Chemometrics prediction model and method for predicting acute oral toxicity of six typical drug skeleton molecules

Abstract

The invention discloses a chemometrics prediction model and method for predicting acute oral toxicity of six typical drug skeleton molecules. Based on interpretable chemometrics and machine learning methods, the invention provides an acute oral toxicity prediction model for six types of drug skeleton molecules (pyrazine, piperazine, thiazole, thiophene, indole and benzimidazole), and aims to reduce animal experiment and drug development costs. According to the method, 12 high-precision QSAR models are constructed by integrating rodent animal toxicity data (rat / mouse LD50) of 1150 compounds, and prediction performance is optimized by combining cross reference (q-RASAR) and seven machine learning algorithms (random forest, SVM, neural network and the like). An early toxicity screening tool is provided for drug research and development, the experiment cost and period are remarkably reduced, meanwhile, environmental risk assessment is supported, and green drug design and chemical supervision decision making are assisted. All models and data are open-sourced and can be applied to toxicity prediction of new compounds through a standardized process.

Description

Technical Field

[0001] The present invention relates to a chemometric prediction model and method for predicting the acute oral toxicity of six typical drug skeleton molecules, and belongs to the field of compound toxicity prediction. Background Art

[0002] The analysis of the core structure of drug molecules has always been the core focus of medicinal chemistry research. Bemis and Murck systematically proposed the theoretical framework of molecular skeletons in 1996: by removing all substituents (R groups) in the compound and retaining the aliphatic connection structure between ring systems, a standardized definition of molecular skeletons was successfully constructed. Based on this methodological innovation, researchers successfully identified 41 of the most representative drug basic skeletons from the Comprehensive Medicinal Chemistry (CMC) database. On this basis, various methods for defining the core structure or ring system of drugs have been successively proposed in the academic community, and their core objectives are: (1) to reveal the core configurations that frequently appear in drug molecules; (2) to explore the structure-activity relationship between drug skeletons and specific biological activities; (3) to provide theoretical guidance for the research and development of innovative drugs. In particular, it is worth noting that the skeleton analysis of drugs included in the FDA Orange Book by Taylor et al. shows that approximately 70% of the marketed drugs are constructed using known ring system skeletons, and the probability of newly discovered drug ring systems in recently marketed new drugs is still relatively low. This discovery fully demonstrates the key position of skeleton analysis in drug design.

[0003] The innovative application of drug skeleton analysis methods has achieved remarkable breakthroughs in the field of targeted therapy, and various characteristic molecular skeletons have shown excellent therapeutic advantages. In the contemporary drug R & D system, heterocyclic compounds occupy a core position as the basic structural skeleton, and their privileged scaffolds play a key role in optimizing pharmacological properties. Taking the research and development of anti-tumor drugs as an example, the chromone skeleton has been confirmed as a privileged pharmacophore through systematic research, and its significantly improved target affinity and optimized pharmacokinetic properties have been widely verified. Pyrazole derivatives, especially molecular variants containing pyrimidine groups, show unique advantages in the screening of antibacterial drugs due to their excellent bioisosteric effects and metabolic stability. In current medicinal chemistry practice, classical heterocyclic skeletons such as quinoline, isoquinoline, pyrazine, piperazine, and benzimidazole are widely used. Through flexible modification site design, these core structures can be derived into compound libraries covering multiple pharmacological activities such as antibacterial, anti-inflammatory, antioxidant, antiviral, and anti-tumor. It is worth noting that the therapeutic potential of such skeletons has extended to the field of complex nervous system diseases and shown clinical value in the intervention of depression and neurodegenerative diseases.

[0004] Drug R & D is a highly complex systematic project. The long cycle and high attrition characteristics have become an industry consensus - approximately 90% of candidate drugs fail in the clinical trial and approval stages, and nearly 30% of the R & D failure cases can be attributed to compound toxicity problems. This highlights the decisive position of toxicity assessment in drug development: candidate compounds must meet the strict safety thresholds established by regulatory agencies such as the FDA and EMA in order to pass the drugability verification. It is worth noting that the drug molecular skeleton, as the core structural unit, not only dominates the pharmacodynamic activity but also serves as the structural basis for toxicity generation. Taking the quinoline / indole skeleton as an example, such aromatic compounds may exhibit the typical toxicity characteristics of polycyclic aromatic hydrocarbons (PAHs). Such molecules can bind to blood proteins (such as albumin) to achieve systemic distribution and generate toxic intermediates such as phenols, quinones, and epoxides through metabolic transformation. These reactive metabolites can covalently bind to biological macromolecules such as DNA, triggering gene mutations and carcinogenic processes. In addition to carcinogenic and mutagenic effects, studies have confirmed that such compounds can also cause acute toxicity through oral ingestion or inhalation, posing a health threat to tissues and organs such as the throat, skin, and eyes.

[0005] In addition to the potential threat to human health, the large-scale application of drug skeleton molecular derivatives in the pharmaceutical industry has led to their release into the environment, having a continuous impact on the ecosystem. With the continuous increase in environmental exposure levels, it has become an urgent task to establish a more stringent drug molecular safety standard system. Against this background, clarifying the structure-activity relationship between the structural characteristics of different drug skeletons and acute oral toxicity is of great value for innovative drug R & D and environmental monitoring of potentially highly toxic compounds.

[0006] With the continuous exploration of the pharmacological activities of drug skeletons, the demand for new drug design and development has become increasingly urgent. However, the traditional animal experiment toxicology evaluation system relied on in new drug research and development has inherent defects such as long cycle and high cost. The 3R principle (Replacement, Reduction, Refinement) proposed by Russell and Burch in 1959 provides a theoretical framework for reducing animal experiments. In response, the US Environmental Protection Agency (EPA) plans to cut animal experiments by 30% by 2025 and advocates the use of computational methods such as Quantitative Structure-Activity Relationship (QSAR) to re-evaluate pesticide chemicals lacking safety data, with the ultimate goal of completely terminating animal experiments by 2035. QSAR models can significantly improve drug discovery efficiency while reducing experimental costs by predicting the activities and toxicities of compounds containing pharmacophoric skeletons. As the core tool of Computer-Aided Drug Design (CADD), QSAR technology has accelerated the discovery and optimization processes of lead compounds in multiple disease areas. Typical examples include the Alzheimer's disease treatment drug E2020 (donepezil, trade name AricePt) designed and synthesized through QSAR research. In addition to the already marketed drugs, there are currently a large number of clinical potential compounds based on QSAR modeling in the R & D stage: in the anti-tumor field, QSAR technology has promoted the development of targeted therapies such as kinase inhibitors and multidrug resistance reversers; for major infectious diseases such as HIV, drug-resistant tuberculosis, and viral hepatitis, QSAR has accelerated the discovery of new inhibitors through the collaboration of virtual screening and molecular optimization strategies; in the research of neurodegenerative diseases (such as Alzheimer's disease and Parkinson's disease), QSAR combined with multi-target pharmacophore models has successfully developed candidate drugs with both high blood-brain barrier permeability and low peripheral toxicity. In addition, QSAR also plays a key role in drug design in the fields of metabolic diseases, cardiovascular diseases, etc., demonstrating its multi-dimensional value as the core computational framework of modern drug research.

[0007] As an important tool for predicting the toxicological characteristics of drugs and drug-like compounds, QSAR provides a computational framework for the screening of safe candidates in early drug discovery. The EU REACH regulation clearly supports the use of QSAR systems to evaluate the risks of approximately 30,000 chemicals in terms of annual production / usage / import volume; the US Agency for Toxic Substances and Disease Registry (ATSDR) has also incorporated QSAR into the chemical toxicity assessment system. In cardiac toxicity prediction, the QSAR strategy that combines hybrid descriptors and machine learning algorithms can accurately identify hERG potassium channel inhibitors (a key inducer of drug-induced arrhythmia); for genotoxicity, QSAR combined with deep learning can predict the DNA response mechanism (such as mutagenicity) of novel scaffold compounds, while rule-based systems can label high-risk functional groups such as aromatic amines and epoxides; acute toxicity and developmental toxicity prediction benefit from models trained with cross-species data (such as zebrafish embryo toxicity) to achieve rapid screening of safe compounds. It is worth noting that organ-specific QSAR models for the liver / kidney, etc., can identify drugs withdrawn from the clinic with a sensitivity of 82% by incorporating mechanisms such as mitochondrial dysfunction and transporter inhibition. Overall, QSAR-driven toxicity assessment is reshaping the drug safety paradigm, enabling researchers to avoid risks at a lower cost and at an earlier stage. The present invention is hereby proposed. Summary of the Invention

[0008] The object of the present invention is to establish a chemometric prediction model and method for predicting the acute oral toxicity of drug scaffold molecules. The present invention collects the acute oral toxicity data of derivatives of six common drug scaffold molecules in rats and mice. On the basis of strictly following the OECD guidelines, an interpretable, transferable, and reproducible QSAR model is provided and strictly statistically verified. All the established QSAR models show good robustness and excellent external prediction performance.

[0009] In addition, the present invention combines the q-RASAR method and the machine learning method to further improve the quality of the 2D-QSAR model and analyze the toxicity mechanism of the key descriptors in the optimal model. The present invention also uses the optimal model to perform high-throughput toxicity prediction on real external compounds and verifies the external prediction quality by applying the application domain and prediction reliability indicator method. The results show that the model established by the present invention has wide applicability in filling data gaps and predicting the toxicity of novel or unknown compounds.

[0010] Based on rodents, two different acute oral toxicity prediction models are provided for each drug scaffold molecule. In each model, pLD 50 represents the acute oral toxicity, which is obtained by converting the LD 50 value to the molar unit (mol / kg) and then performing a negative logarithm conversion, and is positively correlated with the toxicity level; the remaining parameters represent the 2D molecular descriptors and cross-reference similarity descriptors of the drug scaffold molecule.

[0011] The prediction models provided for pyrazine drug skeleton molecules are the model shown in formula (1) obtained based on rats or the model shown in formula (2) obtained based on mice:

[0012] pLD 50 = 3.4823 - 0.227GATS3M - 0.6638(nCb-) - 0.7063C-005 + 0.9514F02[O-O](1)

[0013] pLD 50 = 1.5275 + 0.2929MATS7m + 0.3377GATS4p + 0.6792B01[C-O] - 0.9758B04[N-O](2)

[0014] The prediction models provided for piperazine drug skeleton molecules are the model shown in formula (3) obtained based on rats or the model shown in formula (4) obtained based on mice:

[0015] pLD 50 = 1.8553 - 0.2479ATSC4i - 0.2787C-025 - 0.2908GATS5s + 0.9064Uc (3)

[0016] pLD 50 = 68.0917 - 0.5144SpDiam_X + 0.4637P_VSA_p_3 + 0.4761SHED_AL - 0.3586F09[N-O] - 0.5085F10[C-C] + 0.1252SHssNH (4)

[0017] The prediction models provided for thiazole drug skeleton molecules are the model shown in formula (5) obtained based on rats or the model shown in formula (6) obtained based on mice:

[0018] pLD 50 = 1.8183 + 0.3523MATS7m + 0.4705MATS1p + 0.4107P_VSA_charge_6 - 0.4928F02[O-O] + 0.6562F09[C-S] + 0.2496LOGP99 (5)

[0019] pLD 50 = 2.

[0020] The prediction models provided for thiophene drug skeleton molecules are the model shown in formula (7) obtained based on rats or the model shown in formula (8) obtained based on mice:

[0021] pLD 50 = 1.577 - 0.3599MATS4e - 0.289GATS5p + 0.5252GATS6i + 0.635P_VSA_s_1(7)

[0022] pLD 50 = 0.435 + 0.3796nHM - 0.1023MATS7s + 0.5386GATS1m + 0.2042B03[N - N] - 0.2415F09[C - N] + 0.1397Neoplastic - 50 + 0.4879nBase (8)

[0023] The prediction models provided for indole drug skeleton molecules are the model shown in formula (9) obtained based on rats or the model shown in formula (10) obtained based on mice:

[0024] pLD 50 = 1.8762 + 0.4911H - 052 + 0.4885MATS8m + 0.4926BCUTw - 1h - 0.2225maxHCsats + 0.7426maxssssC (9)

[0025] pLD 50 = 1.238 - 0.2741nCconj - 0.3191MaxsOH + 0.2356SHED_DL + 0.4495B04[O - O] + 0.2323LLS_02 + 0.6119VC - 5 (10)

[0026] The prediction models provided for benzimidazole drug skeleton molecules are the model shown in formula (11) obtained based on rats or the model shown in formula (12) obtained based on mice:

[0027] pLD 50 = 2.0496 + 0.1485MATS4m + 0.1688MATS1s - 0.3895MATS4s - 0.6328minssssC - 0.283T(O..CI) + 0.1601sSCI (11)

[0028] pLD 50= 1.6587 + 0.2544O% + 0.3437GATS6i - 0.9299P_VSA_MR_7 + 0.9732P_VSA_charge_6 - 0.3227C - 044 + 0.4968CATS2D_04_DA (12)

[0029] Based on the prediction model, the present invention also provides a method for predicting the acute oral toxicity of a drug scaffold molecule, comprising the following steps:

[0030] Obtain the 2D molecular descriptors of the drug scaffold molecule, and according to the chemometric prediction model, obtain the acute oral toxicity of the drug scaffold molecule, that is, realize the prediction of the acute oral toxicity of the drug scaffold molecule;

[0031] Use PaDEL and Alvadesc software to obtain the 2D molecular descriptors of the drug scaffold molecule.

[0032] The present invention also provides a prediction device for the acute oral toxicity of a drug scaffold molecule applying the chemometric prediction 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 acute oral toxicity.

[0034] The present invention still further provides a computer storage medium, on which a computer program is stored, characterized in that: when the computer program is executed by a processor, the method for predicting the acute oral toxicity is implemented.

[0035] For the first time, the present invention has established a dual-species (rat / mouse) acute oral toxicity database covering 1150 compounds for 6 types of drug core skeletons (pyrazine / piperazine / thiazole / thiophene / indole / benzimidazole). The provided QSAR / q-RASAR / ML models can achieve high-throughput prediction of LD 50 values, reduce the need for animal experiments, and comply with the international 3R principle and the EPA animal experiment reduction plan. The present invention integrates QSAR (12 models), q-RASAR (cross-reference optimization) and 7 machine learning algorithms to form a complementary model system: external validation of the rat model R 2Reaching 0.7376 - 0.9685 for the rat model and 0.7473 - 0.8915 for the mouse model; the q - RASAR model has a 35 - 47% reduction in MAE for thiophene / benzimidazole - type predictions; neural network / SVM and other ML models perform optimally in most scaffolds, with the prediction error reduced by 15 - 30% compared to traditional QSAR. The present invention reveals key toxicity - influencing factors through SHAP value analysis: electronic parameters (GATS series descriptors) dominate 53% of the models; specific functional groups (such as C - S bonds, N - O structures) are positively correlated with toxicity; hydrophobicity (LogP99) and surface charge distribution (P_VSA) significantly affect bioavailability. The present invention realizes the toxicity prediction of 23,592 untested compounds, with the application domain coverage rate exceeding 97%. The present invention establishes a drug scaffold - toxicity correlation map, provides a prediction reliability grading system (PRI), with the proportion of "high - quality predictions" exceeding 95%, providing data support for regulatory authorities. The present invention for the first time provides a triple - verification system of "chemical space - application domain - reliability index", proposes an optimal strategy of multi - dimensional weighted MCDM model, improves the robustness of the model by integrating 9 statistical parameters, and open - sources 12 executable prediction models (smd format) for direct use by drug R & D institutions. The intelligent toxicity prediction system established by the present invention shortens the toxicity screening cycle in the initial stage of new drug R & D by 60 - 80%, and is expected to reduce the pre - clinical R & D cost by more than 30%, providing key technical support for green drug design and environmental pollutant early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a schematic diagram of the process of the present invention.

[0037] Figure 2 shows the chemical space distribution of the training set / test set / real external set in the models provided by the present invention; Figure A is the pyrazine rat model; Figure B is the piperazine rat model; Figure C is the thiazole rat model; Figure D is the thiophene rat model; Figure E is the indole rat model; Figure F is the benzimidazole rat model; Figure G is the pyrazine mouse model; Figure H is the piperazine mouse model; Figure I is the thiazole mouse model; Figure J is the thiophene mouse model; Figure K is the indole mouse model; Figure L is the benzimidazole mouse model.

[0038] Figure 3 for the predicted pLD in each rat model 50 and the experimental pLD 50 between the linear correlation diagram (Figure A - Figure F) and the Williams diagram (Figure G - Figure L).

[0039] Figure 4 are the key visualization results of each rat model; bubble diagram (Figure A, Figure D, Figure G, Figure J, Figure M, Figure P), loading diagram (Figure B, Figure E, Figure H, Figure K, Figure N, Figure Q) and SHAP diagram (Figure C, Figure F, Figure I, Figure L, Figure O, Figure R).

[0040] Figure 5 To predict pLD in each mouse model 50 and the experimental pLD 50 The linear correlation graphs (Graphs A - F) and Williams graphs (Graphs G - L) between them.

[0041] Figure 6 Key visualization results for each rat model; bubble graphs (Graphs A, D, G, J, M, P), loading graphs (Graphs B, E, H, K, N, Q), and SHAP graphs (Graphs C, F, I, L, O, R).

[0042] Figure 7 Violin plots of the pLD 50 values of rats (Graph A) and mice (Graph B) in the true external compound sets predicted by each model.

[0043] Figure 8 Insubria plots of the distributions of compounds in the training set, test set, and true external set within the application domain for each model; Graph A is for the pyrazine rat model (LSVR); Graph B is for the piperazine rat model (q - RASAR); Graph C is for the thiazole rat model (LSVR); Graph D is for the thiophene rat model (q - RASAR); Graph E is for the indole rat model (NN); Graph F is for the benzimidazole rat model (q - RASAR); Graph G is for the pyrazine mouse model (NN); Graph H is for the piperazine mouse model (QSAR); Graph I is for the thiazole mouse model (SVM); Graph J is for the thiophene mouse model (q - RASAR); Graph K is for the indole mouse model (SGD); Graph L is for the benzimidazole mouse model (q - RASAR).

[0044] Figure 9 PRI evaluation results for each rat model (Graph A) and mouse model (Graph B), where the blue band in the circle represents the proportion of compounds predicted to be of "good" quality, and the red band represents the proportion of compounds predicted to be of "medium" quality.

[0045] Figure 10 Toxicity values and the values of their main descriptors for highly toxic compounds in the true external sets of each model; Graph A is for the pyrazine rat model; Graph B is for the piperazine rat model; Graph C is for the thiazole rat model; Graph D is for the thiophene rat model; Graph E is for the indole rat model; Graph F is for the benzimidazole rat model; Graph G is for the pyrazine mouse model; Graph H is for the piperazine mouse model; Graph I is for the thiazole mouse model; Graph J is for the thiophene mouse model; Graph K is for the indole mouse model; Graph L is for the benzimidazole mouse model.

[0046] Figure 11 Explanation of the toxicity mechanism in the rat model (Graph A) and mouse model (Graph B).

[0047] Figure 12 Results of 2000 Y randomization tests for the optimal QSAR model for each skeletal compound; Panel A is the pyrazine rat model; Panel B is the piperazine rat model; Panel C is the thiazole rat model; Panel D is the thiophene rat model; Panel E is the indole rat model; Panel F is the benzimidazole rat model; Panel G is the pyrazine mouse model; Panel H is the piperazine mouse model; Panel I is the thiazole mouse model; Panel J is the thiophene mouse model; Panel K is the indole mouse model; Panel L is the benzimidazole mouse model.

[0048] Figure 13 Heat map comparison of each model; Panel A is the pyrazine rat model; Panel B is the piperazine rat model; Panel C is the thiazole rat model; Panel D is the thiophene rat model; Panel E is the indole rat model; Panel F is the benzimidazole rat model; Panel G is the pyrazine mouse model; Panel H is the piperazine mouse model; Panel I is the thiazole mouse model; Panel J is the thiophene mouse model; Panel K is the indole mouse model; Panel L is the benzimidazole mouse model. Specific implementation manners

[0049] Unless otherwise specified, the experimental methods used in the following examples are all conventional methods.

[0050] Unless otherwise specified, the materials, reagents, etc. used in the following examples are all available from commercial sources.

[0051] The work flow is as Figure 1 shown.

[0052] I. Materials and methods

[0053] 1. Data collection and processing

[0054] The present invention systematically evaluates the acute oral toxicity of six typical drug skeletal molecules of pyrazine, piperazine, thiazole, thiophene, indole and benzimidazole to rats and mice. The toxicity data are from the TOXNET-ChemIDplus database of the National Library of Medicine, USA (access date: September 16, 2021) and the PubChem compound library (access date: October 8, 2024), and all data are characterized by the median lethal dose (LD 50 ). To meet the modeling requirements, the original LD 50 values (mg / kg) are first converted to mol / kg units through molar mass conversion, and then subjected to negative logarithmic transformation to obtain the toxicity effect endpoint pLD 50 (i.e., pLD 50 = -logLD 50 ). Different from LD 50 , pLD 50The value is positively correlated with the toxicity of the compound. The higher the value, the stronger the toxicity. This data processing method strictly follows the definition requirements of the validation criterion I of the Organization for Economic Co-operation and Development (OECD) for a single clear toxicity endpoint.

[0055] By strictly screening, compounds containing metal ions, inorganic substances, salts, complexes, mixtures, and compounds for which descriptors cannot be calculated are excluded. Finally, 1150 qualified compounds of six types of drug-eluting stents are selected for modeling. The pLD50 values of these compounds range from 0.9524 to 6.1323 (a span of about 6 logarithmic units), and the average range of toxicity values for each dataset is 3 - 4 logarithmic units. Figure 1 The flow chart is shown.

[0056] 2. Descriptor calculation and dataset partitioning

[0057] Molecular descriptors, as the core elements characterizing the physicochemical and structural properties of compounds, form the basis for modeling. In this invention, the two-dimensional molecular structures (in SDF format) of 1150 compounds are obtained from the PubChem database, and descriptors are calculated using PaDEL 2.18 and Alvadesc software respectively. Among them, PaDEL 2.18 provides a total of 1444 descriptors in 63 categories such as atomic electrotopological state, autocorrelation coefficient, atomic class, and bonding class; Alvadesc contains 5636 descriptors in 30 categories such as basic descriptors, ring system descriptors, topological indices, connectivity indices, three-dimensional matrix descriptors, and druglikeness indices. Considering the complexity of three-dimensional conformations, only two-dimensional descriptors with clear physicochemical meanings are selected for the study. To avoid problems of high autocorrelation and overfitting caused by excessive descriptor numbers, QSARINs 2.2.4 software is used for screening: constant descriptors and variables with internal correlation > 95% are removed to obtain the final modeling descriptor set.

[0058] According to OECD criterion IV, the model needs to be validated after construction to evaluate its robustness, goodness of fit, and external prediction ability. For this purpose, the dataset is divided into a training set (for model development) and a test set (for model validation) in a ratio of 3:1. Specifically, three partitioning strategies are adopted: (1) Partitioning by effect value (ORes): The samples of the test set are selected at intervals after arranging the pLD 50 values in descending order; (2) Partitioning by structure (OStr): Based on the scores of the first principal component (PC1) of the descriptors arranged in descending order; (3) Random partitioning (Ran): Completely random sampling.

[0059] To ensure that the training set fully covers the characteristics of the test set, the pLD 50Or the extreme value compounds of PC1 to optimize the application domain (AD). In the specific operation, starting from the second compound, every three samples are assigned to the test set, and the rest are assigned to the training set. All partitioning processes are executed through the QSARINs 2.2.4 software, and each model is independently optimized based on the preliminary results to obtain the best prediction performance.

[0060] 3. Descriptor Selection and Model Construction

[0061] This invention strictly follows the requirements of OECD Guideline II regarding "explicit algorithms", uses the genetic algorithm (GA) for descriptor screening, and constructs a multiple linear regression (MLR) model through ordinary least squares (OLS), which is finally presented in an explicit MLR equation. After completing the initial variable screening, the full subset tool of the QSARINs 2.2.4 software is used to calculate low-dimensional models containing 2 - 3 descriptors, aiming to identify key variables early and avoid the completely random start of GA. The descriptor subsets selected at this stage constitute the core architecture of the initial genetic algorithm population chromosome (GAPC).

[0062] GA simulates the natural biological evolution mechanism and optimizes with the leave-one-out cross-validation correlation coefficient as the fitness function. The key parameter settings are: the maximum number of variables is 7, the population size is 50, the mutation rate is 50%, and the number of iterations is 2000 times. According to the empirical criterion, the number of model descriptors needs to be less than one-fifth of the number of compounds in the training set to prevent overfitting. After repeating the verification process, the optimal model set is finally obtained.

[0063] 4. Model Verification and Evaluation

[0064] This invention strictly follows the comprehensive verification requirements of OECD Guideline IV regarding "goodness of fit, robustness, and predictive ability", verifies the goodness of fit and robustness of the model through internal verification, and ensures the external predictive performance of the model through external verification, all using internationally recognized verification parameters. At the internal verification level, the coefficient of determination R 2 and the adjusted coefficient of determination are used as dual indicators to evaluate the goodness of fit of the model, and at the same time, the leave-one-out cross-validation correlation coefficient is used to quantify the robustness of the model. To exclude the accidental correlation between variables, 2000 Y randomization tests are carried out: by randomly rearranging the dependent variable Y and reconstructing the model based on the original independent variable matrix X, when the correlation coefficient of the obtained random model and the cross-validation coefficient are significantly lower than the original model values (p < 0.01), it can be confirmed that the model is not randomly generated.

[0065] In the variable screening stage, the QUIK rule (Q Under Influence of K) of QSARINs 2.2.4 software was applied, and the decision threshold was set at 0.05. This rule requires that the correlation between the descriptor and the effect value (K_XY) must be significantly higher than the internal correlation between the descriptors (K_XX), so as to effectively exclude the interference of highly collinear variables on the model and realize the pre-screening evaluation of the model quality.

[0066] In the external validation stage, the model established from the training set was applied to an independent test set, and the focus was on examining the prediction ability of the model for compounds not involved in model building. The adopted validation index system includes: the extended prediction correlation coefficient based on the test set Coefficient of determination Consistency correlation coefficient CCC test And the predicted squared correlation coefficient At the same time, the root mean square error RMSE test And the mean absolute error MAE test were used to quantify the prediction accuracy. The final qualified model needs to meet the following criteria simultaneously: R 2 、 CCC test and other positive indicators reach the high value threshold (usually > 0.7), while indicators such as RMSE test and MAE test [[ID=ID=27]]remain at a low level. Given that the performance of each validation parameter may vary, the present invention uses a multi-dimensional weighted evaluation method to comprehensively determine the overall performance of the model. All statistical parameters are shown in Table 1.

[0067] Table 1 Statistical parameters for internal and external validation

[0068]

[0069]

[0070] 5. Application domain

[0071] According to OECD Principle 3, the importance of "clear scope of application" is emphasized. Therefore, the application domain (AD) of the developed QSAR model must be clearly defined. The prediction of compounds within the application domain is considered reliable.

[0072] The leverage distance method and the standardized residual method are used to define the application domain. The leverage method calculates the leverage value (h) for each compound, which is the diagonal value of the descriptor hat matrix of the training set, ranging from 0 to 1. This value measures the spatial distance of the compound from the central structure of the compounds in the training set. If the h value of a certain compound exceeds the threshold h*, it is regarded as a structural outlier, indicating that the compound is outside the applicable range and its predicted value may be unreliable. The calculation formula for the threshold h is: h* = 3(p + 1) / n, where p is the number of model descriptors and n is the number of compounds in the training set.

[0073] In addition, the standardized residual method is also applied to identify effect outliers. If the standardized residual of a certain compound - that is, the difference between the experimental value and the predicted value - is greater than ±3 standard deviation units, the compound is regarded as an effect outlier. To visualize the application domain, a Williams plot is constructed using the leverage value and the standardized residual. For clarity, the predicted values of the compounds in the training set / test set and the true external set are plotted together with the leverage value (h) on an Insubria plot.

[0074] 6. Select the optimal model using MCDM values

[0075] After the modeling process, the multi-criteria decision analysis (MCDM) technique in QSARINs 2.2.4 software is used to select the optimal QSAR model. The MCDM method simultaneously evaluates multiple statistical criteria and summarizes the overall performance of each model into a score ranging from 0 to 1, with a score of 1 indicating the highest validation criterion. MCDM test values are calculated by maximizing and CCC test , while MCDM fit values are determined by maximizing CCC tr and minimizing . After comprehensive evaluation of MCDM test and MCDM fit , the final MCDM value is obtained. Finally, the model with the highest MCDM value, strictly following the OECD principles, meeting the internal and external validation criteria, and having the fewest descriptors is selected as the optimal QSAR model.

[0076] 7. Cross-reference, calculation of RASAR descriptors, and establishment of q-RASAR models

[0077] When the number of available experimental data points is limited, the statistical reliability of QSAR models may be reduced. To overcome the limitations of applying QSAR on smaller datasets, the European Chemicals Agency (ECHA) and some researchers advocate the use of "read-across" (RA) as an alternative assessment tool. Different from QSAR modeling, RA does not require model development but fills information gaps by evaluating the similarity between source compounds and target compounds. By extrapolating the hazard identification results of source compounds to one or more target compounds, RA can accurately predict the effect values of new compounds using limited data.

[0078] The RA process consists of two sequential steps: (1) determining the 10 training set compounds most similar to each test set compound; (2) using these 10 training set compounds for weighted average prediction. Similarity can be calculated by methods such as Euclidean distance, Gaussian kernel function, and Laplace kernel function. To optimize the external validation of the RA process, hyperparameter optimization is first performed using "AutoRA Optimize v1.0" (available at the following link: https: / / sites.google.com / jadavpuruniversity.in / dtc-lab-software / home, accessed on February 24, 2024). This step determines the optimal parameters for RA to achieve the best prediction quality, including the sigma value in the Gaussian kernel function method, the gamma value in the Laplace kernel function, and the number of similar compounds. The optimized parameters are shown in Table 2. Subsequently, the RA work is carried out using "Read-Across-v4.1".

[0079] The q-RASAR method combines the advantages of RA and QSAR prediction techniques and is considered to provide more accurate predictions than traditional QSAR or RA methods. RASAR descriptors - including similarity, error, and consistency measures such as the RA function, CV act and SD Similarity - can effectively characterize the dataset through similarity measures. Using "RASAR-Desc-Calc-v3.0.2", 18 RASAR descriptors were calculated. Finally, these RASAR descriptors, together with physicochemical descriptors, were used for modeling, and the descriptor screening and model building processes were repeated through QSARINs 2.2.4 software to develop the q-RASAR model.

[0080] Table 2 Optimized RA hyperparameters for each model

[0081]

[0082] 8. Interpretable machine learning-driven QSAR models

[0083] In the present invention, seven types of machine learning models were developed for each 2D-QSAR model, namely: Random Forest (RF), k-Nearest Neighbors (kNN), Support Vector Machines (SVM), Linear Support Vector Machines (LSVM), Gradient Boosting (GB), Neural Networks (NN), and Stochastic Gradient Descent (SGD). These models were constructed based on the Orange3 37.0 version software, using descriptors screened by the 2D-QSAR or q-RASAR model, depending on which model showed better prediction performance.

[0084] Subsequently, MAE tr 、 and MAE test and other metrics were used to evaluate the robustness and prediction performance of the models.

[0085] To improve the interpretability of the machine learning models, the present invention introduced the SHapley Additive exPlanations (SHAP) method to visualize the impact of each descriptor on the model prediction results and evaluate its importance.

[0086] By analyzing the SHAP values of the models, the effectiveness of the selected QSAR descriptors can be further verified. The present invention used the "Explain model" component in the Orange3 37.0 version to complete the SHAP analysis based on the machine learning model with the best performance in each basic model.

[0087] The optimal QSAR / q-RASAR / machine learning (ML) model was used to predict the acute oral toxicity of untested external datasets. These external datasets consisted of 23,592 drug scaffold molecules, with approximately 2,000 compounds in each dataset, and the toxicity data of these compounds were not yet available. All compounds were sourced from the PubChem database and were not used in the development and validation of the QSAR model.

[0088] 9. Prediction of real external sets lacking experimental values

[0089] The coverage of the application domain of the model on a real external dataset was evaluated using an Insubria plot. To further verify the reliability of the prediction results on the real external dataset, the present invention uses the "Prediction Reliability Indicator" software tool. The prediction quality of each molecule was classified as "good" (comprehensive score = 3), "medium" (comprehensive score = 2), or "poor or unreliable" (comprehensive score = 1). In the real external dataset, the more compounds rated as "good", the higher the reliability of the model in external prediction.

[0090] II. Results and Discussion

[0091] 1. Chemical Space Analysis

[0092] To improve the prediction performance, it is crucial to ensure that the compounds in the dataset used for modeling have sufficient diversity. Higher diversity can enhance the potential for wide applicability of the model, enabling it to be applied to a broader range. Therefore, the present invention uses the chemical space defined by the molecular weight (Molecular Weight) and ALOGP (lipophilicity partition coefficient) of the modeling compounds to visualize the diversity of the compounds.

[0093] Figure 2 The chemical space distributions of the compounds in the training set, test set, and real external set of the 12 optimal stoichiometric models are shown. The extensive distribution of the compounds in the chemical space verifies that the modeling compounds have sufficient diversity. It is worth noting that the distributions of the test set and external validation set of all models in the chemical space are similar to that of the training set, indicating that the models developed based on these training set compounds have a wide range of applicability.

[0094] 2. QSAR Model

[0095] Models were constructed based on Alvadesc and PaDEL descriptors, and the best QSAR models were selected through the MCDM program. The performance, statistical parameters, and model equations of all 12 QSAR models are summarized in Table 3.

[0096] Both the internal validation and external validation of the qualified models follow the strictest validation criteria. When the model simultaneously meets the following conditions, it is considered to have strong prediction performance: R 2 > 0.7; in the randomization test and are significantly lower than the original model values; CCC test > 0.85; RMSE test and MAE test values are as low as possible.

[0097] Among the 12 models constructed based on six types of drug skeleton molecules, the goodness-of-fit parameter R 2 ranged from 0.7127 to 0.8959 in the rat model and from 0.7015 to 0.8120 in the mouse model. The adjusted coefficient of determination of the rat model ranged from 0.6789 to 0.8760, and that of the mouse model ranged from 0.6868 to 0.7806. The cross-validation coefficient ranged from 0.6340 to 0.8592 in the rat model and from 0.6328 to 0.7412 in the mouse model, indicating that the models had good goodness-of-fit and robustness.

[0098] In terms of external validation, for the rat model, it ranged from 0.7376 to 0.9685, and for the mouse model, it ranged from 0.7473 to 0.8915; while for the rat model, it ranged from 0.7371 to 0.9591, and for the mouse model, it ranged from 0.7296 to 0.8629. The CCC test ranged from 0.8505 to 0.9775 in the rat model and from 0.8557 to 0.9205 in the mouse model, indicating that the models had strong external prediction ability.

[0099] In addition, as Figure 12 shown, both and observed during the Y randomization process were significantly lower than the original model values, further verifying the reliability of the established models and the absence of accidental correlations.

[0100] Table 3 Equations and important validation parameters of all rat and mouse QSAR models

[0101]

[0102]

[0103] 2.1 Rat QSAR model

[0104] a) Pyrazine rat model

[0105] Figure 3 Shows the linear correlation plot between the predicted pLD 50 values and the experimental pLD 50 values in each rat model, as well as the corresponding Williams plot. The concordance correlation plot( Figure 3Figure A shows that there is a strong correlation between the experimental and predicted values of the compounds in the training and test sets, indicating that the pyrazine rat model has good fitting and strong prediction capabilities.

[0106] The application domain of the model was demonstrated by the Williams plot ( Figure 3 Figure G). The two horizontal dashed lines in the figure represent the range of standardized residuals of ±3, and the vertical solid line represents the leverage threshold (h*). The area enclosed by these lines defines the applicable range of the model, and compounds outside this range are considered outliers. Most of the compounds in the training and test sets of the pyrazine rat model are within the defined applicable range, and only compound 10 (CID: 9272) exceeds the leverage threshold and is identified as a structural outlier. However, the predicted residual of this compound is extremely small, further confirming the reliability of the model prediction results.

[0107] According to OECD Principle 5, the generated QSAR model should be comprehensively explained. The optimal model contains four descriptors: GATS3m, nCb-, C-005, and F02[O-O]. The correlation heatmap of the descriptors ( Figure 13 Figure A) shows that the correlation between these descriptors is weak, indicating that they are independent of each other.

[0108] The bubble plot ( Figure 4 Figure A) shows the importance of the descriptors. In the figure, the horizontal axis represents the descriptor number, the vertical axis shows the standardized coefficient of each descriptor in the model equation, and the bubble size reflects the VIP (Variable Importance in Projection) score. The larger the bubble, the higher the VIP value. Red bubbles indicate that the descriptor has a positive contribution to toxicity, and blue bubbles indicate a negative contribution. Among them, the VIP score of F02[O-O] is greater than 1, which is the most important descriptor, and the importance of the other three descriptors is relatively low.

[0109] In addition, the loading plot ( Figure 4 Figure B) was introduced to further clarify the relationship between the descriptors and the toxicity values. This figure uses the W* value, that is, the descriptor resolution score (loading weight) based on the first two principal components, and its numerical size is proportional to the toxicity value. The results of the loading plot are consistent with the conclusions of the bubble plot: F02[O-O] is closest to the pLD 50 value point, indicating that it has a significant impact on toxicity; while nCb- is close to the origin, and C-005 and GATS3m are far from the pLD 50 projection point, indicating that their impact on toxicity is relatively small.

[0110] The SHAP summary plot ( Figure 4Figure C further verifies the effectiveness of the descriptors. Based on the SHAP analysis results of the optimal machine learning model, the descriptors are sorted by importance (average absolute value of SHAP values), stacked vertically, and the horizontal axis represents the SHAP values. Positive SHAP values indicate a possible increase in toxicity values, while negative SHAP values indicate a possible decrease in toxicity values. Each point in the figure represents a compound, and the color gradually changes from low (blue) to high (red) according to the descriptor values. The SHAP plot shows the effectiveness of the selected descriptors. As mentioned before, F02[O-O] is the most important descriptor in the model, and its value has a positive impact on toxicity. The summary plot shows that most of the values of F02[O-O] are low (blue) and concentrated in the negative half of the X-axis, indicating that the low values of this descriptor play an important role in the low toxicity results predicted by the model, which is consistent with the conclusion of its positive contribution. Correspondingly, the low values of the negative descriptors nCb- and C-005 promote the increase of the toxicity prediction value, further verifying the previous analysis conclusion.

[0111] Based on the descriptor coefficients in the model equation (see Table 3), combined with the analysis of the bubble plot and the loading plot, the four descriptors in the piperazine rat model can be classified as positive and negative descriptors. F02[O-O] is a positive descriptor, that is, its value is positively correlated with toxicity; nCb-, C-005, and GATS3m are negative descriptors, that is, their values are negatively correlated with toxicity. For the detailed analysis of the action mechanisms of these descriptors, please refer to the descriptor analysis identified in the QSAR or q-RASAR models proposed in 5.

[0112] b) Piperazine rat model

[0113] Consistency correlation plot ( Figure 3 Figure B in) shows a strong correlation between the experimental and predicted values of the compounds in the test set and the training set, indicating that the model fits well and has high predictive ability. Williams plot ( Figure 3 Figure H in) shows the application domain of the model. Most of the compounds in the training set and the test set fall within the applicable range, and only compounds 1 (CID: 2170), 75 (CID: 3067108), 42 (CID: 67151), 55 (CID: 120984), and 40 (CID: 66057) exceed the leverage threshold (h*). However, the residuals of these compounds are extremely small, indicating that the prediction results of the model are still highly reliable.

[0114] Descriptor correlation heat map ( Figure 13 Figure B in) shows that the correlation between the descriptors is weak, further indicating that the descriptors are independent of each other. Bubble plot ( Figure 4 Figure D in) identifies Uc as the descriptor with the greatest impact. Loading plot ( Figure 4The E graph) further verified the significant impact of the UC value on toxicity, while the SHAP summary graph ( Figure 4 the F graph) emphasized the important role of Uc in model prediction, especially showing its positive correlation with toxicity.

[0115] These results further indicate that the descriptor Uc is positively correlated with toxicity, while the other descriptors (ATSC4i, GATS5s, and C-025) are negatively correlated with toxicity.

[0116] c) Thiazole rat model

[0117] The concordance correlation graph ( Figure 3 the C graph) highlighted the good fit and excellent predictive performance of the thiazole rat model. The Williams graph ( Figure 3 the I graph) showed that most compounds in the training set and test set fell within the application domain, and only compounds 25 (CID: 37113), 30 (CID: 53421), 41 (CID: 85506), and 38 (CID: 74034) exceeded the leverage threshold.

[0118] The descriptor correlation heat map ( Figure 13 the C graph) showed that the correlation between descriptors was weak, indicating that they were independent of each other. The loading graph ( Figure 4 the H graph) identified P_VSA_charge_6 and MATS1p as the most important descriptors, and the bubble graph ( Figure 4 the G graph) further verified that these two descriptors had a significant impact on toxicity. The SHAP summary graph ( Figure 4 the I graph) further confirmed the key role of P_VSA_charge_6 in predicting toxicity and showed its positive correlation with the toxicity value.

[0119] The thiazole rat model showed that P_VSA_charge_6, LOGP99, MATS1p, MATS7m, and F09[C-S] were positively correlated with toxicity, while F02[O-O] was negatively correlated with toxicity.

[0120] d) Thiophene rat model

[0121] The concordance correlation graph ( Figure 3 the D graph) showed a strong correlation between the experimental toxicity values and predicted toxicity values of compounds in the training set and test set. The Williams graph ( Figure 3 the J graph) highlighted the application domain of the model, and all compounds in the training set and test set were within the applicable range, indicating that the model had good external application potential.

[0122] The descriptor correlation heat map ( Figure 13Figure D shows a weak correlation between descriptors, further indicating their independence. The loading plot ( Figure 4 Figure K) indicates that P_VSA_s_1 is the most influential descriptor, and the bubble plot ( Figure 4 Figure J) also confirms the important role of GATS4e and GATS6i in toxicity effects. The SHAP summary plot ( Figure 4 Figure L) further emphasizes the importance of P_VSA_s_1 in model prediction and its positive correlation with toxicity.

[0123] These results show that the descriptors P_VSA_s_1 and GATS6i are positively correlated with toxicity, while MATS4e and GATS5p are negatively correlated with toxicity.

[0124] e) Indole rat model

[0125] The linear correlation plot ( Figure 3 Figure E) shows a strong correlation between the experimental toxicity values and predicted toxicity values of the modeled compounds in the indole rat model. The Williams plot ( Figure 3 Figure K) indicates that most compounds fall within the application domain, and only compound 35 (CID: 3074013) slightly exceeds the leverage threshold h*, and its predicted residual is extremely small.

[0126] The descriptor correlation heat map ( Figure 13 Figure E) shows a weak correlation between model descriptors, further confirming their independence. The bubble plot ( Figure 4 Figure M) indicates that maxssssC is the most influential descriptor, and the loading plot ( Figure 4 Figure N) further supports the conclusion that maxsssC has a significant impact on toxicity. The SHAP summary plot ( Figure 4 Figure O) confirms that maxsssC plays a key role in model prediction and shows its positive correlation with toxicity.

[0127] The analysis results show that the descriptors H-052, MATS8m, BCUTw-1h, SHsOH, and maxssssC are positively correlated with toxicity, while the descriptor maxHCsats is negatively correlated with toxicity.

[0128] f) Benzimidazole rat model

[0129] The linear correlation plot ( Figure 3 Figure F) shows a very strong correlation between the experimental values and predicted values of the modeled compounds in the benzimidazole rat model. The Williams plot ( Figure 3The Chinese L figure) shows the application domain of the model, where most of the compounds in the training set and the test set are within the domain, and only compounds 146 (CID: 135506210), 99 (CID: 156310), 112 (CID: 198158), and 140 (CID: 212898) exceed the leverage threshold h*. Nevertheless, the prediction residuals of these compounds are extremely small, indicating that the prediction results of the model are still reliable.

[0130] Descriptor correlation heatmap ( Figure 13 The Chinese F figure) shows a weak internal correlation between descriptors, further indicating that the descriptors are independent of each other. Bubble plot ( Figure 4 The Chinese Q figure) highlights minssssC as the most important descriptor, and the loading plot ( Figure 4 The Chinese P figure) further verifies the conclusion that minssssC has a significant impact on toxicity. SHAP summary plot ( Figure 4 The Chinese R figure) further confirms the significant impact of minssssC on the model prediction and emphasizes its negative correlation with toxicity.

[0131] The benzimidazole rat model shows that the descriptors MATS4m, MATS1s, and SsCI are positively correlated with toxicity, while MATS4s, minssssC, and T(O..CI) are negatively correlated with toxicity.

[0132] 2.2, Mouse QSAR model

[0133] a) Pyrazine mouse model

[0134] Figure 5 The Chinese A figure shows that all points in the training set and the test set are evenly distributed on both sides of the trend line, verifying the good fit and excellent prediction ability of the model. Williams plot ( Figure 5 The Chinese G figure) shows the application domain of the model. In the pyrazine mouse model, almost all compounds in the training set and the test set are within the AD, and only compounds 10 (CID: 9362) and compound 14 (CID: 48882) exceed the leverage threshold h*. However, the prediction residuals of these two compounds are extremely small, indicating that the prediction results of the model are still highly reliable.

[0135] Descriptor correlation heatmap ( Figure 13 The Chinese G figure) shows a weak correlation between descriptors, further confirming their independence. Bubble plot ( Figure 6 The Chinese A figure) identifies B04[N-O] and GATS4p as the most important descriptors, with larger bubble diameters and relatively large standardized coefficients, while the importance of the other two descriptors is lower. Loading plot ( Figure 6Figure B) further verified the conclusion of the bubble plot: B04[N-O] is closest to the projection point of pLD in the corresponding quadrant, indicating its significant impact on toxicity; while MATS7m is close to the origin and B01[C-O] is far from the pLD projection point, suggesting that these two descriptors have relatively little impact on toxicity. 50 The projection point, indicating that it has a significant impact on toxicity; while MATS7m is close to the origin and B01[C-O] is far from the pLD 50 Projection point, indicating that these two descriptors have relatively little impact on toxicity.

[0136] SHAP summary plot ( Figure 6 Figure C) also reached the same conclusion. In the SHAP plot, B04[N-O] significantly improves the model prediction value when its value is low, while significantly reducing the predicted toxicity value when its value is high. This phenomenon further confirms the negative correlation between B04[N-O] and the toxicity level, highlighting the key role of this descriptor in model prediction. On the contrary, the other three positive descriptors show the opposite trend, consistent with the previous conclusion.

[0137] In the pyrazine mouse model, the descriptors MATS7m, GATS4p, and B01[C-O] (in decreasing order of importance) are positively correlated with toxicity, while the descriptor B04[N-O] is negatively correlated with toxicity. For a more detailed analysis of the mechanism of action of these descriptors, see the descriptor analysis identified in the QSAR or q-RASAR models proposed in 5.

[0138] b) Piperazine mouse model

[0139] Figure 5 Figure B shows that all modeled compounds are evenly distributed on both sides of the trend line, indicating that the model has a good fit. Williams plot ( Figure 5 Figure H) shows that most of the training set and test set compounds are within the application domain, and only compounds 32 (CID:9362), 120 (CID:205041), and 12 (CID:16104) exceed the leverage threshold h*. However, the prediction residuals of these compounds are small, indicating that the model prediction results are still reliable. Descriptor correlation heat map ( Figure 13 Figure H) shows weak correlations between descriptors, further verifying the independence of descriptors. Bubble plot ( Figure 6 Figure D) identified SpDiam_X and P_VSA_P_3 as the most important descriptors, and the loading plot ( Figure 6 Figure E) also confirmed that these two descriptors have a significant impact on toxicity. SHAP summary plot ( Figure 6 Figure F) shows that SpDiam_X has a negative effect on toxicity prediction, while P_VSA_P_3 has a positive effect.

[0140] c) Thiazole mouse model

[0141] Linear correlation plot ( Figure 5 Figure C) shows a strong correlation between the experimental toxicity values and predicted values of the compounds in the training set and test set, indicating that the model has high predictive ability. Williams plot ( Figure 5 Figure I) shows that all compounds are within the application domain. Descriptor correlation heatmap ( Figure 13 Figure I) indicates a weak correlation between descriptors. Bubble plot ( Figure 6 Figure G) highlights B02[N-N], s3_relPathLength, and SHBint5 as the most influential descriptors, and the loading plot ( Figure 6 Figure H) further confirms the important role of these descriptors in toxicity prediction. SHAP summary plot ( Figure 6 Figure I) further confirms that B02[N-N] and s3_relPathLength are positively correlated with toxicity, while SHBint5 has a significant negative impact on toxicity.

[0142] d) Thiophene mouse model

[0143] Figure 5 Figure D shows a strong linear relationship between experimental toxicity and predicted toxicity, supporting the high predictive accuracy of the model. Williams plot ( Figure 5 Figure J) shows that most compounds are within the AD, and only compound 80 (CID: 3053655) slightly exceeds the leverage threshold h*, but its residual is extremely small, indicating that the model is still reliable. Descriptor correlation heatmap ( Figure 13 Figure J) shows a weak correlation between descriptors. Bubble plot ( Figure 6 Figure J) indicates that nBase and GATS1m are the most important descriptors, and the loading plot ( Figure 6 Figure K) also shows that these two descriptors have a significant impact on toxicity. SHAP summary plot ( Figure 6 Figure L) further confirms the significant positive impact of nBase and GATS1m on toxicity prediction.

[0144] e) Indole mouse model

[0145] Figure 5 Figure E shows that all modeled compounds are evenly distributed around the trend line, showing good goodness of fit and predictive performance. Williams plot ( Figure 5 Figure K) shows that most compounds are within the application domain, and only compounds 139 (CID: 5280954), 140 (CID: 5282183), 160 (CID: 16132269), and 171 (CID: 76959818) exceed the leverage threshold h*. However, the prediction residuals of these compounds are small, indicating reliable model prediction. Descriptor correlation heatmap ( Figure 13The K diagram shows a weak correlation between descriptors. The bubble chart ( Figure 6 M) identifies VC-5 and B04[N-O] as key descriptors, and the loading plot ( Figure 6 N diagram) also emphasizes their significant impact on toxicity. The SHAP summary plot ( Figure 6 O diagram) further confirms a strong negative correlation between VC-5 and B04[N-O] and toxicity.

[0146] f) Imidazole mouse model

[0147] Figure 5 The F diagram shows a strong correlation between the experimental and predicted values of the compounds in the training and test sets, indicating that the model has good predictive accuracy. The Williams plot ( Figure 5 L diagram) shows that almost all compounds fall within the applicable range, and only compound 5 (CID: 5798) slightly exceeds the leverage threshold h*. Nevertheless, the predicted residual of this compound is extremely small, indicating that the model prediction is still accurate. The descriptor correlation heat map ( Figure 13 L diagram) shows a weak correlation between the selected descriptors. The bubble chart ( Figure 6 P diagram) identifies GATS6i as the most important descriptor, and the loading plot ( Figure 6 Q diagram) further emphasizes the importance of GATS6i in toxicity prediction. The SHAP summary plot ( Figure 6 R diagram) further confirms the positive contribution of GATS6i to toxicity prediction.

[0148] 3. q-RASAR and machine learning models

[0149] The q-RASAR constructed based on 2D-QSAR and the machine learning (ML) model have significantly improved the prediction quality of the model. Table 4 compares the performance of 2D-QSAR, q-RASAR, and the best ML model, covering the number of descriptors and multiple validation metrics.

[0150] According to these criteria, the q-RASAR model shows the best external prediction performance in the rat models of piperazine, thiophene, and benzimidazole skeletons and the mouse models of thiophene and benzimidazole; for the remaining drug skeleton models, the ML-driven models are generally superior to other methods, with the only exception being the piperazine mouse model, in which the 2D-QSAR model performs best.

[0151] Overall, the external prediction performance of both the q-RASAR and ML-driven models is superior to the original 2D-QSAR model, demonstrating obvious advantages.

[0152] Table 4 Important validation parameters for each QSAR, q-RASAR, and machine learning model

[0153]

[0154] 4. Prediction of the acute oral toxicity of untested drug scaffold molecules compounds

[0155] One of the main objectives of the present invention is to use the established model to predict the toxicity of new or unknown compounds, thus filling the existing data gaps. The study retrieved a dataset of 23,592 external compounds from the PubChem database; these compounds lack known acute oral toxicity values in rats or mice and were not involved in the model construction or testing process. Subsequently, the established best model was used to predict the toxicity values of these compounds.

[0156] The reliability and coverage of the model predictions were evaluated using two methods: application domain and PRI. By plotting the violin plot of the predicted pLD 50 value for each external compound ( Figure 7 ), the data distribution was visualized and the toxicity differences between different scaffold groups were highlighted.

[0157] The analysis results showed that there were no significant differences in the average toxicity values of compounds with different drug scaffolds, with the only exception being benzimidazoles compounds, which showed significant differences in the rat model and smaller differences in the mouse model.

[0158] Figure 8 The Insubria plot shown indicates that in each true external dataset, only a few compounds fall outside the model AD. In addition, for most compounds located outside the AD, their leverage value (h) only slightly exceeds the threshold (h*).

[0159] In the rat model, the AD coverage of the model for the true external dataset ranges from 97.31% to 99.05%; in the mouse model, the coverage is also high, ranging from 96.91% to 99.63%. These results highlight the excellent external prediction ability of the best model constructed by the present invention and its application potential in filling the toxicity data gaps.

[0160] The model also performed well in the PRI evaluation. Figure 9 The prediction quality evaluation results of each true external compound by the PRI software are shown.

[0161] In the rat model, in all models except the thiazole scaffold model, the prediction quality of more than 95% of the compounds was rated as "good", and 93.68% of the compounds in the thiazole scaffold model were rated as "good". Only a very small number of molecules were rated as "medium", and no molecules were rated as "poor".

[0162] In the mouse model, more than 95% of the compounds in all models except the benzimidazole skeleton model were also rated "good".

[0163] The above double-verification method shows that the models provided by the present invention have wide applicability and can perform reliable toxicity predictions on a large number of new or untested compounds. This ability is particularly important in filling key data gaps and is expected to play a key role in the lead stage of drug discovery. By helping researchers identify highly toxic molecules, these models can significantly shorten the time required for drug development and provide important guidance throughout the discovery process. In addition, the models established by the present invention will also support policymakers in formulating more effective regulatory policies, regulating the production and use of potentially harmful chemicals, and contributing to environmental risk assessment work.

[0164] To further verify the accuracy and external applicability of the descriptors selected by the models, the present invention selected a representative highly toxic compound from the true external set of each model. Figure 10 Showed the molecular structures of these compounds, the descriptor values included in the models, and the pLD 50 values predicted by the models. Through these visualization results, the correlation between the toxicity values and the descriptor values can be observed, and it is clear which features lead to the high toxicity of the molecules.

[0165] For example, in the true external set of the pyrazine rat model, the compound with the highest toxicity (CID:7469241) was predicted to have a pLD 50 value of 4.5089. It was found that the most important positive descriptor F02[O-O] of this compound had the highest value, which can be considered the main factor leading to its high toxicity.

[0166] 5. Analysis of the descriptors identified in the proposed QSAR or q-RASAR models

[0167] Figure 11 Showed the descriptors included in the 12 optimal models, where the positive descriptors were marked with red boxes and the negative descriptors were marked with blue boxes. For each descriptor, its meaning and the corresponding representative compound were shown in the outermost circular box.

[0168] For example, in the pyrazine mouse model, compounds 18 (CID:25915), 10 (CID:9272), and 6 (CID:7159) had high values for the three positive descriptors - MATS7m, GATS4p, and B01[C-O] - and their pLD 50 values were the highest in the modeling dataset, showing relatively high toxicity. In contrast, for the negative descriptor B04[N-O], compound 36 (CID:202963) had a high value (B04[N-O]=1), and its pLD50 The value is relatively low (pLD 50 = 2.11), indicating relatively low toxicity.

[0169] These examples further verified that the descriptors selected in the QSAR / q-RASAR / ML models can accurately characterize the relationship between the physicochemical properties of compounds and their toxicity values.

[0170] Among all the descriptors selected by the models, autocorrelation type descriptors appeared most frequently, with a total of 18 occurrences, including GATS3m and JGI4. Notably, most of the autocorrelation type descriptors (12 out of 18) were positively correlated with the toxicity value, mainly reflecting the charge distribution, electronegativity, and ionization energy of the molecule. These descriptors characterized the internal electronic properties of the molecule, such as electrical differences and charge distribution, which would affect the interaction between the molecule and biological targets. The reason for the positive correlation between the autocorrelation descriptors and the toxicity value is that higher ionization energy differences and larger charged regions usually enhance the reactivity of the molecule, thus increasing toxicity.

[0171] Descriptors related to the structure or geometric properties of compounds also appeared frequently, including two-dimensional atom pair descriptors (such as F02[O-O], F09[C-S]), atom-centered fragment descriptors (such as C-005, C-044), functional group descriptors (such as nCb-, nCconj-), and shape index type descriptors (such as Kier3). These descriptors usually directly indicate the relationship between specific groups or structures in the compound and toxicity. For example, in the thiazole rat model, the positive effect of F09[C-S] on the toxicity score may mean that the toxicity of molecules containing the C-S bond increases.

[0172] Descriptors characterizing the surface properties of compounds also showed relatively high importance, such as P_VSA type descriptors (P_VSA_charge_6) and electrotopological state atom type descriptors (such as SHsOH, maxHCsats). These descriptors are related to the surface chemical properties of the molecule, especially polarity, hydrophobicity, and hydrogen bonding. Generally, a larger charged surface area or highly hydrophobic region will enhance the reactivity and toxicity of the molecule, while stronger polarity may reduce the binding of the molecule to biological targets, thus reducing toxicity.

[0173] Descriptors with relatively low frequencies of occurrence include drug-like index descriptors, such as Inflammat-80 and Neoplastic-50, which reflect the anti-inflammatory and anti-tumor activity indices of the molecule respectively. These descriptors are very important in drug discovery and can screen out compounds with high toxicity but pharmacological activity at an early stage of research and development.

[0174] Generally speaking, the acute oral toxicity of the six types of drug skeleton molecules selected in this invention in rats and mice is mainly related to electronegativity, polarity, hydrophobicity, and the presence of specific molecular groups. The detailed mechanism explanations of the descriptors selected for each model are as follows:

[0175] Pyrazine skeleton model: For the pyrazine skeleton rat model, the optimal model contains four molecular descriptors: GATS3m, nCb-, C-005, and F02[O-O]. Among them, the first three descriptors are negatively correlated with compound toxicity, while F02[O-O] is positively correlated with toxicity. GATS3m is a Moran autocorrelation descriptor with a lag of 4 and weighted by atomic mass. It is negatively correlated with toxicity because it is related to carbon skeleton structures similar to CH3X, and such structures help reduce toxicity. In contrast, F02[O-O] represents the frequency of O-O bonds with a topological distance of 2, and this descriptor is positively correlated with toxicity. This indicates that the presence of O-O bonds may promote the formation of toxic oxides or free radicals, which in turn lead to cell damage or oxidative stress, thereby increasing toxicity.

[0176] For the pyrazine skeleton mouse model, the descriptors of the optimal model include: MATS7m, GATS4p, B01[C-O], and B04[O-N]. MATS7m is a Moran autocorrelation descriptor with a lag of 4 and weighted by atomic mass, which is used to quantify the spatial relationship between atoms in a molecule. It is positively correlated with toxicity, indicating that in complex structures, high-quality atom pairs may enhance toxicity. GATS4p is a two-dimensional autocorrelation descriptor, representing the Geary autocorrelation index with a lag of 4 and weighted by atomic polarizability. It describes the spatial distribution of atomic polarizabilities in a molecule. Higher polarizability may enhance the interaction with biological macromolecules (such as proteins or DNA), thereby leading to increased toxicity by interfering with key biological functions. B01[C-O] represents the presence or absence of C-O bonds with a topological distance of 1, while B04[O-N] represents the presence or absence of O-N bonds with a topological distance of 4. These two descriptors are negatively correlated with toxicity, probably because O-N bonds with a topological distance of 4 endow the molecule with higher stability, thus reducing its interaction with biological targets and ultimately resulting in reduced toxicity.

[0177] Piperazine skeleton model: For the piperazine skeleton rat model, the optimal model is the q-RASAR model, which contains three descriptors positively correlated with toxicity: RA function (LK), nHet, and GATS3e. The RA function (LK) descriptor is derived from the RASAR method and integrates various structural features of the molecular structure. RA function (LK) is positively correlated with toxicity, verifying the effectiveness of RA and RASAR descriptors in accurately characterizing the structural features of compounds. The nHet descriptor is used to quantify the number of heteroatoms in the molecule; GATS3e is an autocorrelation descriptor weighted by electronegativity, representing the electronegativity correlation between atom pairs with a topological distance of 3. These descriptors together indicate that the toxicity of compounds is closely related to their molecular polarity: the stronger the polarity, the higher the likelihood of adverse interactions with biomolecules, which may form stronger hydrogen bonds or induce greater oxidative stress, thereby leading to increased toxicity.

[0178] For the piperazine skeleton mouse model, the optimal model contains three descriptors negatively correlated with toxicity: SpDiam_X, B09[N-O], and F10[C-C], and three descriptors positively correlated with toxicity: P_VSA_p_3, SHED_AL, and SHssNH. As mentioned above, SpDiam_X is related to molecular size; B09[N-O] represents the presence of an N-O bond with a topological distance of 9; F10[C-C] represents the presence of a C-C bond with a topological distance of 10. These descriptors are negatively correlated with toxicity, indicating that compounds with these structural features and larger molecules may exhibit lower toxicity. P_VSA_p_3 represents the surface area of positively charged molecules in the tertiary volume; SHED_AL is based on the entropy value of the atom pair distribution, reflecting the structural complexity and diversity of the molecule; SHssNH measures the surface tension of nitrogen-hydrogen bonds. These positively correlated descriptors emphasize the relationship between molecular polarity, structural complexity, and the presence of positively charged groups and toxicity.

[0179] Thiazole skeleton model: The thiazole skeleton rat model contains five descriptors positively correlated with toxicity: MATS7m, MATS1p, P_VSA_charge_6, F09[C-S], and LOGP99, and one descriptor negatively correlated with toxicity: F02[O-O]. LOGP99 represents the partition coefficient (LogP) of the molecule between water and octanol and is used to quantify its hydrophobicity. The positively correlated descriptors in the model (such as molecular mass distribution, polarizability, charge surface area, carbon-sulfur bond, and hydrophobicity) are all related to enhanced interaction with biological targets or increased chemical reactivity of the molecule, and these factors may jointly lead to increased toxicity. In contrast, F02[O-O] is negatively correlated with toxicity, indicating that molecules containing such oxygen bonds may have higher stability and lower reactivity, thereby reducing toxicity.

[0180] In the thiazole skeleton mouse model, descriptors such as nR06, B02[N-N], Inflammat-80, and s3_relPathLength are positively correlated with toxicity. nR06 characterizes the number of six-membered rings in the molecule; Inflammat-80 is related to the inflammatory response and may indicate the ability of the molecule to initiate or regulate inflammation; s3_relPathLength measures the average relative distance from atoms in the molecule to a specific path and is related to the molecular size; B02[N-N] indicates the presence of N-N bonds with a topological distance of 2. B02[N-N] is negatively correlated with toxicity, indicating that this type of hydrogen bond interaction may enhance molecular stability, reduce its reactivity, and thus decrease toxicity.

[0181] Thiophene skeleton model: The thiophene skeleton rat model contains four descriptors: gm_class(LK), GATS5p, GATS6i, and P_VSA_s_1. Among them, GATS5p is negatively correlated with toxicity, and the other three descriptors are positively correlated with toxicity. The gm_class(LK) descriptor is related to the consistency coefficient of Banerjee and Roy. When gm equals 1, m is positive; if gm equals 0, then m is positive and m_class is negative; conversely, if MaxPos is greater than MaxNeg, then g and m_class take the value of 1, and vice versa. GATS5p characterizes the polarizability correlation between atom pairs with a topological distance of 5. Its negative correlation with toxicity indicates that molecules with a uniform electronegativity distribution and small polarizability differences may have higher stability, fewer interactions with biomolecules, and thus lower toxicity. GATS6i characterizes the ionization energy correlation between atom pairs with a topological distance of 6; P_VSA_s_1 is used to quantify the surface area size of the negatively charged part of the molecule. Molecules with significant ionization energy differences and large negatively charged surface areas may have higher reactivity and thus higher toxicity.

[0182] For the thiophene skeleton mouse model, the optimal descriptors include four descriptors positively correlated with toxicity: nHM, GATS1m, Neoplastic-50, and nBase, and two descriptors negatively correlated with toxicity: CVact(LK) and F09[C-N]. nHM represents the number of heavy atoms in the molecule; ATS1m characterizes the mass correlation between atom pairs with a topological distance of 1; Neoplastic-50 is related to tumorigenesis and is used to evaluate the potential ability of the molecule to induce tumors; nBase represents the number of basic groups or bases in the molecule and is related to its ability to interact with biological targets. And those negatively correlated with toxicity: CVact(LK) represents the variability of the compound effect value in the RASAR descriptor; F09[C-N] indicates the presence of C-N bonds with a topological distance of 9.

[0183] Indole Skeleton Model: The indole skeleton rat model contains five descriptors that are positively correlated with toxicity: H-052, MATS8m, BCUTw-1h, SHsOH, and maxssssC, and one descriptor that is negatively correlated with toxicity: maxHCsats. H-052 describes the structural configuration where a hydrogen atom in the molecule is connected to an sp 3 hybridized carbon atom (C0), and the adjacent carbon of this carbon atom is connected to a 1X atom; BCUTw-1h is an eigenvalue based on the molecular graph, used to quantify the electron distribution in the molecule, with hydrogen atoms weighted during calculation; SHsOH measures the surface tension related to the hydroxyl group (-OH), reflecting the polarity of the molecule and its ability to form hydrogen bonds; maxssssC represents the maximum number of carbon atoms with four single bonds; these positively correlated descriptors are related to the structural complexity, polarity, electronic effects, and reactivity of the molecule, and may enhance the binding of the molecule to biological targets or biological activity, thereby increasing toxicity. The descriptor negatively correlated with toxicity: maxHCsats represents the maximum number of saturated carbon-hydrogen bonds in the molecule, indicating that the molecular structure is more stable and the reactivity is lower, thus reducing toxicity.

[0184] The indole skeleton mouse model contains two descriptors that are negatively correlated with toxicity: nCconj and MaxsOH, and four descriptors that are positively correlated with toxicity: SHED_DL, B04[O-O], LLS_02, and VC-5. nCconj represents the number of non-aromatic conjugated carbon atoms in the molecule; MaxsOH represents the maximum value of the number of hydroxyl groups (–OH) in the molecule. Hydroxyl groups are polar groups, often related to hydrogen bonding and solubility; these two negatively correlated descriptors indicate that an increase in the number of non-aromatic conjugated carbons and hydroxyl groups can enhance molecular stability, thereby reducing toxicity. The descriptors positively correlated with toxicity include: SHED_DL measures the entropy value of the atom pair distribution, reflecting the complexity and diversity of the molecular structure; B04[O-O] represents the number of oxygen-oxygen bond pairs with a topological distance of 4, and the O-O bond is related to the stability and oxidative stress of the molecule; LLS_02 and VC-5: among them, VC-5 measures the fifth-order principal component of the molecular volume, used to reflect the overall size and structural characteristics of the molecule; these descriptors are positively correlated with toxicity, indicating that a more complex structure, stronger oxygen reactivity, low-frequency electron characteristics, and larger molecular volume may enhance the interaction of the molecule with biological targets or increase its reactivity, thereby increasing toxicity.

[0185] Benzimidazole skeleton model: The benzimidazole skeleton rat model contains four descriptors that are positively correlated with toxicity: MATS1s, Eig02_AEA(dm), minHBd, and RA function(LK), and one descriptor that is negatively correlated with toxicity: F02[O-O]. MATS1s measures the electrical correlation between atom pairs with a topological distance of 1; Eig02_AEA(dm) is an edge-connectivity index descriptor; minHBd represents the lowest energy state of hydrogen bond donors; RA function(LK) is a descriptor obtained based on the results of RASAR analysis; these positively correlated descriptors indicate that characteristics such as the charge distribution of the molecule, the number of hydrogen bond donors, and the molecular geometry may enhance its reactivity or binding ability to biological targets, thereby increasing toxicity. While F02[O-O] is negatively correlated with toxicity, indicating that the presence of the oxygen-oxygen bond may stabilize the molecular structure, reduce its reactivity, and thus lower toxicity.

[0186] The benzimidazole skeleton mouse model contains three descriptors that are positively correlated with toxicity: GATS6i, P_VSA_charge_6, and CATS2D_04_DA, and two descriptors that are negatively correlated with toxicity: P_VSA_MR_7 and C-044. GATS6i characterizes the ionization energy correlation between atom pairs with a topological distance of 6; P_VSA_charge_6 measures the size of the charged region on the molecular surface at the 6th layer charge level; CATS2D_04_DA is used to evaluate the interaction between donor-acceptor groups with a topological distance of 4; these positively correlated descriptors indicate that a higher ionization energy difference, the presence of a charged region, and the interaction between donors and acceptors may enhance the binding or reactivity of the molecule to biological targets, thereby increasing toxicity. P_VSA_MR_7 characterizes the hydrophobic surface area region of the molecule at the 7th layer, which is related to the hydrophobicity and partition coefficient of the molecule; C-044 indicates the presence of a specific carbon structure X–CX...X in the molecule; in addition, gm*Avg.Sim is a similarity consistency index in the RASAR model, where gm represents the similarity degree of the molecule to its similar source compounds. This descriptor reflects the hydrophobicity of the molecule and indicates that the presence of a specific carbon structure and a high similarity to low-toxicity compounds may contribute to reducing toxicity.

[0187] 6. Comparison with other models reported in the literature

[0188] The present invention also analyzes the quality of the constructed model and compares it with previous toxicity modeling studies. Literature review reveals that several computational toxicology studies have reported toxicity models for different drug scaffold derivatives. However, due to differences in dataset composition, scope definition strategies, descriptor attributes and quantities, and target species, it is not possible to make a completely direct comparison between the QSAR and q-RASAR models provided by the present invention and these existing models. Nevertheless, Table 5 compares the main statistical parameters and provides a basic reference for the prediction reliability of different models.

[0189] The model provided by the present invention is constructed in strict accordance with the five principles of OECD QSAR validation, demonstrating excellent external prediction ability.

[0190] The main features of the present invention include: (1) systematically evaluating the acute oral toxicity of multiple drug scaffold molecules in two species for the first time; (2) using a dataset larger in scale than most existing studies; (3) validating the model using multiple advanced validation metrics; (4) further improving the prediction quality of the QSAR model by integrating the q-RASAR method and machine learning techniques; (5) applying the optimal model to the high-throughput toxicity prediction of a real external compound set for the first time; (6) conducting a detailed analysis of the relationship between descriptors and acute oral toxicity; (7) providing all established models in the'smd' format for users to predict the acute oral toxicity of drug scaffold derivatives in rats / mice for non-commercial purposes.

[0191] The above advantages indicate that the model provided by the present invention can serve as an effective tool in drug R & D, regulatory decision-making, and risk assessment.

[0192] Table 5 Comparison of the QSAR models of the present invention with models in the existing literature

[0193]

[0194] The present invention selects six common drug scaffolds and collects the acute oral toxicity data of their molecular derivatives in rats and mice. On the basis of strictly following the OECD guidelines, interpretable, transferable, and reproducible QSAR models are provided and strictly statistically validated. The results confirm that all established QSAR models of the present invention exhibit good robustness and excellent external prediction performance.

[0195] In addition, the present invention combines the q-RASAR method and machine learning method to further improve the quality of the 2D-QSAR model and analyzes the toxicity mechanism of key descriptors in the optimal model. Finally, the optimal model is used to perform high-throughput toxicity prediction on real external compounds, and the external prediction quality is verified by the application domain and prediction reliability indicator methods.

[0196] The research results of the present invention show that the established model has wide applicability in filling data gaps and predicting the toxicity of new or unknown compounds. The model provided by the present invention provides a valuable tool for predicting the toxicity of various drug scaffold derivatives, helps drug discovery, improves the R & D efficiency, and supports the environmental monitoring and risk assessment of potentially highly toxic compounds.

Claims

1. A chemometric prediction model for predicting the acute oral toxicity of drug skeleton molecules, which is any one of the following formulas (1)-(12): When the drug skeleton molecule is a pyrazine drug skeleton molecule, it is the following formula (1) or formula (2): pLD 50 = 3.4823 - 0.227GATS3M - 0.6638(nCb-) - 0.7063C-005 + 0.9514F02[O-O] (1) pLD 50 = 1.5275 + 0.2929MATS7m + 0.3377GATS4p + 0.6792B01[C - O] - 0.9758B04[N - O] (2) When the drug skeleton molecule is a piperazine drug skeleton molecule, it is the following formula (3) or formula (4): pLD 50 = 1.8553 - 0.2479ATSC4i - 0.2787C - 025 - 0.2908GATS5s + 0.9064Uc (3) pLD 50 = 68.0917 - 0.5144SpDiam_X + 0.4637P_VSA_p_3 + 0.4761SHED_AL - 0.3586F09[N - O] - 0.5085F10[C - C] + 0.1252SHssNH (4) When the drug skeleton molecule is a thiazole drug skeleton molecule, it is the following formula (5) or formula (6): pLD 50 = 1.8183 + 0.3523MATS7m + 0.4705MATS1p + 0.4107P_VSA_charge_6 - 0.4928F02[O - O] + 0.6562F09[C - S] + 0.2496LOGP99(5) pLD 50 = 2.2533 + 0.3193nR06 - 0.3043GATS6m + 0.5062B02[N-N] + 0.3117Inflammat-80 + 0.3914s3_relPathLength - 0.2153nHBint - 0.2606SHBint5 (6) When the drug skeleton molecule is a thiophene drug skeleton molecule, it is the following formula (7) or formula (8): pLD 50 = 1.577 - 0.3599MATS4e - 0.289GATS5p + 0.5252GATS6i + 0.635P_VSA_s_1 (7) pLD 50 = 0.435 + 0.3796nHM - 0.1023MATS7s + 0.5386GATS1m + 0.2042B03[N-N] - 0.2415F09[C-N] + 0.1397Neoplastic-50 + 0.4879nBase (8) When the drug skeleton molecule is an indole drug skeleton molecule, it is the following formula (9) or formula (10): pLD 50 = 1.8762 + 0.4911H-052 + 0.4885MATS8m + 0.4926BCUTw-1h - 0.2225maxHCsats + 0.7426maxssssC (9) pLD 50 = 1.238 - 0.2741nCconj - 0.3191MaxsOH + 0.2356SHED_DL + 0.4495B04[O - O] + 0.2323LLS_02 + 0.6119VC - 5(10) When the drug skeleton molecule is a benzimidazole drug skeleton molecule, it is the following formula (11) or formula (12): pLD 50 = 2.0496 + 0.1485MATS4m + 0.1688MATS1s - 0.3895MATS4s - 0.6328minssssC - 0.283T(O..CI) + 0.1601sSCI (11) pLD 50 = 1.6587 + 0.2544O% + 0.3437GATS6i - 0.9299P_VSA_MR_7 + 0.9732P_VSA_charge_6 - 0.3227C - 044 + 0.4968CATS2D_04_DA (12) In each formula, pLD 50 represents the acute oral toxicity of the drug skeleton molecule, and the remaining parameters represent the 2D molecular descriptors of the drug skeleton molecule.

2. The application of the chemometric prediction model according to claim 1 in predicting the acute oral toxicity of drug skeleton molecules; The drug skeleton molecules include pyrazine drug skeleton molecules, piperazine drug skeleton molecules, thiazole drug skeleton molecules, thiophene drug skeleton molecules, indole drug skeleton molecules and benzimidazole drug skeleton molecules.

3. A method for predicting the acute oral toxicity of drug skeleton molecules, comprising the following steps: Obtain the 2D molecular descriptors of the drug skeleton molecules, and according to the chemometric prediction model of claim 1, obtain the acute oral toxicity of the drug skeleton molecules, that is, realize the prediction of the acute oral toxicity of the drug skeleton molecules.

4. The prediction method according to claim 3, wherein: Use PaDEL and Alvadesc software to obtain the 2D molecular descriptors of the drug skeleton molecules.

5. A prediction device for the acute oral toxicity of drug skeleton molecules applying the chemometric prediction model according to claim 1.

6. 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 acute oral toxicity as described in claim 3 or 4.

7. A computer storage medium, on which a computer program is stored, characterized in that: When the computer program is executed by the processor, it implements the method for predicting acute oral toxicity as described in claim 3 or 4.

Images