A method for predicting bioconcentration factor content using qicar modeling

Abstract

The present application relates to the technical field of ecological risk assessment test strategy, in particular to a method for predicting the content of biological enrichment factor by using QICAR modeling. The present application uses known BCF, metal physicochemical properties and corresponding regional soil physicochemical properties as data, uses the training of QICAR coupled machine learning model to obtain the prediction value of unknown heavy metal BCF in the target region, and screens the test set data through multiple rounds of cross-validation 2 and average absolute error MAE to evaluate the performance of the model to obtain the optimal model, and then evaluates the result through feature analysis and Shapley weighted explanation to obtain accurate and scientific prediction value, which provides effective support for the prediction of heavy metal enrichment capacity of crops and guarantees food safety.

Description

Technical Field

[0001] This invention relates to the field of ecological risk assessment testing strategies, specifically a method for predicting the content of bioaccumulation factors using QICAR modeling. Background Technology

[0002] Heavy metals, as typical persistent pollutants in agricultural soils, pose a serious threat to global food security due to their toxicity, widespread presence, non-biodegradability, and bioavailability. Rice, in particular, is a major route of human exposure to heavy metals; its grains readily absorb accumulated heavy metals from the soil through their roots, directly impacting human health.

[0003] Bioaccumulation factor (BCF) is a key parameter for assessing the ability of rice to accumulate heavy metals, representing the ratio of heavy metal content in organisms to that in the environment. However, traditional measurement methods require detecting pollutant concentrations in both organisms and the environment, which is time-consuming and costly, making it difficult to meet the needs of large-scale and rapid assessment. While stepwise multiple linear regression in existing technologies quantifies bioaccumulation capacity by constructing linear models of BCF with soil pH, soil organic matter (SOM), and clay content, it often focuses on single heavy metals and lacks models applicable to predicting multiple heavy metals.

[0004] Quantitative structure-activity relationship (QSAR) models have been widely used for BCF prediction of organic compounds, such as models using the octanol-water partition coefficient logKow as the key indicator, but their application to heavy metals remains inconclusive. While quantitative ion characteristic-activity relationship (QICAR) models have been applied in research on heavy metal ecological risks in water and soil environments, their application in BCF prediction of crops is still lacking.

[0005] Based on the above reasons, this invention designs a method for predicting the content of bioaccumulation factors using QICAR modeling. Known BCF data, metal physicochemical properties, and soil physicochemical properties of the corresponding region are used to obtain predicted values ​​of unknown bioaccumulation factors by coupling three machine learning models with QICAR. The optimal results are obtained by combining multiple rounds of cross-validation screening, thereby making accurate and scientific predictions of the bioaccumulation factors of crops in unknown regions. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for predicting the content of bioaccumulation factors using QICAR modeling. Known BCF data, metal physicochemical properties, and soil physicochemical properties of the corresponding region are used to obtain predicted values ​​of unknown bioaccumulation factors by coupling three machine learning models with QICAR. The optimal results are obtained by combining multiple rounds of cross-validation screening, thereby making accurate and scientific predictions of the bioaccumulation factors of crops in unknown regions.

[0007] To achieve the above objectives, this invention provides a method for predicting the content of bioaccumulation factors using QICAR modeling, comprising the following steps:

[0008] S1. Input 26 feature factors into the prediction model to obtain the feature importance ranking, and arrange the physicochemical properties of 23 metals from high to low according to the importance score;

[0009] S2, the arranged metal physicochemical features are cyclically input into the prediction model in an n+1 manner. Each feature subset includes soil physicochemical properties and metal physicochemical properties.

[0010] S3 uses a network search method with 5-fold cross-validation for each feature subset to screen the optimal combination of parameters and feature subsets by minimizing the negative mean square error.

[0011] S4. The data is randomly divided into training and test sets in a 4:1 ratio. Based on the selected optimal feature subset and the optimized hyperparameters, the prediction model is retrained on the training set. The model stability is evaluated by leave-one-out cross-validation (LOOCV) on the training set. Finally, the generalization ability of the model is verified using an independent test set.

[0012] S5 interprets the feature importance by generating feature coefficients through feature analysis of each prediction model, and uses feature analysis and Shapley weighted interpretation (SHAP) to evaluate the results, thereby enhancing the interpretability of machine learning models and analyzing important influencing factors affecting heavy metal accumulation in rice.

[0013] S6. Based on the optimal model obtained after steps S1 to S5, the input feature factors are changed to variables of the optimal feature subset. Among them, the soil physicochemical properties are replaced with the soil physicochemical properties of other regions, and the metal physicochemical properties are replaced with the physicochemical properties of the predicted metals. Running the prediction model can obtain the content of unknown heavy metal bioaccumulation factors in rice grains of other regions.

[0014] The 26 characteristic factors in S1 include soil physicochemical properties and metal physicochemical properties.

[0015] Soil physicochemical properties include: soil pH, soil clay content (Clay), and soil organic matter (SOM);

[0016] The physicochemical properties of metals include: atomic number AN, atomic mass AW, atomic radius AR, covalent radius CR, and ionic radius r, which characterize geometric configuration; melting point MP, boiling point BP, and density D at 300K, which characterize thermal stability; ionization potential difference ΔIP, electronegativity Xm, standard redox potential ΔE0, potential IP, atomic ionization potential AN / ΔIP, soft index Σp, and electron density AR / AW, which characterize electrophilicity; and the first hydrolysis constant |logKOH| and covalent bond index Xm, which characterize hydrophilicity. 2 r, ion valence state Z, three polarization force parameters Z 2 / r、Z / r 2 Z / r and two quasi-polarization force parameters Z / AR 2 Z / AR.

[0017] Clay, SOM, and the biological enrichment factor BCF are converted into commonly used logarithmic forms to compress data size, thereby simplifying calculations without altering the properties of the data.

[0018] Prediction models include random forest, CatBoost (class boosting model), and XGBoost (extreme gradient boosting model);

[0019] The random forest model is as follows:

[0020] Where f represents the result, and B represents the number of guide samples or the number of iterations, f b (x ′ ) represents the summation of each data point x. ′ The function applied;

[0021] The CatBoost model is as follows:

[0022] Among them, L(y) i Let y, γ) be the loss function. i The true label is γ, and the initial predicted value is γ.

[0023] The XGBoost model is specifically as follows:

[0024] Where N represents the number of data points; It is a loss function used to measure the predicted value at each data point. Compared with the actual value y j The difference between them, K represents the number of trees or components in the set, Ω(f j ) is a regularization term used to penalize the behavior of a single tree f. j Related complex model structures.

[0025] In S3, 5-fold cross-validation involves randomly dividing the preprocessed dataset into 5 mutually exclusive subsets, each containing approximately 1 / 5 of the data. Four subsets are used as the training set, and the remaining subset is used as the validation set.

[0026] The network search in S3 finds the parameters that optimize model performance by traversing preset parameter combinations. It then filters the model configurations by minimizing the negative mean square error (NMSE) and selecting the configurations that minimize the squared error between the predicted and actual values. The smaller the mean square error (MSE) value, the larger the NMSE value, which results in a smaller deviation between the model's prediction and the actual value, and thus higher accuracy.

[0027] In S4, LOOCV (Local Optical Cross-Validation) is a commonly used validation method for QICAR models. It uses one sample from the original dataset as a validation variable and the remaining samples as training elements to prevent overfitting and evaluate the model's generalization ability. Its minimum acceptable value is 0.5, and its formula is as follows:

[0028] in, To output the actual value of the label, To output the predicted values ​​for the labels, This is the average value of the output labels.

[0029] The defining formulas for feature analysis and Shapley weighted interpretation are:

[0030] Where F is the set of all features; S is any subset of features excluding feature i; f(S) represents the predicted value of sample x when the model uses only subset S; fraction It is the weight of the subset, calculated based on the combined probabilities of all possible subsets.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] Known BCF data, metal physicochemical properties, and corresponding regional soil physicochemical properties are used to obtain predicted values ​​of unknown bioaccumulation factors BCF by coupling three machine learning models with QICAR. The optimal results are obtained through multiple rounds of cross-validation screening, thereby making accurate and scientific predictions of crop bioaccumulation factors in unknown regions.

[0033] This invention uses three machine learning models to screen feature factors affecting prediction results. The results show that only 4 to 7 feature factors are needed to achieve optimal prediction performance. Among them, the soft index Σp is a key influencing factor determining the accumulation effect of different heavy metals in crops. Based on the results obtained through the screening method of this invention, subsequent related research can directly focus on core feature factors, effectively avoiding problems such as increased data processing costs, reduced model computation efficiency, and wasted time caused by introducing too many redundant features, thus significantly improving the efficiency of research and application. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the model flow of the present invention.

[0035] Figure 2 The results of the three machine learning models of this invention are shown in the figure.

[0036] Figure 3 This is a ranking chart showing the importance of the results from the three machine learning models of this invention.

[0037] Figure 4 This is a feature SHAP value distribution and partial dependency graph of the CAT model of the present invention.

[0038] Figure 5 Box plots showing the prediction results for eight metals in this invention.

[0039] Figure 6 This is an interpolated distribution map of the BCF values ​​of eight heavy metals in rice, taken as an example of an embodiment of the present invention using Hainan Island as an example.

[0040] Figure 7 This is a schematic diagram of the entire process of the method of the present invention. Detailed Implementation

[0041] See Figures 1 to 7 The present invention will now be further described with reference to the accompanying drawings. The present invention provides a method for predicting the content of bioaccumulation factors using QICAR modeling:

[0042] The prepared data includes 529 sets of soil-rice corresponding data, including six heavy metals / metalloids: Cd, Pb, Cu, Zn, Ni, and As, as well as the corresponding soil physicochemical properties in the environment, pH range 4.23-7.61, organic matter content (SOM) 0.44-10.10%, and clay content (Clay) 2.04-74.61%. The physicochemical properties of the metals are shown in Table 1 below.

[0043] Table 1

[0044]

[0045]

[0046] After preprocessing, the data was input into three machine learning models. The optimal subset number and optimal parameter results were selected step by step according to the model workflow. The optimal parameter results are shown in Table 2 below.

[0047] Table 2

[0048]

[0049]

[0050] according to Figure 2 Figure (a) shows that the optimal number of features differs among different machine learning models: RF has an optimal number of 4 features, XGB has an optimal number of 5 features, and CAT has an optimal number of 7 features. It is not true that more input factors necessarily lead to better performance for all models; redundant features may introduce noise or multicollinearity, resulting in decreased model performance. The performance of the three machine learning models is shown in [Figure A]. Figure 2 As shown in (b) to (d), the optimal model is CAT(R). 2 =0.9144, MSE=0.0620), followed by RF(R) 2 =0.8634, MSE=0.0989), and finally the XGB model (R 2 =0.7854, MSE=0.1701).

[0051] The model results show that ( Figure 2 As shown, selecting only 4-7 factors from 26 feature factors as input to the model allows for good model performance, explaining 78%-91% of the logBCF. This is because there is a high correlation between the physicochemical properties of metals of the same type, resulting in redundancy in the model. Furthermore, CATBoost performs better, likely because its symmetric tree structure and adaptive learning rate design significantly improve generalization performance, making it more effective than RF and XGB in capturing nonlinear relationships in complex categories. Further investigation into the feature contributions of metal accumulation in rice reveals the main controlling factors through feature importance analysis, followed by SHAP analysis to determine the contribution direction of each factor. Figure 3 The ranking and weights of features corresponding to three machine learning models are shown. It can be seen that Σp has the greatest impact on the accumulation of heavy metals in rice, and its contribution to the model is significantly higher than that of other variables. This shows that Σp is an important factor in determining the accumulation effect of different heavy metals.

[0052] Figure 4Figure (a) shows the SHAP values ​​of the seven input features in the CATBoost model. The horizontal position represents the SHAP value for each data point, and the color reflects the magnitude of the feature value. SHAP values ​​to the left of zero indicate a negative impact on the model output, while points to the right indicate a positive impact. The CATBoost model shows that Σp, Z... 2 / r, SOM and Z / AR 2 The high values ​​of α and β are mainly located to the left of the zero SHAP value, indicating that these characteristics have a negative impact on the absorption of heavy metals by rice. Conversely, the high values ​​of CR are mostly distributed to the right of the zero SHAP value, indicating that the covalent radius has a positive impact on the model. In contrast, pH and Clay do not show a clear trend in the model, suggesting that the contribution of soil physicochemical properties to the enrichment of heavy metals in rice is complex.

[0053] To better observe whether there is a relationship between soil physicochemical properties and metallic properties, SHAP partial dependence plots were drawn between pairwise properties. Among them, pH and Σp show a relatively clear practical significance, such as... Figure 4 As shown in Figure (b), when pH < 6, a smaller Σp value contributes more to the model. As Σp increases, the model's SHAP value gradually decreases, even becoming negative. This indicates that in regions with lower pH, a smaller Σp value has a more significant positive impact on the model's predictions; when pH > 6, in regions with higher pH, a larger Σp value corresponds to a higher SHAP value, while a smaller Σp value leads to a lower SHAP value. Therefore, in regions with lower pH, it is advisable to focus more on the enrichment effect of metals with low soft indices, such as Hg, Cd, and Cu; and in regions with higher pH, focus more on the enrichment effect of metals with high soft indices, such as Ni, Co, and Pb.

[0054] Subsequently, to apply the s-QICAR model to the field, Hainan Island was selected as an example. Soil physicochemical properties, including pH, Clay, and SOM, were obtained, along with properties of eight metals / types, including Cd, Cu, As, and Pb (included in the training model) and Hg, Mn, Ni, and Co (not included in the training model). The seven optimal features were input into the CAT model for training, resulting in the prediction of 7280 BCF data points for each heavy metal. The box plot of the prediction results is shown below. Figure 5 As shown.

[0055] from Figure 5 It can be seen that the BCF values ​​of Ni, Co, Pb, Mn, and As are in a smaller order of magnitude, and the order of cumulative amount is Cd>Hg>Cu>Mn>As>Co>Pb>Ni.

[0056] Figure 6The map shows the interpolated distribution of predicted BCF values ​​on Hainan Island. It can be seen that high BCF values ​​for Ni, Mn, and Co are mainly distributed in Baisha and Qionghai areas, which have low pH and high SOM soils, potentially leading to easier accumulation of Ni, Mn, and Co. Hg, Cu, and Cd accumulate mainly in rice paddies along the western coast of Hainan, where the lower elevation results in higher soil pH than the rest of the island, but lower Clay and SOM content. As differs from other metals, accumulating more easily in rice paddies at higher altitudes in the central part of the island, where soil physicochemical properties show low pH, high SOM, and high Clay. Pb accumulation factors are low across the island, suggesting that more attention should be paid to the coastal areas of Changjiang and Wenchang.

[0057] The above are merely preferred embodiments of the present invention, intended only to aid in understanding the method and core ideas of this application. The scope of protection of the present invention is not limited to the above embodiments; all technical solutions falling within the scope of the present invention's concept are within its protection. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

[0058] This invention comprehensively addresses the shortcomings of existing measurement methods, which require the detection of pollutant concentrations in organisms and the environment, resulting in long cycles, high costs, and difficulty in meeting the needs of large-scale rapid assessment. Furthermore, these methods often focus on single heavy metals, lacking applicability to multiple heavy metal predictions, and fail to fully utilize the quantitative ion characteristic-activity relationship (QICAR) model in crop BCF prediction. By using known BCF values ​​and the corresponding soil and metal physicochemical properties as data, this invention employs a QICAR-coupled machine learning model to train and obtain predicted BCF values ​​for unknown heavy metals in the target area. Multiple rounds of cross-validation are used to screen the test set data, and the prediction is achieved through the coefficient of determination R... 2 The model's performance was evaluated using the mean absolute error (MAE) to obtain the optimal model. Then, through feature analysis and Shapley weighted interpretation of the evaluation results, accurate and scientific predictions were obtained, providing effective support for predicting the heavy metal enrichment capacity of crops and ensuring food security.

Claims

1. A method for predicting the content of bioaccumulation factors using QICAR modeling, characterized in that, Includes the following steps: S1. Input 26 feature factors into the prediction model to obtain the feature importance ranking, and arrange the physicochemical properties of 23 metals from high to low according to the importance score; S2, the arranged metal physicochemical features are cyclically input into the prediction model in an n+1 manner, and each feature subset includes soil physicochemical properties and metal physicochemical properties; S3, For each set of feature subsets, a network search method with 5-fold cross-validation is used to screen the optimal combination of parameters and feature subsets by minimizing the negative mean square error; S4. The data is randomly divided into training set and test set in a 4:1 ratio. Based on the selected optimal feature subset and the hyperparameters of the tuned parameters, the prediction model is retrained on the training set. The stability of the model is evaluated by leave-one-out cross-validation (LOOCV) on the training set. Finally, the generalization ability of the model is verified by the independent test set. S5 interprets the feature importance by generating feature coefficients through feature analysis of each prediction model, and uses feature analysis and Shapley weighted interpretation (SHAP) to evaluate the results, thereby enhancing the interpretability of machine learning models and analyzing important influencing factors affecting the accumulation of heavy metals in rice. S6. Based on the optimal model obtained after steps S1 to S5, the input feature factors are changed to variables of the optimal feature subset. The soil physicochemical properties are replaced with the soil physicochemical properties of other regions, and the metal physicochemical properties are replaced with the predicted metal physicochemical properties. Running the prediction model can obtain the content of unknown heavy metal bioaccumulation factors in rice grains of other regions. The 26 characteristic factors in S1 include soil physicochemical properties and metal physicochemical properties; The soil physicochemical properties include: soil pH, soil clay content (Clay), and soil organic matter (SOM); The physicochemical properties of the metal include: atomic number AN, atomic mass AW, atomic radius AR, covalent radius CR, and ionic radius r characterizing the geometric configuration; melting point MP, boiling point BP, and density D at 300K characterizing thermal stability; ionization potential difference ΔIP, electronegativity Xm, standard redox potential ΔE0, potential IP, atomic ionization potential AN / ΔIP, soft index Σp, and electron density AR / AW characterizing electrophilicity; and the first hydrolysis constant |logKOH| and covalent bond index Xm characterizing hydrophilicity. 2 r, ion valence state Z, three polarization force parameters Z 2 / r、Z / r 2 Z / r and two quasi-polarization force parameters Z / AR 2 Z / AR.

2. The method for predicting the content of bioaccumulation factors using QICAR modeling according to claim 1, characterized in that, The Clay, SOM, and the biological enrichment factor BCF are converted into common logarithmic forms to compress data size, thereby simplifying calculations without altering the data properties.

3. The method for predicting the content of bioaccumulation factors using QICAR modeling according to claim 1, characterized in that, The prediction models include the random forest model, the CatBoost class boosting model, and the XGBoost extreme gradient boosting model. The random forest model is specifically as follows: ;in Indicates the result, This indicates the number of guide samples or the number of iterations. This represents the summation of each data point. The function applied; The CatBoost model is specifically as follows: ;in, For loss function, For real labels, These are the initial predicted values; The XGBoost model is specifically as follows: Where N represents the number of data points; It is a loss function used to measure the predicted value at each data point. Compared with actual value The difference between them, where K represents the number of trees or components in the set, It is a regularization term used to penalize a single tree. Related complex model structures.

4. The method for predicting the content of bioaccumulation factors using QICAR modeling according to claim 1, characterized in that, The 5-fold cross-validation in S3 involves randomly dividing the preprocessed dataset into 5 mutually exclusive subsets, each containing 1 / 5 of the data. Four subsets are used as the training set, and the remaining subset is used as the validation set.

5. The method for predicting the content of bioaccumulation factors using QICAR modeling according to claim 1, characterized in that, The network search in S3 involves traversing preset parameter combinations to find the parameters that optimize model performance. By minimizing the negative mean square error (NMSE), the model configuration that minimizes the squared error between the predicted and actual values ​​is selected as the best performing configuration. The smaller the mean square error (MSE) value, the larger the NMSE, and the smaller the deviation between the model prediction result and the actual value, resulting in higher accuracy.

6. The method for predicting the content of bioaccumulation factors using QICAR modeling according to claim 1, characterized in that, The LOOCV cross-validation in S4 is a method that uses one sample from the original sample as a validation variable and the remaining samples as training elements to prevent overfitting of the equation and evaluate the generalization ability of the model. Its minimum acceptable value is 0.5, and its formula is as follows: ;in, To output the actual value of the label, To output the predicted values ​​for the labels, This is the average value of the output labels.

7. The method for predicting the content of bioaccumulation factors using QICAR modeling according to claim 1, characterized in that, The definition formulas for feature analysis and Shapley weighted interpretation are as follows: Where F is the set of all features; S is any subset of features that does not contain feature c; This represents the predicted value of sample x when the model uses only a subset S; fraction It is the weight of the subset, calculated based on the combined probabilities of all possible subsets.

Images