger the corresponding carcinogenicity (pTD 50 8.63 and 5.23, respectively); and it can also be seen in compound 69 (Pigment Red 3) and compound 10 (Benzoin) that the carcinogenicity decreases with the decrease of descriptor value. Another group of descriptors is negatively correlated with carcinogenicity, including SpMax2_Bh (v), SpMaxA_B (s), minaaaC and J_D ( Figure 8(B)). For example, compound 7 (4-aminobiphenyl), compound 22 (chlorophenanthane), compound 21 (2,2-bis(4-chlorophenyl)-1,1-dichloroethylene) and compound 2 (2-acetylaminofluorene) have low SpMax_Bh(v) (3.671), SpMaxA_B(s) (0.464), minaaaC(0) and J_D (1.6116), respectively, and instead show high carcinogenic potency. On the contrary, the higher the value of these descriptors, the lower the corresponding carcinogenic potency, as shown by compound 69 (Pigment Red 3), compound 10 (benzoin), compound 51 (N-phenyl-2-naphthylamine) and compound 23 (decabromodiphenyl ether).

[0115] The linear correlation diagram between the predicted value and the experimental value and the Williams diagram used to characterize the definition of AD are respectively Fig. 9 (A) and 9(B). It is noteworthy here that no structural outliers and effect outliers were found in the established MF carcinogenicity model.

[0116] 2.4. Male mouse (MM) QSAR model

[0117] The MM model (N) obtained based on OStr division tr =36, N test =10) works best, as shown in equation (4), it can explain 67.53% of the training set variance (Q 2 LOO 50%), and 88.02% variance (R 2 ext ). Based on the MAE criterion, the prediction quality of the model was evaluated as “good”. Seven descriptors were included in the model, namely CIC4, MATS4s, GATS7e, P_VSA_MR_7, B10[CC], VCH-6, and maxwHBa.

[0118] TD 50 =–2.7777+0.7292CIC4–5.3604MATS4s+1.5330GATS7e+0.0085P_VSA_MR_7–1.2361B10[CC]+17.2964VCH-6+1.1910maxwHBa (4)

[0119] N tr =36, N test =10, R 2 =0.6753, Q 2 LOO =0.5000, R 2 ext =0.8802, Q 2F1 =0.7952, Q 2 F2 =0.7860, Q 2 F3 =0.9486, CCC test =0.9133, RMSE test =0.4002, MAE test =0.3488, r 2 m(test) =0.6866, Δr 2 m(test) =0.1437,TSR=9.2230,MA E95％test +3*SD 95％test =0.8581, Quality Test =Good

[0120] According to VIP analysis ( Figure 3 D), the most important descriptors in the model are maxwHBa, P_VSA_MR_7 and MATS4s, and the secondary descriptors are CIC4, GATS7e, VCH-6 and B10[CC]. Loading Figure ( Figure 4 D) also reflects the importance of the variables, and the results are consistent with VIP analysis.

[0121] Similarly, the descriptors that appeared in the MM model were divided into two categories. The first category of descriptors were maxwHBa, P_VSA_MR_7, VCH-6, GATS7e, and CIC4, which were positively correlated with carcinogenicity. Fig.10 As shown in A, compound 56 (rotenone) is a typical mitochondrial respiratory chain inhibitor, contains many oxygen atoms that serve as hydrogen bond acceptors, and has a high carcinogenicity to male mice (pTD 50 =5.71); P_VSA_MR_7 belongs to the P_VSA-like descriptor, which is related to the lipophilicity characteristics. Compound 22 (2,2-bis(p-chlorophenyl)-1,1,1-trichloroethane) with a high P_VSA_MR_7 value shows strong carcinogenicity; similarly, compound 21 (2,2-bis(4-chlorophenyl)-1,1-dichloroethylene), compound 2 (2-acetylaminofluorene) and compound 31 (4-amino-4'-fluorobiphenyl) with strong carcinogenicity have high CIC4, VCH values-6 and GATS7e, respectively. Another type of descriptor is MATS4s and B10[CC], which are negatively correlated with carcinogenicity ( Fig.10 (B)). Lower values ​​of MATS4s and B10[CC] were observed in compounds with high carcinogenic potency, such as compound 36 (clofenac methyl ester) and compound 31 (4-amino-4'-fluorobiphenyl).

[0122] From the linear correlation diagram between the predicted value and the experimental value ( Fig. 9 (C)) and the Williams diagram used to characterize and define AD ( Fig. 9 It can be intuitively observed in (D) that two compounds were identified as outliers, one of which is compound 56 (rotenone) which is a structural outlier and the other is compound 59 (2,3,7,8-tetrachlorodibenzo-p-dioxin) which is an effect outlier. It is worth noting here that although the h value of 56 (0.8268) is greater than the warning h* (0.6667), its prediction residual is very small.

[0123] 2.5. Rat (R)QSAR model

[0124] Based on the R model obtained by OStr division (N tr =40, N test =11) works best, as shown in equation (5), it can explain 62.77% of the training set variance (Q 2 LOO 60.23%), and 79.21% of the test set variance (R 2 ext ). Based on the MAE criterion, the prediction quality of the R model was evaluated as “good”. The model has a total of 7 descriptors, namely SHaaCH, SIC1, J_D, MATS5m, SpMad_EA, SM03_EA(bo) and nArOH.

[0125] TD 50 =–17.7427–0.6262SHaaCH+4.4089SIC1–4.6565J_D+3.7901MATS5m+25.0407SpMAD_EA–1.8307SM03_EA(bo)–0.5179nArOH (5)

[0126] N tr =40, N test =11, R 2 =0.6277, Q 2 LOO =0.6023, R 2 ext =0.7921, Q 2 F1 =0.7843, Q 2 F2 =0.7809, Q 2 F3 =0.8710, CCC test =0.8882, RMSE test =0.6382, MAE test=0.4930, r 2 m(test) =0.7121, Δr 2 m(test) =0.0051, TSR = 9.5990, MAE 95％test +3*SD 95％test =1.5860, Quality Test =Good

[0127] According to VIP analysis ( Figure 3 (E)), the most important descriptors in the model are SHaaCH and MATS5m, and the secondary descriptors are J_D, SIC1, nArOH, SM03_EA(bo) and SpMAD_EA. As shown in the loading diagram ( Figure 4 (E)) As shown, SHaaCH and MATS5m contributed most to rat carcinogenicity, consistent with the VIP analysis.

[0128] The descriptors that appeared in the R model were divided into two categories. The first category of descriptors were positively correlated with carcinogenicity, including MATS5m, SIC1, and SpMad_EA. It is not surprising that MATS5m appeared in the R model because it is also a key variable in the RF model. Fig.11 As shown in (A), compound 49 (4,4'-oxydianiline) (with high MATS5m value), compound 58 (mycin) (with high SIC4 value) and compound 33 (N-hydroxy-2-acetylaminofluorene) (with high SpMad_EA value) showed strong carcinogenicity. Another class of descriptors was negatively correlated with carcinogenicity, including SHaaCH, J_D, nArOH and SM03_EA(bo). Similarly, SHaaCH was the most important descriptor in the RM model ( Figure 7 (B)), such as Fig.11 As shown in (B), higher values ​​of these descriptors generally lead to weaker carcinogenic potency, such as compound 57 (sulfasalazine) (with a high SHaaCH value), compound 17 (chlorobenzyl salt) (with a high J_D value), compound 54 (quercetin) (with four aromatic hydroxyl groups) and compound 23 (decabromodiphenyl ether) (with a high SM03_EA(bo) value).

[0129] From the linear correlation diagram between the predicted value and the experimental value ( Figure 6 (E)) and the Williams diagram used to characterize and define AD ( Figure 6 (F)) It can be intuitively observed that only one compound 64 (Pigment Yellow 12) in the test set has an h value greater than the warning h*, which is a structural outlier, but its prediction residual is very small, indicating that the model's prediction is reliable. We also found an effect outlier (Compound 25) whose prediction standardized residual is greater than 3.

[0130] Mouse (M) QSAR model

[0131] Based on the M model obtained by Rnd partition (N tr =44,N test =11) has the best effect, as shown in formula (6), which can explain 61.61% of the training set variance (Q 2 LOO 54.72%), and 71.02% of the test set variance (R 2 ext ). Using the MAE-based criterion, the prediction quality of the model was evaluated as “good”. The established M model includes eight descriptors, namely MATS6, SpMax8_Bh(e), P_VSA_LogP_5, P_VSA_MR_7, SpMaxA_EA(dm), CATS2D_07_PL, minaaCH, and ETA_dbeta.

[0132] TD 50 =3.9044+1.587MATS6e–1.1674SpMax8_Bh(e)+0.0192P_VSA_LogP_5+0.0092P_VSA_M R_7–8.423SpMaxA_EA(dm)+0.3554CATS2D_07_PL+1.2319minaaCH–0.1702ETA_dBeta (6)

[0133] N tr =44,N test =11, R 2 =0.6161, Q 2 LOO =0.5472, R 2 ext =0.7102, Q 2 F1 =0.6526, Q 2 F2 =0.6325, Q 2 F3 =0.8062, CCC test =0.8290, RMSE test =0.7119, MAE test =0.6404, r 2 m(test) =0.6015, Δr 2 m(test) =0.1599, TSR = 9.2230, MAE 95％test +3*SD 95％test=1.5133, Quality Test =Good

[0134] like Figure 3 As shown in (F), descriptors with VIP scores greater than 1.0 are considered to be the most influential descriptors, including P_VSA_MR_7, ETA_dBeta, and minaCH. In contrast, descriptors with VIP scores less than 1.0 are considered to be less influential descriptors, including P_VSA_LogP_5, MATS6e, SpMax8_Bh(e), SpMaxA_EA(dm), and CATS2D_07_PL. Figure 4 (F))The results are consistent.

[0135] The descriptors that appeared in the M model were divided into two categories. The first category of descriptors were positively correlated with carcinogenicity, including P_VSA_MR_7 (also appeared in the MM model), P_VSA_LogP_5, CATS2D_07_PL, minaaCH, and MATS6e. Obviously, the first three descriptors are closely related to lipophilicity. Chemicals with lipophilicity are more likely to penetrate cell membranes and react with DNA or proteins, resulting in increased carcinogenicity. For example, Fig.12 As shown in (A), compounds 22 (2,2-bis(p-chlorophenyl)-1,1,1-trichloroethane), compound 2 (2-acetylaminofluorene) and compound 31 (4-amino-4'-fluorobiphenyl) have high values ​​of this type of descriptor and therefore have strong carcinogenic potency. Interestingly, minaaCH, as a descriptor positively correlated with toxicity, also appeared in the RM model, and compound 21 (2,2-bis(4-chlorophenyl)-1,1-dichloroethylene) with a high minaaCH value showed high carcinogenicity in mice. MATS6e, which is associated with electronegativity, also has a positive effect on carcinogenicity, as shown by compound 7 (4-aminobiphenyl). The second type of descriptors are negatively correlated with carcinogenicity ( Fig.12 (B)), including ETA_dBeta, SpMax8_Bh(e) and SpMaxA_EA(dm). The higher the value of these descriptors, the lower the carcinogenicity of the corresponding compound, as shown in compound 51 (N-phenyl-2-naphthylamine), compound 13 (bisphenol A) and compound 10 (benzoin).

[0136] From the linear correlation diagram between the predicted value and the experimental value ( Fig. 9 (E)) and the Williams diagram used to characterize and define AD ( Fig. 9 (F)), it was observed that no structural outliers were found and only one compound 25 was identified as an effect outlier as its predicted standardized residual was greater than 3.

[0137] 3. iQCCR Model of PAHs Carcinogenicity in Rats and Mice

[0138] The present invention constructs two interspecies quantitative carcinogenicity-carcinogenicity relationship (iQCCR) models.

[0139] 3.1. Rat-mouse (RM) iQCCR modeling

[0140] Rat-mouse (RM) iQCCR model for predicting carcinogenicity of PAHs (N tr =28,N test =9) As shown in formula (7), various internal and external validation parameters are met, so it is highly robust and externally predictive. The RM iQCCR model can explain 72.9% of the training set variance (Q 2 LOO 65.66%), and 74.46% of the test set variance (R 2 ext ). Q 2 Yscr and R 2 Yscr Having low values ​​indicates that there is no chance correlation in the model building process.

[0141] TD 50 (Rat) = 0.3553 + 0.8987 pTD 50 (Mouse) (7)

[0142] N tr =28,N test =9, R 2 =0.7290, Q 2 LOO =0.6566, R 2 ext =0.7446, Q 2 F1 =0.7351, Q 2 F2 =0.7119, Q 2 F3 =0.8494, Q 2 Yscr =-0.1311, R 2 Yscr =0.0365, CCC test =0.8305,

[0143] RMSE test =0.6500, MAE test =0.5003, r 2 m(test) =0.6467, Δr2 m(test) =0.1602, TSR=8.7784, MAE 95％test +3*SD 95％test =2.1217, Quality Test =Moderate

[0144] As with the QSAR model, the prediction quality of the iQCCR model was evaluated based on the MAE criterion. The prediction quality of the R-MiQCCR model was determined to be “moderate” because it met the following conditions: MAE 95% test (0.4237)<0.1×TSR(0.8778); MAE95% test +3×SD95% test (2.1217)>0.2×TSR(1.7556). However, the prediction quality of each compound in the test set was evaluated using the "Prediction Reliability Index" tool and was rated as "good". Fig.13 (A) It can be seen that the compounds in the training set and the test set are evenly distributed around the trend line. Williams plot ( Fig.13 (B)) shows that only compound 59 was identified as a structural outlier; however, its prediction residual was relatively low, which also indicates that the model prediction accuracy is high.

[0145] 3.2. Mouse-rat (MR) iQCCR modeling

[0146] Mouse-rat (MR) iQCCR model for predicting carcinogenicity of PAHs (N tr =29, N test =8) As shown in formula (8), it meets various internal and external validation indicators and is therefore highly robust and externally predictive. The MR iQCCR model can explain 72.98% of the training set variance (Q 2 LOO 69.14%) and 79.9% of the test set variance (R 2 ext ). Q 2 Yscr and R 2 Yscr Having low values ​​indicates that there is no chance correlation in the model building process.

[0147] TD 50 (Mouse)=0.3426+0.8451pTD 50 (Rat) (8)

[0148] N tr =29, N test =8, R2 =0.7298, Q 2 LOO =0.6914, R 2 ext =0.7990, Q 2 FI =0.7251, Q 2 F2 =0.7171, Q 2 F3 =0.9103, Q 2 Yscr =-0.1329, R 2 Yscr =0.0355, CCC test =0.8815, RMSE test =0.4899, MAE test =0.4687, r 2 m(test) =0.6877, Δr 2 m(test) =0.1737, TSR=8.4514, MAE 95％test +3*SD 95％test =0.8209, Quality Test =Good

[0149] Based on the MAE criteria, the prediction quality of the MR iQCCR model was evaluated as “good” because it met the following conditions: MAE 95% test (0.4351)<0.1×TSR(0.8451);

[0150] MAE 95% test +3×SD95% test (0.8209) < 0.2 × TSR (1.6903). In addition, all compounds in the test set were evaluated for prediction using the “Prediction Reliability Index” tool and obtained a “good” prediction quality. Fig.13 (C) It can be seen that the training and test compounds are evenly distributed around the trend line. Similar to the RM iQCCR model, the Williams plot ( Fig.13 (D)) also shows that only compound 59 was identified as a structural outlier; however, its prediction residual was very low, indicating the reliability of the model prediction.

[0151] 3.4 Filling data gaps using established QSAR and iQCCR models

[0152] One of the main purposes of the QSAR and iQCCR models established in the present invention is to predict the carcinogenicity of PAHs in the absence of experimental values ​​to fill the gap in carcinogenicity data. For each data set, the present invention also collected real external set compounds from CPDB (Table 1) and predicted the corresponding carcinogenicity endpoints. Fig.14 As shown in the figure, the insubria plot shows that the six QSAR models and the two iQCCR models have significant prediction coverage for the real external set compounds. The prediction coverage of the RF, RM, MF, MM, R, M, RM and MR models is 84.5%, 87.1%, 87.1%, 90.7%, 93.5%, 91.7%, 100% and 100%, respectively. In addition, the prediction quality of each compound was evaluated as "good", "moderate" or "poor / unreliable" according to the "Prediction Reliability Index" tool. According to the prediction results listed in Table 5, overall, most of the compounds are "good" (>83.8%) or "good and moderate" (>90.8%). Only 13 out of 142 compounds in the RF model, only 13 out of 132 compounds in the RM model, and only 2 out of 129 compounds in the MM model were predicted as "bad / unreliable".

[0153] Table 5 Prediction quality of the real external set of established QSAR and iQCCR models

[0154]

[0155] The model established in the present invention was strictly verified by QSAR according to the five principles of OECD, showing good external prediction ability, and was applied for the first time to predict the carcinogenicity of real external set compounds without experimental values, indicating that the model of the present invention can be used to predict the carcinogenicity of new or untested PAHs, and can also be used as a useful tool for making regulatory decisions and hazard assessments.

Claims

1. A method for predicting the carcinogenicity of polycyclic aromatic hydrocarbons, comprising the following steps: Obtain the 2D molecular descriptors of PAHs, and obtain the carcinogenicity of PAHs based on the carcinogenicity quantitative structure-activity relationship model or the interspecies quantitative carcinogenicity-carcinogenicity relationship model, that is, to predict the carcinogenicity of PAHs; The carcinogenicity quantitative structure-activity relationship model is any one of the following formulas (1) to (6): The carcinogenicity QSAR model applicable to female rats is as follows (1): p TD 50 =4.8314+5.1819MATS5m+7.2761MATS6v-0.7063minsCH3-0.0866nAtomP+3.3684nFG12Ring (1) The carcinogenicity QSAR model applicable to male rats is as follows (2): p TD 50 =9.7694–3.3927J_D+0.0194ATSC3s-0.5947GATS5s-4.4007ATSc4-0.8934SHaaCH+1.6534minaaCH (2) The carcinogenicity QSAR model applicable to female mice is as follows (3): p TD 50 =58.8706-3.4972J_D-9.3578SpMaxA_B(s)+0.3697GATS8s-12.5216SpMax2_Bh(v)+0.9240Cl-089-1.7211minaaaC (3) The carcinogenicity QSAR model applicable to male mice is as follows (4): p TD 50 =–2.7777+0.7292CIC4-5.3604MATS4s+1.5330GATS7e+0.0085P_VSA_MR_7-1.2361B10[CC]+17.2964VCH-6+1.1910maxwHBa(4) The carcinogenicity QSAR model applicable to rats is as follows (5): p TD 50=-17.7427-0.6262SHaaCH+4.4089SIC1-4.6565J_D+3.7901MATS5m+25.0407SpMAD_EA-1.8307SM03_EA(bo)-0.5179nArOH(5) The carcinogenicity QSAR model applicable to mice is as follows (6): p TD 50 =3.9044+1.587MATS6e-1.1674SpMax8_Bh(e)+0.0192P_VSA_LogP_5+0.0092P_VSA_MR_7-8.423SpMaxA_EA(dm) + 0.3554CATS2D_07_PL + 1.2319minaaCH – 0.1702ETA_dBeta (6) In various p TD 50 represents the carcinogenic potency of PAHs, and the remaining parameters represent the 2D molecular descriptors of PAHs; The interspecies quantitative carcinogenicity-carcinogenicity relationship model is any one of the following formulas (7) to (8): TD 50 (Rat) =0.3553+0.8987 p TD 50 (Mouse)(7) p TD 50 (Mouse) =0.3426+0.8451 p TD 50 (Rat)(8) In the formula, p TD 50 (Mouse) indicates the carcinogenicity of PAHs to mice. p TD 50 (Rat) indicates the carcinogenic potency of PAHs in rats; DRAGON and PaDEL-Descriptor software were used to obtain 2D molecular descriptors of PAHs.

2. A computer device comprising a processor and a memory storing a computer program; the processor is configured to execute the computer program to implement the method for predicting the carcinogenicity of polycyclic aromatic hydrocarbons as claimed in claim 1.

3. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for predicting the carcinogenicity of polycyclic aromatic hydrocarbons as claimed in claim 1 is implemented.

Images