A method and system for interspecies toxicity prediction based on machine learning

Abstract

This invention relates to a machine learning-based method and system for predicting interspecific toxicity. The method acquires toxicity data of the same chemical to both alternative and predicted species, calculates its molecular descriptors, and preprocesses them to select descriptors suitable for modeling. Combining the toxicity data of the alternative species with the selected descriptors as independent variables and the toxicity data of the predicted species as the dependent variable, a machine learning algorithm is used to construct an interspecific toxicity relationship prediction model. The model's performance is then evaluated and validated through cross-validation. For a chemical to be predicted, only its toxicity data to an alternative species and related molecular descriptors need to be acquired. The validated final interspecific relationship prediction model is then used to predict the toxicity of the chemical to the predicted species. Compared to existing technologies, this invention utilizes a reliable model to efficiently and accurately predict the potential toxicity of a chemical to a target predicted species, effectively reducing the cost and time of traditional cross-species toxicity testing.

Description

Technical Field

[0001] This invention relates to the field of environmental toxicology and ecological risk assessment technology, specifically to a machine learning-based method and system for predicting interspecific biotoxicity. Background Technology

[0002] In the work of chemical ecological risk assessment and environmental baseline derivation, there is a common problem of lack of biotoxicity test data. In order to obtain the required toxicity data, a variety of model prediction methods have been developed, among which quantitative structure-activity relationship (QSAR) method and interspecific relationship prediction (ICE) method are two major technical approaches.

[0003] The QSAR method predicts the toxicity of unknown substances by establishing a quantitative relationship between the structural parameters of chemical substances and their biotoxic effects. For example, reference [1] discloses a freshwater acute baseline prediction method based on the quantitative structure-activity relationship of metals. This method predicts the toxic endpoint of unknown metals based on the quantitative relationship between the structural characteristics of heavy metal ions and the acute toxic effects on aquatic organisms, and derives the dangerous concentration for protecting different proportions of aquatic organisms by combining the sensitivity distribution analysis of different species. However, the limitation of the QSAR method is that it requires a large amount of experimental data of the species to be predicted when modeling. For species with less experimental data, the prediction effect is poor, there is a large uncertainty, and it is even impossible to complete the modeling work.

[0004] The ICE method employs a different approach, using existing species' toxicity data to establish a simple linear relationship to predict the toxicity of other species. The U.S. Environmental Protection Agency has developed the online ICE platform Web-ICE, which can be used to predict the toxicity of aquatic organisms and wildlife. Traditional ICE methods typically use linear regression models, employing only the toxicity data of surrogate species as independent variables for prediction. However, this method requires strict validation conditions, such as a cross-validation success rate ≥80%, mean squared error (MSE) ≤0.54, and R0.05. 2 ≥0.78, etc. For two species with large taxonomic differences, traditional linear ICE models often fail to meet the above validation criteria, resulting in insufficient predictive reliability.

[0005] Reference [2] discloses a system and method for predicting the chronic toxicity of bioaccumulative pollutants. This system acquires tissue residual toxicity data and standardizes it to construct a multi-dimensional feature matrix that includes the physicochemical properties of pollutants, biological characteristics of species, and experimental conditions. It then uses machine learning algorithms to optimize the toxicity prediction model. Although this method uses machine learning technology, it is mainly aimed at chronic toxicity and tissue residual data, which differs from the application scenario of acute toxicity prediction.

[0006] In summary, existing QSAR methods have limited predictive capabilities when experimental data on the target species are scarce, while traditional linear ICE methods often fail to achieve a cross-validation success rate below the standard requirement of 80% when dealing with species pairs with large taxonomic differences. These shortcomings fail to meet the growing demands for chemical eco-risk assessment and environmental benchmark derivation. Therefore, a new technical solution is needed to achieve higher accuracy in interspecific toxicity prediction under conditions of limited target species data.

[0007] Citations: [1] CN104820873A, A method for predicting acute baseline in freshwater based on quantitative structure-activity relationship of metals, announcement date December 26, 2017.

[0008] [2] CN120822659A, A system and method for predicting the chronic toxicity of bioaccumulated pollutants, published on October 21, 2025. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a machine learning-based method and system for predicting interspecific biotoxicity. This addresses the problems of the QSAR method having limited predictive ability when there is limited experimental data for the target species, and the traditional linear ICE method having a cross-validation success rate of less than 80% when dealing with species pairs with large taxonomic distances, thus failing to meet the needs of chemical ecological risk assessment and environmental benchmark derivation.

[0010] The objective of this invention can be achieved through the following technical solutions: The first aspect of this invention provides a machine learning-based method for predicting interspecific biotoxicity, comprising the following steps: S1. Obtain toxicity endpoint data for the same chemical to alternative and predicted species; S2. Obtain the structure information file of the chemical and calculate the molecular descriptor based on the structure information file; S3. Preprocess all the calculated molecular descriptors to select the molecular descriptors for modeling. S4. Based on the toxicity endpoint data of the alternative species and the screened molecular descriptors, an independent variable group is formed, and the toxicity endpoint data of the predicted species is used as the dependent variable. A machine learning method is used to establish an interspecific relationship prediction model, and the effectiveness of the interspecific relationship prediction model is evaluated using cross-validation. S5. For the chemical to be predicted, obtain its toxicity data and molecular descriptor to the alternative species, and use the validated final interspecific relationship prediction model to predict the toxicity of the chemical to be predicted to the predicted species.

[0011] Furthermore, S1 specifically includes the following steps: Collect publicly available acute or chronic toxicity endpoint data of the same chemical to alternative and predicted species from interspecific relationship prediction databases; After the data collection, a data cleaning step is performed, which includes removing data on mixtures, metals, salts, and ionic chemicals to ensure that the obtained toxicity endpoint data are all test data for a single organic chemical. The surrogate species are known species whose toxicity data are used as independent variables, and the predicted species are species whose toxicity data are used as dependent variables.

[0012] Furthermore, S2 specifically includes the following steps: Based on the chemicals identified in step S1, obtain or draw the two-dimensional or three-dimensional chemical structural formulas of the chemicals from public databases or professional software; Save the chemical structural formulas in a uniform format as a specified structured file. The structured file is input into molecular descriptor calculation software, which is configured to calculate at least a set of one-dimensional and two-dimensional molecular descriptors for the chemical to numerically characterize the structural features of the chemical.

[0013] Furthermore, in S3, all calculated molecular descriptors are preprocessed, specifically including the following sub-steps executed sequentially: Iterate through all the calculated molecular descriptors in S2, identify and delete descriptor variables that have null or invalid values ​​in all chemical samples; After null value removal, the variance of each variable is calculated in the descriptor variable set, and near-zero variance variables with variances below a preset threshold are removed to exclude features that are not discriminative. Calculate the pairwise linear correlation coefficients between the remaining molecular descriptor variables and remove variables with correlation coefficients greater than 0.9 with other variables to ensure that there is no high collinearity among the variables in the final set of independent variables for modeling.

[0014] Furthermore, in S3, the step of deleting variables with correlation coefficients greater than 0.9 with other variables adopts the following specific process: Calculate the Pearson correlation coefficient matrix between each pair of the remaining molecular descriptor variables; In the Pearson correlation coefficient matrix, for any pair of variables with a correlation coefficient greater than 0.9, one of the variables is deleted according to a preset variable importance ranking or a random principle, and the variable set is updated. Repeat this process until the Pearson correlation coefficient between any two variables is no greater than 0.9, thus completing the screening of highly correlated variables.

[0015] Furthermore, in S4, the machine learning method includes one of lightGBM, XGBoost, and CatBoost.

[0016] Furthermore, in S4, the method of evaluating the effectiveness of the interspecific relationship prediction model using cross-validation specifically involves using leave-one-out cross-validation to evaluate the predictive performance of the established model, and the performance is measured by the cross-validation success rate index.

[0017] Furthermore, in S4, the cross-validation success rate is calculated by comparing the model's predicted value for a single test sample with its actual measured value in each round of leave-one-out cross-validation. If the prediction error is less than a preset biologically reasonable threshold, the prediction is considered successful. Finally, the proportion of successful predictions in all rounds is statistically analyzed.

[0018] Furthermore, S5 specifically includes the following steps: For the target chemical to be predicted, its known toxicity data to the said alternative species are obtained through literature or experiments; Following the same procedure as S2, obtain the structural information file of the target chemical and calculate its molecular descriptor; Based on the preprocessing rules established in S3 and the standard that the final variable screening results are consistent, a subset of descriptors that completely correspond to the variables used in modeling are selected from the molecular descriptors of the target chemical. The toxicity data of the alternative species and the selected subset of descriptors are combined into an independent variable vector, which is then input into the final interspecific relationship prediction model trained and validated in step S4. The model outputs the predicted toxicity endpoint value of the target chemical on the predicted species.

[0019] A second aspect of the present invention provides a machine learning-based interspecific biotoxicity prediction system, comprising: The data acquisition and cleaning module acquires toxicity endpoint data for the same chemical on alternative and predicted species. The chemical structure processing and descriptor calculation module obtains the structure information file of the chemical substance and calculates the molecular descriptor based on the structure information file. The descriptor preprocessing and filtering module preprocesses all calculated molecular descriptors to filter out molecular descriptors for modeling. The model training and validation module uses the toxicity endpoint data of the alternative species and the screened molecular descriptors to form a set of independent variables, and uses the toxicity endpoint data of the predicted species as the dependent variable. It employs machine learning methods to establish an interspecific relationship prediction model and uses cross-validation to evaluate the effectiveness of the interspecific relationship prediction model. The toxicity prediction module acquires the toxicity data and molecular descriptors of the chemical to be predicted to the alternative species, and uses the validated final interspecific relationship prediction model to predict the toxicity of the chemical to be predicted to the predicted species.

[0020] Compared with the prior art, the present invention has the following beneficial effects: (1) By supplementing molecular descriptors as independent variables with the traditional method of using only surrogate species toxicity data, and applying machine learning algorithms capable of handling nonlinear relationships, the performance of the interspecific relationship prediction model was significantly improved, achieving a cross-validation success rate of 82.53%-87.50%, which is 10-18 percentage points higher than the 64.52%-79.21% of the traditional linear model. This meets the EPA Web-ICE validation standards (cross-validation success rate ≥80%, MSE ≤0.54, R... 2 ≥0.78); (2) R values ​​were obtained from the large Daphnia-mysis species pair using the lightGBM method. 2 =0.8939, MSE=0.33, and the R² values ​​were obtained using the CatBoost method on the rainbow trout-sheephead sardine species pair. 2 =0.8439 and MSE=0.20, both of which meet the EPA standard, enabling the establishment of reliable interspecific relationship prediction models for more species pairs with large biological taxonomic differences, and providing more predictive data on the biotoxicity of species for ecological risk assessment and environmental benchmark derivation; (3) Feature importance analysis shows that the toxicity of the alternative species contributes 60-70% of the importance, and the supplementary molecular descriptor contributes 30-40% of the importance. The combination of the two using nonlinear machine learning methods produces a synergistic effect, which can handle more complex nonlinear relationships and the prediction accuracy is significantly better than the traditional simple linear relationship model. (4) The method is simple and easy to use, using widely available software tools (PaDEL-Descriptor, R / Python), with low requirements for computer hardware, which can be implemented on common personal computers. No machine learning expertise is required, making it easy for environmental toxicology workers to learn and highly practical. (5) By using alternative species data to predict the toxicity of target species, the direct testing requirements for target species are reduced, the number of experimental animals used is reduced, which has ethical advantages. At the same time, it accelerates the process of chemical ecological risk assessment and environmental benchmark derivation, and can be widely applied in disciplines such as environmental ecology and toxicology. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a machine learning-based method for predicting interspecific relationships. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Component models, material names, connection structures, control methods, algorithms, and other features not explicitly described in this technical solution are considered common technical features disclosed in the prior art.

[0023] The specific embodiments of the present invention will be described in detail with reference to the accompanying drawings. It should be noted that the terms used in the following embodiments are defined as follows: Cross-validation success rate refers to the percentage of samples in leave-one-out cross-validation where the deviation between the predicted and actual values ​​is within 10 times the total sample size; Mean squared error (MSE) refers to the sum of the squares of the logarithmic differences between the predicted and actual values ​​divided by the number of samples; R... 2 The coefficient of determination indicates how well the model fits the data; the molecular descriptor refers to numerical parameters calculated based on the chemical structure, used to quantitatively describe the physicochemical properties, topological structure, and electronic characteristics of a molecule. (See attached figure.) Figure 1 This is a flowchart illustrating the method of the present invention, showing the complete process from data collection, molecular descriptor calculation, preprocessing, machine learning modeling, model validation to toxicity prediction.

[0024] Example 1 In this embodiment, the interspecific biotoxicity prediction method based on machine learning is described in [reference needed]. Figure 1 This includes the following steps: S1. Obtain toxicity endpoint data for the same chemical to alternative and predicted species; In specific implementation, S1 includes the following steps: Collect publicly available acute or chronic toxicity endpoint data of the same chemical to alternative and predicted species from interspecific relationship prediction databases; After the data collection, a data cleaning step is performed, which includes removing data on mixtures, metals, salts, and ionic chemicals to ensure that the obtained toxicity endpoint data are all test data for a single organic chemical. The surrogate species are known species whose toxicity data are used as independent variables, and the predicted species are species whose toxicity data are used as dependent variables.

[0025] S2. Obtain the structure information file of the chemical and calculate the molecular descriptor based on the structure information file; In specific implementation, S2 includes the following steps: Based on the chemicals identified in step S1, obtain or draw the two-dimensional or three-dimensional chemical structural formulas of the chemicals from public databases or professional software; Save the chemical structural formulas in a uniform format as a specified structured file. The structured file is input into molecular descriptor calculation software, which is configured to calculate at least a set of one-dimensional and two-dimensional molecular descriptors for the chemical to numerically characterize the structural features of the chemical.

[0026] S3. Preprocess all the calculated molecular descriptors to select the molecular descriptors for modeling. In specific implementation, S3 involves preprocessing all calculated molecular descriptors, specifically including the following sub-steps executed sequentially: Iterate through all the calculated molecular descriptors in S2, identify and delete descriptor variables that have null or invalid values ​​in all chemical samples; After null value removal, the variance of each variable is calculated in the descriptor variable set, and near-zero variance variables with variances below a preset threshold are removed to exclude features that are not discriminative. Calculate the pairwise linear correlation coefficients between the remaining molecular descriptor variables and remove variables with correlation coefficients greater than 0.9 with other variables to ensure that there is no high collinearity among the variables in the final set of independent variables for modeling.

[0027] In specific implementation, the step of deleting variables with a correlation coefficient greater than 0.9 with other variables in S3 adopts the following specific process: Calculate the Pearson correlation coefficient matrix between each pair of the remaining molecular descriptor variables; In the Pearson correlation coefficient matrix, for any pair of variables with a correlation coefficient greater than 0.9, one of the variables is deleted according to a preset variable importance ranking or a random principle, and the variable set is updated. Repeat this process until the Pearson correlation coefficient between any two variables is no greater than 0.9, thus completing the screening of highly correlated variables.

[0028] S4. Based on the toxicity endpoint data of the alternative species and the screened molecular descriptors, an independent variable group is formed, and the toxicity endpoint data of the predicted species is used as the dependent variable. A machine learning method is used to establish an interspecific relationship prediction model, and the effectiveness of the interspecific relationship prediction model is evaluated using cross-validation. In specific implementation, in S4, the machine learning method includes one of lightGBM, XGBoost, and CatBoost.

[0029] In specific implementation, in S4, the method of using cross-validation to evaluate the effectiveness of the interspecific relationship prediction model specifically involves using leave-one-out cross-validation to evaluate the predictive performance of the established model, and the performance is measured by the cross-validation success rate index.

[0030] In specific implementation, in S4, the cross-validation success rate is calculated as follows: in each round of leave-one-out cross-validation, the model's predicted value for a single test sample is compared with its actual measured value. If the prediction error is less than a preset biologically reasonable threshold, the prediction is considered successful. Finally, the proportion of successful predictions in all rounds is statistically analyzed.

[0031] S5. For the chemical to be predicted, obtain its toxicity data and molecular descriptor to the alternative species, and use the validated final interspecific relationship prediction model to predict the toxicity of the chemical to be predicted to the predicted species.

[0032] In practice, for the target chemical to be predicted, known toxicity data to the alternative species are obtained through literature or experiments. Following the same procedure as S2, obtain the structural information file of the target chemical and calculate its molecular descriptor; Based on the preprocessing rules established in S3 and the standard that the final variable screening results are consistent, a subset of descriptors that completely correspond to the variables used in modeling are selected from the molecular descriptors of the target chemical. The toxicity data of the alternative species and the selected subset of descriptors are combined into an independent variable vector, which is then input into the final interspecific relationship prediction model trained and validated in step S4. The model outputs the predicted toxicity endpoint value of the target chemical on the predicted species.

[0033] In summary, the principle of this embodiment lies in establishing a quantitative extrapolation relationship of chemical toxicity between alternative and predicted species through machine learning technology, thereby achieving cross-species toxicity prediction. The approach is as follows: first, collect the toxicity endpoint data of the same series of chemicals on two types of species and calculate diverse molecular descriptors for each chemical to comprehensively characterize its structural features; then, preprocess and screen the massive descriptors to remove redundancy and noise and retain the subset of key features most relevant to the toxicity mechanism; then, combine the actual toxicity data of the alternative species with the screened molecular descriptors to form a multidimensional set of independent variables, and use the corresponding toxicity data of the predicted species as the dependent variable target, and use machine learning algorithms to train a mathematical model that can profoundly capture the complex nonlinear laws of toxicity transmission between species.

[0034] During the modeling process, cross-validation was used to repeatedly evaluate and optimize the model to ensure its stability and generalization ability and prevent overfitting. The final validation model essentially internalized the mapping function from the chemical's structural characteristics and the toxic response of alternative species to the toxic endpoint of the target species. Therefore, for a novel chemical to be predicted, only its basic molecular structure information and the toxicity test results of alternative species need to be input into the model to automatically, quickly and accurately infer its potential toxicity to the target species that is difficult to determine experimentally. This significantly reduces the dependence and cost of direct toxicity testing of rare or protected species in ecological risk assessment.

[0035] Application Example 1 This application example provides a method for predicting interspecific relationships using datasets of the freshwater crustacean Daphnia magna and the marine crustacean Americamysis bahia.

[0036] Acute toxicity data of the same chemical on the freshwater crustacean *Daphnia magna* and the marine crustacean *Americamysis bahia* were collected from the Web-ICE website (248 sets). The collected data were screened to exclude mixtures, elemental metals, salts, and ionic compounds, leaving 229 valid data sets. *Americamysis bahia* was designated as the predicted species, and *Daphnia magna* as the surrogate species. Toxicity data for the predicted species were used as the dependent variable (Y), and toxicity data for the surrogate species were used as the independent variable (X). Toxicity data were expressed as the half-maximal effective concentration (EC50). 50 The value is expressed in μg / L and is used for modeling after taking the logarithm.

[0037] The chemical structure files of 229 chemicals were searched and downloaded from the ChemSpider chemical structure database in .mol ​​format. .mol ​​format files contain atomic coordinates, chemical bond types, and stereochemical information for the chemicals, and can be directly read by molecular descriptor calculation software.

[0038] Molecular descriptors were calculated using PaDEL-Descriptor software. Using 229 .mol ​​files as input, both one-dimensional (1D) and two-dimensional (2D) molecular descriptors were calculated, resulting in 1444 sets of molecular descriptors. One-dimensional descriptors included parameters reflecting the overall properties of the molecule, such as molecular weight, number of atoms, number of chemical bonds, number of hydrogen bond donors, number of hydrogen bond acceptors, number of rotatable chemical bonds, and number of aromatic rings. Two-dimensional descriptors included parameters reflecting the molecular topology and electronic distribution, such as topological indices, connectivity indices, shape indices, and charge distribution indices.

[0039] The 1444 molecular descriptors underwent preprocessing. The preprocessing consisted of three steps: First, variables containing null values ​​were removed. Null values ​​indicate that the software could not calculate specific descriptors for certain chemicals, leading to modeling failure. Second, near-zero variance variables were removed. Near-zero variance variables are those that take almost identical values ​​across all samples; these variables contain no useful information and increase model complexity. Third, the Pearson correlation coefficients between all variables were calculated, and one variable in each pair with a correlation coefficient greater than 0.9 was removed. This step was used to eliminate multicollinearity and prevent highly correlated variables from repeatedly providing the same information. After these three preprocessing steps, the number of molecular descriptors was reduced from 1444 to 501.

[0040] A predictive model was built using the lightGBM machine learning algorithm. lightGBM is a gradient boosting decision tree-based machine learning algorithm capable of handling high-dimensional data and nonlinear relationships. 501 preprocessed molecular descriptors and toxicity data X of the substitute species *Daphnia magna* were used as independent variables, and toxicity data Y of the predicted species *Americamysis bahia* was used as the dependent variable. Leave-one-out cross-validation was employed to evaluate model performance. The leave-one-out cross-validation process involved repeatedly selecting one sample from 229 samples as the test set, using the remaining 228 samples as the training set, building the model, and predicting the toxicity value of the test set samples. This process was repeated 229 times until every sample was predicted once. The deviation between the predicted and actual values ​​was calculated; a prediction was considered successful when the deviation was within 10 times. The cross-validation success rate was the number of successfully predicted samples divided by the total number of samples.

[0041] As shown in Table 1, the leave-one-out cross-validation results of the lightGBM algorithm demonstrate excellent model performance.

[0042] Table 1: Comparison of leave-one-out cross-validation results of the ICE model for Daphnia magna and Americamysis bahia. Note: R 2 The coefficient of determination is denoted as , MSE is the mean squared error, and CV success rate is the cross-validation success rate.

[0043] As can be seen from Table 1, the R of the lightGBM algorithm 2A cross-validation score of 0.8939 indicates that the model can explain 89.39% of the toxicity data variation, demonstrating a good fit. The mean squared error (MSE) is 0.33, lower than the EPA Web-ICE standard requirement of 0.54, indicating a small prediction error. The cross-validation success rate is 82.53%, higher than the EPA Web-ICE standard requirement of 80%, meeting the conditions for interspecific relationship prediction. In contrast, traditional linear models using only the surrogate species toxicity X as a single independent variable have a cross-validation success rate of only 64.52%, failing to meet the 80% standard and thus unsuitable for interspecific relationship prediction. The lightGBM algorithm, by supplementing the independent variable with molecular descriptors and utilizing gradient boosting decision trees to handle nonlinear relationships, improves the cross-validation success rate by 18.01 percentage points, representing a performance improvement of 27.9%.

[0044] The mechanism behind the improved cross-validation success rate lies in the fact that traditional linear models assume a simple linear relationship between the toxicity of the surrogate species and the toxicity of the predicted species, i.e., Y=aX+b, which can only capture the overall trend of toxicity among species. However, in reality, the differences in toxicity of different chemicals to different species are influenced by various factors such as chemical structure, mechanism of action, and metabolic pathway, resulting in complex nonlinear characteristics in the toxicity relationship among species. The lightGBM algorithm, by introducing molecular descriptors, incorporates the structural information of chemicals into the model, enabling it to identify the toxicity difference patterns of chemicals with different structural types. For example, for aromatic compounds, which are highly hydrophobic and have a high accumulation capacity in lipid membranes, their toxicity to crustaceans is mainly generated through membrane perturbation mechanisms; in this case, the octanol-water partition coefficient LogP is an important predictor. For organophosphate compounds, their toxicity is mainly generated by inhibiting the production of acetylcholinesterase; in this case, the electron density distribution and stereoconformation of the molecule are important predictors. The decision tree structure of the lightGBM algorithm can automatically learn these complex structure-toxicity relationships. Through the combination of multi-layer decision trees, the chemical space is divided into multiple sub-regions, and local linear relationships are established within each sub-region, thereby achieving accurate fitting of the global nonlinear relationship.

[0045] As shown in Table 2, the variable importance analysis reveals the contribution of each variable to the prediction results.

[0046] Table 2: The 10 most important variables Note: Variable importance is calculated using the gain method, representing the cumulative contribution of a variable to the improvement of model performance across all decision tree splits. As shown in Table 2, toxicity of *Daphnia magna* is the most important variable, with an importance of 0.7008, accounting for 70.08% of the total importance. This indicates that toxicity of alternative species remains the primary basis for prediction, which aligns with the basic principles of interspecific relationship prediction. The supplementary molecular descriptions contribute 29.92% of the total importance. Specifically, *SpMin4_Bhi* contributes 4.45%, representing the fourth smallest Burden eigenvalue in the molecule, reflecting the charge distribution and polarization of the molecule, which is related to the compound's partitioning behavior in biomembranes. *VR1_Dzm* contributes 1.58%, representing the Randic-type topological index, reflecting the degree of branching in the molecule, which is related to the compound's bioavailability. *CrippenLogP* contributes 0.66%, representing the octanol-water partition coefficient calculated by the Crippen method, a classic parameter for assessing compound hydrophobicity. Strongly hydrophobic compounds are more likely to accumulate in lipid tissues and are more toxic to aquatic organisms. The addition of molecular descriptors enables the model to identify the moderating effect of chemical structures on interspecies toxicity differences, thereby improving prediction accuracy.

[0047] Based on cross-validation results, the lightGBM model meets the conditions for predicting interspecies relationships. The final prediction model was trained using all 229 datasets and the optimal hyperparameters determined through cross-validation. The optimal hyperparameters included: learning rate 0.05, maximum tree depth 6, minimum number of samples per leaf node 20, feature sampling ratio 0.8, and sample sampling ratio 0.8. The final model can be used to predict the toxicity of other chemicals to the mysid shrimp *Americamysis bahia*.

[0048] The prediction process is demonstrated using N,N-diethylcyclohexylamine as an example. The CAS number for N,N-diethylcyclohexylamine is 91-65-6, and its molecular formula is C2. 10 H 21 N is an aliphatic tertiary amine compound. First, acute toxicity data of N,N-diethylcyclohexylamine to the substituted species *Daphnia magna* were obtained, and EC50 data were obtained through literature search. 50 The concentration was 5756.23 μg / L, and the logarithm of the lgEC was calculated. 50 The value was 3.7604. Then, the chemical structure formula mol file of N,N-diethylcyclohexylamine was downloaded from the ChemSpider database, and 1D and 2D molecular descriptors were calculated using PaDEL-Descriptor software. From the calculation results, 501 variables were selected for use in the final model. Alternative species toxicity data (lgEC) were then used. 50 =3.7604 and 501 molecular descriptors were input into the final prediction model to obtain the predicted toxicity value lgEC for the mysid shrimp Americamysis bahia. 50=4.5209. Convert the logarithm to the original concentration, EC... 50 =33.18 mg / L. The prediction results indicate that the acute toxicity of N,N-diethylcyclohexylamine to mysid shrimp (Americamysis bahia) is 33.18 mg / L, which belongs to the low toxicity class of compounds.

[0049] Application Example 2 This embodiment provides a method for predicting interspecific relationships using datasets of freshwater rainbow trout (Oncorhynchus mykiss) and marine fish (Cyprinodon variegatus).

[0050] Acute toxicity data of the same chemical on the freshwater fish *Oncorhynchus mykiss* and the marine fish *Cyprinodon variegatus* were collected from the Web-ICE website (101 sets of data). The collected data were screened to exclude mixtures, elemental metals, salt compounds, and ionic compounds, leaving 96 valid data sets. *Cyprinodon variegatus* was designated as the predictive species, and *Oncorhynchus mykiss* as the surrogate species. Toxicity data for the predictive species were used as the dependent variable (Y), and toxicity data for the surrogate species were used as the independent variable (X). Toxicity data were expressed as median lethal concentration (LC50). 50 The value is expressed in μg / L and is used for modeling after taking the logarithm.

[0051] Searched and downloaded the chemical structure files of 96 chemicals from the ChemSpider chemical structure database. The files are in .mol ​​format.

[0052] Molecular descriptors were calculated using PaDEL-Descriptor software. Using 96 mol files as input, both 1D and 2D molecular descriptors were calculated, resulting in a total of 1444 molecular descriptors.

[0053] The 1444 molecular descriptors were preprocessed. The preprocessing consisted of three steps: first, removing variables containing null values; second, removing variables with near-zero variance; and third, calculating the Pearson correlation coefficients among all variables and removing one variable from each pair with a correlation coefficient greater than 0.9. After these three preprocessing steps, the number of molecular descriptors was reduced from 1444 to 500.

[0054] A predictive model was built using the CatBoost machine learning algorithm. CatBoost is a gradient boosting decision tree-based machine learning algorithm that performs well when handling categorical features and small sample data. 500 preprocessed molecular descriptors and toxicity data X of the substitute species rainbow trout (Oncorhynchus mykiss) were used as independent variables, and toxicity data Y of the predicted species (Cyprinodon variegatus) were used as the dependent variable. Leave-one-out cross-validation was used to evaluate model performance. The leave-one-out cross-validation process involved selecting one sample from 96 samples as the test set, using the remaining 95 samples as the training set, building the model, and predicting the toxicity value of the test set samples. This process was repeated 96 times until each sample was predicted once.

[0055] As shown in Table 3, the leave-one-out cross-validation results of the CatBoost algorithm demonstrate excellent model performance.

[0056] Table 3: Comparison of Leave-one-out cross-validation results of the ICE model for rainbow trout (Oncorhynchus mykiss) and mutton (Cyprinodon variegatus). Note: R 2 The coefficient of determination is denoted as , MSE is the mean squared error, and the CV success rate is the cross-validation success rate. As can be seen from Table 3, the R of the CatBoost algorithm is... 2 A cross-validation score of 0.8439 indicates that the model can explain 84.39% of the toxicity data variation, demonstrating a good fit. The mean squared error (MSE) is 0.20, lower than the EPA Web-ICE standard requirement of 0.54, indicating a small prediction error. The cross-validation success rate is 87.50%, higher than the EPA Web-ICE standard requirement of 80%, meeting the conditions for predicting interspecific relationships. In contrast, the traditional linear model has a cross-validation success rate of 79.21%, failing to meet the 80% standard and therefore unsuitable for predicting interspecific relationships. The CatBoost algorithm, by supplementing molecular descriptors as independent variables, improves the cross-validation success rate by 8.29 percentage points, representing a performance improvement of 10.5%.

[0057] The mechanism behind the improved cross-validation success rate lies in the fact that rainbow trout (Oncorhynchus mykiss) and sardine fish (Cyprinodon variegatus) belong to freshwater and marine fish, respectively. These differences in physiological characteristics, metabolic enzyme systems, and osmolarity regulation mechanisms lead to varying toxic responses to the same chemical. Traditional linear models cannot capture the impact of these physiological differences on toxicity, while the CatBoost algorithm, by introducing molecular descriptors, can identify the correlation between chemical structure and species-specific toxicity. For example, for lipid-soluble compounds, their enrichment in fish adipose tissue is related to the molecule's hydrophobicity; in this case, the octanol-water partition coefficient LogP is an important predictor. The decision tree structure of the CatBoost algorithm can automatically learn these chemical structure-species-specific toxicity relationships. Through ordered boosting and a symmetric tree structure, it reduces the risk of overfitting and improves the predictive stability for small sample data.

[0058] As shown in Table 4, the variable importance analysis reveals the contribution of each variable to the prediction results.

[0059] Table 4: The 10 most important variables Note: Variable importance is calculated using the gain method, representing the cumulative contribution of a variable to the improvement of model performance across all decision tree splits. Table 4 shows that the toxicity of rainbow trout (Oncorhynchus mykiss) is the most important variable, with an importance of 0.6073, accounting for 60.73% of the total importance, indicating that the toxicity of alternative species is the main basis for prediction. The supplementary molecular descriptors contribute 39.27% ​​of the importance, with VP.1 contributing 7.45%. This descriptor represents the Kier flexibility index of the molecule, reflecting the conformational flexibility of the molecule and its ability to bind to biomacromolecules. LipoaffinityIndex contributes 1.37%. This descriptor represents the lipid affinity index of the molecule, reflecting the molecule's partitioning tendency in the lipid environment and its bioaccumulation. MLFER_BH contributes 0.90%. This descriptor represents the Abraham hydrogen bond basicity parameter, reflecting the molecule's ability to act as a hydrogen bond acceptor and its solubility and membrane permeability. The addition of molecular descriptors enables the model to identify the regulatory role of chemical structure in the differences in toxicity between freshwater and marine fish, thereby improving prediction accuracy.

[0060] Based on the cross-validation results, the CatBoost model meets the conditions for predicting interspecies relationships. The final prediction model was trained using all 96 datasets and the optimal hyperparameters determined through cross-validation. The optimal hyperparameters included: a learning rate of 0.03, a maximum tree depth of 5, a minimum number of samples per leaf node of 15, and an L2 regularization coefficient of 3. The final model can be used to predict the toxicity of other chemicals to *Cyprinodon variegatus*.

[0061] The prediction process is demonstrated using 7A-ethyldihydro-1H,3H,5H-oxazolo[3,4-c]oxazole as an example. The CAS number for 7A-ethyldihydro-1H,3H,5H-oxazolo[3,4-c]oxazole is 7747-35-5, and its molecular formula is C6H4H2O. 11 NO2 is a heterocyclic compound. First, acute toxicity data of 7A-ethyldihydro-1H,3H,5H-oxazolo[3,4-c]oxazol against the alternative species rainbow trout (Oncorhynchus mykiss) were obtained, and LC-values ​​were obtained through literature search. 50 The concentration is 240,000 μg / L. After taking the logarithm, the lgLC ratio is calculated. 50 The value was 5.3802. Then, the chemical structural formula mol file of 7A-ethyldihydro-1H,3H,5H-oxazolo[3,4-c]oxazole was downloaded from the ChemSpider database. 1D and 2D molecular descriptors were calculated using PaDEL-Descriptor software, and 500 variables were selected from the calculation results for the final model. Alternative species toxicity data were then analyzed using lgLC. 50 =5.3802 and 500 molecular descriptors were input into the final prediction model to obtain the predicted toxicity value lgLC for the species Cyprinodon variegatus. 50 =4.4903. Convert the logarithm to the original concentration, LC... 50 = 30.92 mg / L. Predictive results indicate that 7A-ethyldihydro-1H,3H,5H-oxazolo[3,4-c]oxazol is effective against Cyprinodon dactylus. variegatus lgLC 50 =4.4903, converted to mg / L, then LC 50 =30.92mg / L.

[0062] Comparative Example 1 In the prior art, CN120581096A discloses a method for predicting the toxicity of rare and endangered organisms based on machine learning algorithms and quantitative structure-activity relationships.

[0063] The core concept of this invention is to solve the challenge of toxicity prediction for species with scarce data or difficult-to-experiment species by establishing a toxicity correlation model between alternative species and the target predicted species. The mechanism lies in the fact that it does not directly establish a traditional structure-activity relationship model for the predicted species, but instead introduces a data-rich alternative species as an information bridge. This method systematically collects paired toxicity data of the same chemical on two species, and uses the measured toxicity value of the alternative species and the chemical molecular descriptor together as input features for a machine learning model to predict the toxicity value of the target species. The core principle of model learning is how the chemical structure modulates the quantitative correlation between the toxicity responses of the two species, thereby achieving indirect and reliable extrapolation to the toxicity of the target species through readily available experimental data from the alternative species. This is a two-level prediction mechanism based on data transfer and interspecific relationship learning.

[0064] In contrast to existing technology CN120581096A, its concept is to develop a dedicated toxicity prediction model for a specific group of rare and endangered species. Its mechanism is an extension of traditional quantitative structure-activity relationship (QSAR) modeling. It directly utilizes limited toxicity data about the rare and endangered species themselves, matching it with molecular descriptors of chemicals, and establishing a direct mathematical mapping from chemical structure to the toxic endpoint of that specific organism through machine learning algorithms. The entire process aims to optimize and validate the direct predictive ability for the target organism itself. Its model input is only chemical descriptors, without involving toxicity data from any other species, and it does not explore or utilize interspecies toxicity correlations, fundamentally different from the cross-species extrapolation mechanism adopted in this invention.

[0065] The above description of the embodiments is provided to enable those skilled in the art to understand and use the invention. It will be apparent to those skilled in the art that various modifications can be made to these embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the invention should be within the protection scope of the present invention.

Claims

1. A machine learning-based method for predicting interspecific biotoxicity, characterized in that, Includes the following steps: S1. Obtain toxicity endpoint data for the same chemical to alternative and predicted species; S2. Obtain the structure information file of the chemical and calculate the molecular descriptor based on the structure information file; S3. Preprocess all the calculated molecular descriptors to select the molecular descriptors for modeling. S4. Based on the toxicity endpoint data of the alternative species and the screened molecular descriptors, an independent variable group is formed, and the toxicity endpoint data of the predicted species is used as the dependent variable. A machine learning method is used to establish an interspecific relationship prediction model, and the effectiveness of the interspecific relationship prediction model is evaluated using cross-validation. S5. For the chemical to be predicted, obtain its toxicity data and molecular descriptor to the alternative species, and use the validated final interspecific relationship prediction model to predict the toxicity of the chemical to be predicted to the predicted species.

2. The interspecific biotoxicity prediction method based on machine learning according to claim 1, characterized in that, S1 specifically includes the following steps: Collect publicly available acute or chronic toxicity endpoint data of the same chemical to alternative and predicted species from interspecific relationship prediction databases; After the data is collected, a data cleaning step is performed, which includes removing data about mixtures, metals, salts, and ionic chemicals. The surrogate species are known species whose toxicity data are used as independent variables, and the predicted species are species whose toxicity data are used as dependent variables.

3. The interspecific biotoxicity prediction method based on machine learning according to claim 1, characterized in that, S2 specifically includes the following steps: Based on the chemicals identified in step S1, obtain or draw the two-dimensional or three-dimensional chemical structural formulas of the chemicals from public databases or professional software; Save the chemical structural formulas in a uniform format as a specified structured file. The structured file is input into molecular descriptor calculation software, which is configured to calculate at least a set of one-dimensional and two-dimensional molecular descriptors for the chemical to numerically characterize the structural features of the chemical.

4. The interspecific biotoxicity prediction method based on machine learning according to claim 1, characterized in that, In S3, all calculated molecular descriptors are preprocessed, specifically including the following sub-steps executed sequentially: Iterate through all the calculated molecular descriptors in S2, identify and delete descriptor variables that have null or invalid values ​​in all chemical samples; After null value removal, the variance of each variable is calculated in the descriptor variable set, and near-zero variance variables with variances below a preset threshold are removed to exclude features that are not discriminative. Calculate the pairwise linear correlation coefficients between the remaining molecular descriptor variables and remove variables with correlation coefficients greater than 0.9 with other variables to ensure that there is no high collinearity among the variables in the final set of independent variables for modeling.

5. The interspecific biotoxicity prediction method based on machine learning according to claim 4, characterized in that, In S3, the steps for deleting variables with a correlation coefficient greater than 0.9 with other variables are as follows: Calculate the Pearson correlation coefficient matrix between each pair of the remaining molecular descriptor variables; In the Pearson correlation coefficient matrix, for any pair of variables with a correlation coefficient greater than 0.9, one of the variables is deleted according to a preset variable importance ranking or a random principle, and the variable set is updated. Repeat this process until the Pearson correlation coefficient between any two variables is no greater than 0.9, thus completing the screening of highly correlated variables.

6. The interspecific biotoxicity prediction method based on machine learning according to claim 1, characterized in that, In S4, the machine learning method includes one of lightGBM, XGBoost, and CatBoost.

7. The interspecific toxicity prediction method based on machine learning according to claim 4, characterized in that, In S4, the method of evaluating the effectiveness of the interspecific relationship prediction model using cross-validation specifically involves using leave-one-out cross-validation to evaluate the predictive performance of the established model, and the performance is measured by the cross-validation success rate index.

8. The interspecific biotoxicity prediction method based on machine learning according to claim 7, characterized in that, In S4, the cross-validation success rate is calculated as follows: in each round of leave-one-out cross-validation, the model's predicted value for a single test sample is compared with its actual measured value. If the prediction error is less than a preset biologically reasonable threshold, the prediction is considered successful. Finally, the proportion of successful predictions in all rounds is statistically analyzed.

9. The interspecific biotoxicity prediction method based on machine learning according to claim 7, characterized in that, S5 specifically includes the following steps: For the target chemical to be predicted, its known toxicity data to the said alternative species are obtained through literature or experiments; Following the same procedure as S2, obtain the structural information file of the target chemical and calculate its molecular descriptor; Based on the preprocessing rules established in S3 and the standard that the final variable screening results are consistent, a subset of descriptors that completely correspond to the variables used in modeling are selected from the molecular descriptors of the target chemical. The toxicity data of the alternative species and the selected subset of descriptors are combined into an independent variable vector, which is then input into the final interspecific relationship prediction model trained and validated in step S4. The model outputs the predicted toxicity endpoint value of the target chemical on the predicted species.

10. A machine learning-based interspecific biotoxicity prediction system, comprising: The data acquisition and cleaning module acquires toxicity endpoint data for the same chemical on alternative and predicted species. The chemical structure processing and descriptor calculation module obtains the structure information file of the chemical substance and calculates the molecular descriptor based on the structure information file. The descriptor preprocessing and filtering module preprocesses all calculated molecular descriptors to filter out molecular descriptors for modeling. The model training and validation module uses the toxicity endpoint data of the alternative species and the screened molecular descriptors to form a set of independent variables, and uses the toxicity endpoint data of the predicted species as the dependent variable. It employs machine learning methods to establish an interspecific relationship prediction model and uses cross-validation to evaluate the effectiveness of the interspecific relationship prediction model. The toxicity prediction module acquires the toxicity data and molecular descriptors of the chemical to be predicted to the alternative species, and uses the validated final interspecific relationship prediction model to predict the toxicity of the chemical to be predicted to the predicted species.

Images