Multi-model integrated prediction method and system for toxicity of ionic liquid to human breast cancer cells

Abstract

The invention belongs to the technical field of machine learning, and provides a multi-model integrated prediction method and system for human breast cancer cytotoxicity caused by ionic liquid, and the main scheme is as follows: collecting experimental data and preprocessing the experimental data; splicing the quantifiable features to form an original feature matrix; performing feature dimension reduction processing on the original features to form an input feature matrix; taking the input feature matrix as input data, training a random forest, XGBoost and MLP basic machine learning model and carrying out model parameter adjustment, carrying out model integration on the trained basic machine learning model based on a weighted average fusion strategy, and selecting an optimal integrated model; evaluating the trained basic machine learning model, and applying the optimal integrated model to prediction of the human breast cancer cell toxicity of the ionic liquid of a new sample; and carrying out model deployment on the trained basic machine learning model and the selected optimal integrated model.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a multi-model integrated prediction method and system for the toxicity of ionic liquids to human breast cancer cells. Background Technology

[0002] Ionic liquids (ILs) are a class of salts that melt below 100 degrees Celsius. Due to their unique physicochemical properties, such as negligible vapor pressure, non-flammability, and good physicochemical stability, they are widely used in extraction, adsorption, electrochemistry, biocatalysis, and other fields. However, with the widespread application of ionic liquids, their potential hazards to the environment and human health have also attracted attention. Traditional experimental methods, such as the MTT assay, for determining the toxicity of ionic liquids are time-consuming, labor-intensive, costly, and highly dependent on the operator's skill level. While quantitative structure-activity relationship (QSAR) models and machine learning methods can predict toxicity, traditional QSAR models mainly rely on manually selected molecular descriptors, making it difficult to fully capture the complex structural features of ionic liquids; and traditional machine learning methods struggle to handle high-dimensional features, easily leading to overfitting or underfitting problems, and lack interpretability. Furthermore, there are few existing studies on computational methods for the toxicity of ionic liquids to the human breast cancer cell line (MCF-7).

[0003] Most existing research focuses on toxicity prediction methods based on machine learning. For example, Zhang et al. (2024) proposed a machine learning model based on molecular structure descriptors to predict the toxicity of ionic liquids to *Daphnia magna*. Their study constructed four machine learning models: Multilayer Perceptron (MLP), Random Forest (RF), Gradient Boosting Regression (GBR), and Support Vector Regression (SVR). They used tools such as RDKit, MACCS, and ECFP to compute molecular descriptors and optimized model parameters through grid search and cross-validation. The results showed that the Random Forest model based on RDKit descriptors performed best, with an R² of 0.975 and an RMSE of 0.222. SHAP analysis revealed key molecular features affecting the toxicity of ionic liquids, finding that the substructures around carbon atoms are crucial to toxicity, and that the structure of oxygen-containing atoms can reduce toxicity. While this study improved the accuracy and interpretability of toxicity prediction, it mainly focused on the toxicity prediction of *Daphnia magna*, and the model construction was based solely on traditional molecular descriptors and machine learning methods, without incorporating deep learning techniques.

[0004] Safa Sadaghiyanfam et al. (2025) proposed a hybrid framework combining ChemBERTa chemical structure embeddings with convolutional neural networks (CNNs), XGBoost, and support vector regression (SVR) for predicting the toxicity of ionic liquids. This method extracts chemical structure embeddings from SMILES strings using ChemBERTa, combines them with molecular descriptors and fingerprint features, performs dimensionality reduction using PCA, and then optimizes the model using Optuna. Results show that the CNN model achieves the best performance, with an R² of 0.865, an RMSE of 0.390, and a Pearson correlation coefficient of 0.937. This research innovatively combines deep learning embeddings with traditional machine learning methods, improving the model's predictive accuracy and interpretability; however, the model construction and optimization process is complex and computationally expensive. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-model integrated prediction method and system for the toxicity of ionic liquids to human breast cancer cells, which can achieve accurate and low-cost prediction of the toxicity of ionic liquids to human breast cancer cells.

[0006] The technical solution adopted by this invention to solve its technical problem is as follows: On the one hand, this invention provides a multi-model integrated prediction method for the toxicity of ionic liquids to human breast cancer cells, comprising the following steps: Collect experimental data on the toxicity of ionic liquids to human breast cancer cells and perform pretreatment procedures on them; The molecular structure information of ionic liquids is transformed into quantifiable features, and key information related to molecular structure and toxicity is captured through multi-source feature fusion. The quantifiable features are then spliced ​​together to form the original feature matrix. The original features in the original feature matrix are subjected to feature dimensionality reduction processing to form the input feature matrix; Using the input feature matrix as input data, we train basic machine learning models such as random forest, XGBoost, and MLP and perform model parameter tuning. Based on the weighted average fusion strategy, we integrate the trained basic machine learning models and select the best ensemble model. The trained basic machine learning model was evaluated, and the best ensemble model was applied to predict the cytotoxicity of ionic liquids on human breast cancer cells in new samples. Deploy the trained base machine learning model and the selected best ensemble model.

[0007] As a further optimization, when collecting experimental data on the toxicity of ionic liquids to human breast cancer cells, the collected experimental data covers at least 200 different types of ionic liquids, and for each ionic liquid, the SMILES string of molecular structure and the corresponding toxicity index EC50 value are recorded simultaneously. The local processing operation on the collected experimental data includes: Outlier handling: Outliers in EC50 values ​​are identified using box plots. For samples that exceed the range of [Q1-1.5×IQR, Q3+1.5×IQR], the original experimental records are consulted to confirm whether they are due to operational errors. If they are confirmed as outliers, they are removed. Missing value handling: If the SMILES string is missing, the sample is removed directly. If the EC50 value is missing and the sample size is less than 5% of the total sample, the KNN algorithm is used to fill in the missing values ​​based on the EC50 values ​​of similar molecular structures. Data transformation: Logarithmic transformation of EC50 values ​​is performed using the formula: log10_EC50, which converts the non-linear distribution of toxicity index into an approximately normal distribution.

[0008] As a further optimization, the process of converting the molecular structure information of ionic liquids into quantifiable features includes: calculating molecular descriptors, generating fingerprint features, and ChemBERTa embedding. The molecular descriptors include physicochemical descriptors, 2D descriptors, and combinatorial descriptors. The calculation results of all descriptors are numerical. When the calculation result of a descriptor is not numerical, it is filled with 0. The fingerprint features include MACCS fingerprints and ECFP fingerprints; ChemBERTa embedding refers to: based on the pre-trained ChemBERTa model, converting the SMILES string into a 768-dimensional dense vector, using the AutoTokenizer with Hugging Face to segment SMILES into words with a maximum length of 128, and extracting the mean of the last hidden state as the molecular embedding through AutoModel. The process of capturing key information related to molecular structure and toxicity through multi-source feature fusion and splicing quantifiable features to form an original feature matrix refers to splicing the calculated molecular descriptor, generated fingerprint features, and ChemBERTa embedding according to the sample dimension to form an original feature matrix. In the original feature matrix, each row corresponds to one sample and each column corresponds to one feature.

[0009] As a further optimization, the original features in the original feature matrix are subjected to feature dimensionality reduction to form the input feature matrix, including the following steps: Calculate the variance of all molecular descriptors and fingerprint features, and remove molecular descriptors and / or fingerprint features with variance below the threshold, while retaining molecular descriptors and / or fingerprint features with high variance. The molecular descriptor, fingerprint features, and ChemBERTa embedding were each subjected to PCA dimensionality reduction, with parameters set to retain 95% of the variance.

[0010] As a further optimization, the step of using the input feature matrix as input data to train basic machine learning models such as Random Forest, XGBoost, and MLP, and then performing model parameter tuning, includes the following steps: The feature-label samples after feature dimensionality reduction are divided into training and test sets in an 8:2 ratio. The test set remains independent and does not participate in any training or parameter tuning process. Random Forest, XGBoost, and MLP basic machine learning models are built and trained based on the training set, with a pre-defined hyperparameter combination space. A grid search combined with 5-fold cross-validation is used to traverse the preset hyperparameter combination space. The average R² of each parameter combination over all validation folds is used as the model selection criterion, and the parameter combination with the highest R² is selected as the optimal configuration.

[0011] As a further optimization, in the random forest basic machine learning model, the number of trees n_estimators=300, the maximum depth max_depth=10, and max_features is set to log2; In the XGBoost basic machine learning model, n_estimators is increased to 500 while learning_rate is reduced to 0.05, and a dual sampling mechanism of subsample and colsample_bytree is introduced, both of which are 0.7, and a combination of regularization reg_alpha and reg_lambda is used. In the MLP basic machine learning model, the network structure is changed from the default single hidden layer to a double hidden layer, the activation function is changed from ReLU to tanh, the regularization coefficient alpha is increased from 0.0001 to 0.01, the batch_size is reduced to 16 and the max_iter is increased to 1000.

[0012] As a further optimization, the weighted average fusion strategy includes: Average ensemble: All base machine learning models have equal weights. In this case, the weight calculation formula is: ; MAE-weighted ensemble: The weights are inversely proportional to the MAE of each base machine learning model on the training set. In this case, the weight calculation formula is: , Representing the A machine learning model MAE value; MSE-weighted ensemble: The weights are inversely proportional to the MSE of each base machine learning model on the training set. In this case, the weight calculation formula is: , Representing the A machine learning model MSE value; R²-weighted ensemble: The weights are proportional to the R² of each model on the training set. In this case, the weight calculation formula is: , Representing the The R² value of a machine learning model; Representing the The weights of a machine learning model; The final prediction calculation formula is as follows: , Represents the final predicted value. , These represent the weights and predicted values ​​of the RF model, respectively. , These represent the weights and predicted values ​​of the XGBoost model, respectively. , These represent the weights and predicted values ​​of the MLP model, respectively. When selecting the best ensemble model, the performance of each weighted average fusion strategy is evaluated on an independent test set. RMSE is used as the main evaluation metric, and the underlying machine learning model corresponding to the weighted average fusion strategy with the lowest RMSE is selected as the best ensemble model.

[0013] As a further optimization, the evaluation of the trained base machine learning model includes: Model performance evaluation: Evaluation metrics are calculated on the test set, including mean squared error (MSE), root mean squared error (RMSE) = √MSE, mean absolute error (MAE), coefficient of determination (R²), and Pearson correlation coefficient (r). Feature importance analysis: For the random forest and XGBoost models, the built-in feature importance scores are extracted, and the top 20 important features are selected after ranking to analyze their corresponding molecular structural significance. SHAP value analysis: The fusion model is explained using KernelExplainer from the SHAP library. The SHAP value of each feature on the prediction result is calculated using the test set samples as background data. The positive / negative impact of the feature on the toxicity is shown through the SHAP summary plot. Residual analysis: Calculate the residuals between the predicted values ​​and the actual values ​​of the test set, and draw residual histograms and residual-predicted value scatter plots. The residual mean should be close to 0 and the standard deviation should be <0.5. Application Domain Analysis: The application domain AD is defined based on the leverage value and standardized residuals. Samples with a leverage value > 3p / n or a standardized residual > 3 are considered to exceed the AD. p is the number of features and n is the sample size. For new samples, their leverage value and standardized residuals are calculated. If they exceed the AD, they are marked as "low reliability" and the user is prompted to conduct experimental verification.

[0014] As a further optimization, the deployment of the trained base machine learning model and the selected best ensemble model includes the following steps: Backend system construction: Develop a RESTful API based on the Flask framework, that is, encapsulate the feature extraction, preprocessing, and model prediction process into an interface, with the input being the string SMILES and the output being the prediction result, supporting HTTP / HTTPS requests; Model deployment: The pre-trained ensemble model and PCA converter are serialized using joblib and deployed on a cloud server with 4 cores and 8GB of memory to support parallel computing; Front-end mini-program development: includes 3 core pages: input page, results page, and history page; Interaction design includes input validation, loading animations, and result visualization.

[0015] On the other hand, the present invention provides a multi-model integrated prediction system for the toxicity of ionic liquids to human breast cancer cells, applied to the aforementioned multi-model integrated prediction method for the toxicity of ionic liquids to human breast cancer cells, including: The data collection and preprocessing unit is used to collect experimental data containing the toxicity of ionic liquids to human breast cancer cells and to perform preprocessing operations on them. The feature extraction and fusion unit is used to convert the molecular structure information of ionic liquids into quantifiable features, capture key information related to molecular structure and toxicity through multi-source feature fusion, and splice the quantifiable features to form the original feature matrix. The feature reduction unit is used to reduce the dimensionality of the original features in the original feature matrix to form the input feature matrix. The model training and ensemble unit is used to train basic machine learning models such as random forest, XGBoost and MLP using the input feature matrix as input data and to perform model parameter tuning. It also integrates the trained basic machine learning models based on a weighted average fusion strategy and selects the best ensemble model. The model evaluation and prediction unit is used to evaluate the trained basic machine learning model and apply the best ensemble model to predict the cytotoxicity of ionic liquids on human breast cancer cells in new samples. The model deployment unit is used to deploy the trained base machine learning model and the selected best ensemble model.

[0016] The beneficial effects of this invention are: by tuning the parameters of three machine learning models and integrating them based on a weighted average fusion strategy, this invention can effectively improve the overall prediction accuracy and stability of the model, without requiring high model deployment and parameter tuning costs. Attached Figure Description

[0017] Figure 1 This is a flowchart of the multi-model integrated prediction method for the toxicity of ionic liquids to human breast cancer cells in Embodiment 1 of the present invention; Figure 2 A diagram comparing the performance of the three models; Figure 3 The figure shows a comparison of the results of the optimized RF and the traditional RF in this paper. Figure 4 The figure shows a comparison of the results of the optimized XGBoost and the traditional XGBoost in this paper. Figure 5 The comparison chart of the results of the optimized MLP and the traditional MLP in this paper; Figure 6 A schematic diagram of SHAP analysis in an MLP model; Figure 7 A schematic diagram of the top 20 key features of a random forest; Figure 8 A schematic diagram of the first 20 key features of XGBoost; Figure 9 Scatter plot comparing the three models with the ensemble model; Figure 10 Performance comparison chart of four integration strategies; Figure 11 Histogram of residuals in the ensemble model; Figure 12 Scatter plot of residuals-predictions from the ensemble model; Figure 13 (a) Williams diagram of MLP, (b) Schematic diagram comparing the performance of MLP inside and outside the application domain; Figure 14 (a) Williams diagram of random forest, (b) schematic diagram comparing the performance of random forest inside and outside the application domain; Figure 15 (a) Williams diagram of XGBoost, (b) Schematic diagram comparing the performance of XGBoost inside and outside the application domain; Figure 16 (a) Williams diagram of the ensemble model, (b) schematic diagram comparing the performance of the ensemble model inside and outside the application domain. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0019] Example 1

[0020] This embodiment provides a multi-model integrated prediction method for the toxicity of ionic liquids to human breast cancer cells. See the flowchart below. Figure 1 The method may include the following steps: S1. Collect experimental data on the toxicity of ionic liquids to human breast cancer cells and perform pretreatment procedures on them. S2. The molecular structure information of ionic liquids is transformed into quantifiable features, and key information related to molecular structure and toxicity is captured through multi-source feature fusion. The quantifiable features are then spliced ​​together to form the original feature matrix. S3. Perform feature dimensionality reduction on the original features in the original feature matrix to form the input feature matrix; S4. Using the input feature matrix as input data, train basic machine learning models such as random forest, XGBoost and MLP and perform model parameter tuning. Then, based on the weighted average fusion strategy, integrate the trained basic machine learning models and select the best ensemble model. S5. Evaluate the trained basic machine learning model and apply the best ensemble model to predict the toxicity of ionic liquids to human breast cancer cells in new samples. S6. Deploy the trained basic machine learning model and the selected best ensemble model.

[0021] First, to provide high-quality, standardized foundational data for subsequent model training, and to use the preprocessed dataset as raw input for feature extraction, data quality directly impacts the effectiveness of subsequent feature extraction and model performance. Therefore, in this embodiment, step S1 can be implemented through the following specific steps: (1) Data Collection: Collect experimental data on the toxicity of ionic liquids to MCF-7 cells. The experimental data must meet the following requirements: cover at least 200 different types of ionic liquids (including imidazoles, pyridines, quaternary ammonium salts, etc.). For each ionic liquid, the SMILES string of molecular structure (standard Canonical SMILES format) and the corresponding toxicity index EC50 value (unit uniformly in μM) must be recorded simultaneously. The data source must be publicly available literature or experimental validation data, and the EC50 value must be measured by the standardized MTT method or similar reliable experimental methods, and the experimental conditions (such as culture time, cell concentration) must be consistent.

[0022] (2) Data Preprocessing: Outlier Handling: Outliers in EC50 values ​​are identified using box plots (IQR=1.5). For samples exceeding the range [Q1-1.5×IQR, Q3+1.5×IQR], it is necessary to confirm whether they are operational errors based on the original experimental records. If they are confirmed as outliers, they are removed. Missing Value Handling: If the SMILES string is missing, the sample is removed directly. If the EC50 value is missing and the sample size is less than 5% of the total sample, the KNN algorithm (k=5) is used to fill in the missing values ​​based on EC50 values ​​of similar molecular structures. If the missing value ratio is too high, the dataset is re-screened. Data Transformation: The EC50 values ​​are logarithmically transformed (log10_EC50) to convert the non-linear distribution of the toxicity index into an approximately normal distribution, thereby improving the stability of model training.

[0023] Secondly, since the extracted features are the input for feature dimensionality reduction, the richness and representativeness of the features directly affect the model's predictive ability. Therefore, this embodiment transforms the molecular structure information of ionic liquids into quantifiable features, and comprehensively captures key information related to molecular structure and toxicity through multi-source feature fusion. Therefore, in this embodiment, transforming the molecular structure information of ionic liquids into quantifiable features can include: calculating molecular descriptors, generating fingerprint features, and ChemBERTa embedding. Specifically, this can be achieved through the following steps: (1) Molecular descriptor calculation: Basic physicochemical descriptors were calculated using the RDKit tool, including over 30 physicochemical descriptors such as molecular weight (MolWt), lipid-water partition coefficient (LogP), topological polar surface area (TPSA), and number of hydrogen bond donors / acceptors (NumHDonors / NumHAcceptors); over 1300 2D descriptors (covering topology, electrical properties, fragment counts, etc.) were calculated using the Mordred calculator; and combinatorial descriptors (including constitutive, topological, and electronic types) were calculated using the PaDEL tool. Here, the conditions are: ensure that the SMILES string can be parsed normally by the RDKit / Mordred / PaDEL tools (samples that fail to be parsed need to be removed), and all descriptor calculation results must be numerical (non-numerical results are padded with 0).

[0024] (2) Fingerprint feature generation: MACCS fingerprint: A 166-bit binary fingerprint is generated using the GetMACCSKeysFingerprint function of RDKit, which encodes whether a specific substructure (such as an aromatic ring, hydroxyl group, etc.) exists in the molecule. ECFP fingerprint: A 2048-bit fingerprint is generated using the GetMorganFingerprintAsBitVect function of RDKit, with the parameter set to radius = 2 (ECFP4), which encodes structural features based on the local environment of atoms in the molecule (atoms and bonding relationships within a radius of 2).

[0025] ChemBERTa embedding: Based on a pre-trained ChemBERTa model (77M parameters), the SMILES string is converted into a 768-dimensional dense vector. Hugging Face's AutoTokenizer is used to segment SMILES (maximum length 128 characters), and the mean of the last hidden state is extracted using AutoModel as the molecular embedding. The requirements are: a PyTorch environment (supporting CPU / GPU acceleration) must be configured, the model loading path must be correct, and the SMILES string length must not exceed 128 characters (if too long, it will be truncated).

[0026] (3) Feature fusion: The molecular descriptors (RDKit / Mordred / PaDEL), fingerprint features (MACCS / ECFP), and ChemBERTa embeddings are spliced ​​together according to the sample dimension to form the original feature matrix (each row corresponds to one sample, and each column corresponds to one feature).

[0027] Then, dimensionality reduction is performed on the original features extracted from the feature matrix. The output low-dimensional features provide input for model training, which reduces feature redundancy and dimensionality, lowers model training complexity, and retains key information while improving the model's generalization ability. Therefore, in this embodiment, dimensionality reduction processing of the original features in the original feature matrix to form the input feature matrix may include the following steps: (1) Low-variance feature filtering: Calculate the variance of all features, remove features with variance below the threshold (0.01×(1-0.01)) (applicable to binary fingerprint features, such as MACCS / ECFP), and retain high-variance features (such as RDKit descriptors). Condition: Features need to be processed separately according to type (descriptor / fingerprint) to avoid removing low-variance descriptors with clear chemical meaning (such as functional group counts of certain fixed structures).

[0028] (2) Principal Component Analysis (PCA): PCA dimensionality reduction was performed on the high-dimensional features (MACCS fingerprint, ECFP fingerprint, Mordred descriptor, PaDEL descriptor, ChemBERTa embedding) respectively, with parameters set to retain 95% of the variance (controlled by n_components=0.95). Operation: A PCA transformer was trained separately for each high-dimensional feature (e.g., MACCS fingerprint → MACCS-PCA, ECFP fingerprint → ECFP-PCA) to map the original features to the principal component space. Finally, all principal components were merged with the unreduced RDKit descriptor to form the final input feature matrix.

[0029] Next, model training and ensemble can be carried out. This step aims to learn the nonlinear mapping relationship between the molecular features after dimensionality reduction and the toxicity intensity of the compound (represented by log_EC50) by constructing multiple machine learning models with complementary structures and strong generalization ability, thereby constructing a high-performance and robust toxicity prediction model. The dimensionality reduction feature matrix output from the previous step is used as input data in this step, and the ensemble model trained will serve as the core computational tool for subsequent toxicity assessment and prediction of new compounds.

[0030] During model training, the dataset needs to be partitioned first. In this embodiment, the feature-label sample set after feature dimensionality reduction is divided into a training set and an independent test set in an 8:2 ratio. A random sampling strategy (random seed = 42) is used to ensure experimental repeatability. During training, 5-fold cross-validation is used for hyperparameter tuning and model selection. The test set remains completely independent and does not participate in any training or hyperparameter tuning process. Then, the machine learning model can be built and trained. To achieve high-precision and high-robustness regression modeling tasks, this embodiment constructs three regression models with significantly different structures and complementary learning mechanisms as basic machine learning models: Random Forest, XGBoost, and Multilayer Perceptron Neural Network (MLP). Based on the traditional structure, each model significantly improves prediction performance and generalization ability compared to the traditional architecture through systematic hyperparameter search and structural optimization. The parameter search space and optimal configuration of each model are shown in Table 1.

[0031] The performance comparison chart of the random forest basic machine learning model before and after the improvement is shown in the figure. Figure 3 As shown in the figure, based on the traditional Bootstrap Aggregating framework, optimizing the number of trees (n_estimators=300) and the maximum depth (max_depth=10) effectively controls the risk of overfitting while maintaining model diversity. Setting max_features to 'log2' further enhances the randomness of feature sampling compared to the traditional 'sqrt' or all features, thus improving the model's generalization ability. MSE, RMSE, and MAE are all significantly reduced, especially MSE, indicating that the model in this embodiment significantly reduces the frequency of large errors. The final validation set R² reaches 0.7939 (an improvement of approximately 7%), demonstrating that these adjustments make our random forest more stable in complex feature interaction scenarios.

[0032] See the performance comparison chart of the XGBoost basic machine learning model before and after the improvement. Figure 4This embodiment significantly improves model robustness based on the gradient boosting framework through a multi-dimensional regularization strategy. Here, `n_estimators` is increased to 500 while `learning_rate` is decreased to 0.05, employing a more cautious addition strategy to effectively reduce training error. A dual sampling mechanism of `subsample` and `colsample_bytree` (both 0.7) is introduced to enhance the differentiation between base learners. The combination of regularization `reg_alpha` and `reg_lambda` further controls model complexity, enabling XGBoost to achieve an R² of 0.8171 on the test set (an improvement of approximately 16.9%), with significant decreases in MAE, MSE, and RMSE. This demonstrates that the improved XGBoost model in this embodiment significantly outperforms the traditional architecture, with substantial improvements in accuracy and stability.

[0033]

[0034] This embodiment performs the deepest structural optimization on the MLP basic machine learning model: (1) The network structure is adjusted from the default single hidden layer (100) to a double hidden layer (50,50), which enhances the nonlinear mapping ability of the model and enables the network to learn more complex feature interactions; (2) The activation function is changed from ReLU to tanh, which provides smoother gradient characteristics in regression tasks and alleviates the "death" problem of ReLU neurons; (3) The regularization coefficient alpha is increased from 0.0001 to 0.01, which significantly enhances the L2 regularization strength and effectively controls the risk of overfitting; (4) The batch_size is reduced to 16 and the max_iter is increased to 1000, making the training process more refined and ensuring the full convergence of the model. The performance comparison before and after the improvement is shown in the figure. Figure 5 As can be seen from the figure, these improvements enabled MLP to achieve the best single-model performance, with a test set R² of 0.8691. See the SHAP analysis plot of the MLP model for details. Figure 6 The optimized MLP model shows a significant improvement in nonlinear mapping and feature extraction capabilities compared to traditional shallow MLPs (such as single-hidden-layer structures). For example... Figure 2 As shown, the traditional MLP has a negative R² on the test set, indicating extremely poor performance. In contrast, the optimized MLP achieves an R² value of 0.8691, which is significantly better than the traditional structure. This demonstrates that the improvement in this embodiment is very reasonable and effective.

[0035] Experimental results show that the model structure improvement and systematic parameter tuning strategies adopted in this paper significantly enhance the expressive power and generalization performance of the base model. Furthermore, the ensemble model combines the advantages of different models, exhibiting superior robustness and accuracy in complex regression tasks.

[0036] It should be noted that this embodiment can use grid search combined with 5-fold cross-validation. The system traverses the preset hyperparameter combination space and uses the average R² of each parameter combination over all validation folds as the model selection criterion. The parameter combination with the highest R² is selected as the optimal configuration. The optimization process is entirely based on cross-validation within the training set, avoiding data leakage.

[0037] To fully leverage the strengths of each base model, suppress the bias of a single model, and further improve overall prediction accuracy and robustness, this embodiment employs a weighted average fusion strategy driven by training set performance to integrate the three optimal base models—Random Forest, XGBoost, and MLP—that have already been trained. The specific weighted average fusion strategy is as follows: Average ensemble: All base machine learning models have equal weights. In this case, the weight calculation formula is: ; MAE-weighted ensemble: The weights are inversely proportional to the MAE of each base machine learning model on the training set. In this case, the weight calculation formula is: , Representing the A machine learning model MAE value; MSE-weighted ensemble: The weights are inversely proportional to the MSE of each base machine learning model on the training set. In this case, the weight calculation formula is: , Representing the A machine learning model MSE value; R²-weighted ensemble: The weights are proportional to the R² of each model on the training set. In this case, the weight calculation formula is: , Representing the The R² value of a machine learning model; Representing the The weights of a machine learning model; The final prediction calculation formula is as follows: , Represents the final predicted value. , These represent the weights and predicted values ​​of the RF model, respectively. , These represent the weights and predicted values ​​of the XGBoost model, respectively. , These represent the weights and predicted values ​​of the MLP model, respectively. When selecting the best ensemble model, the performance of each weighted average fusion strategy is evaluated on an independent test set. RMSE is used as the main evaluation metric, and the underlying machine learning model corresponding to the weighted average fusion strategy with the lowest RMSE is selected as the best ensemble model.

[0038] When choosing the best ensemble model, see [link / reference]. Figure 10 The performance of four ensemble strategies was evaluated on an independent test set, with RMSE as the primary evaluation metric. The ensemble strategy with the lowest RMSE was selected as the final model. In this embodiment, through... Figure 10 The MSE-weighted integration strategy was found to perform best.

[0039] from Figure 6 The results show that the weighted average ensemble strategy performs best on multiple metrics such as weighted MSE and weighted R², with an R² weighted value of 0.8507, which is better than single models and other ensemble methods. This is because the ensemble strategy ensures that the model that performs better on the training set receives higher weights, achieving performance-driven adaptive ensemble, effectively improving the overall prediction accuracy and stability of the model, and thus further demonstrating the rationality and effectiveness of our ensemble strategy.

[0040] In this embodiment, it can be achieved through methods such as... Figure 9 The scatter plot shown compares the prediction performance of the three base models and the best ensemble model on the test set. Figure 9 It can be seen that the predicted values ​​of Best_Ensemble (R²=0.8561) are more closely distributed near the diagonal of the actual values, and its goodness of fit is better than any single model (RandomForest R²=0.7939; XGBoost R²=0.8171). This indicates that the ensemble model effectively reduces variance and bias through complementary mechanisms, and obtains more stable and accurate prediction results.

[0041] Next, the trained basic machine learning model can be evaluated. The evaluation results guide whether the model can be used for subsequent predictions, interpret the results to enhance the model's credibility, and comprehensively evaluate the model's performance, explain the prediction logic, clarify the model's applicable scope, and provide a basis for model optimization and practical application. The specific evaluation process can be implemented through the following steps: (1) Model performance evaluation: Calculate the evaluation metrics on the test set: mean square error (MSE), root mean square error (RMSE = square root of MSE), mean absolute error (MAE), coefficient of determination (R²), and Pearson correlation coefficient (r).

[0042] (2) Feature Importance Analysis: For the Random Forest and XGBoost models, extract the built-in feature importance scores (such as feature_importances in XGBoost), see [link to relevant documentation]. Figure 7 and Figure 8 After sorting, the top 20 important features were selected, and their corresponding molecular structural significance (such as specific functional groups and topological features) was analyzed.

[0043] (3) SHAP value analysis: The KernelExplainer of the SHAP library is used to interpret the fusion model: with the test set samples as background data, the SHAP value of each feature on the prediction result is calculated, and the positive / negative impact of the feature on the toxicity is shown by the SHAP summary plot (e.g., the presence of a certain functional group will increase the toxicity).

[0044] (4) Residual analysis: Calculate the residuals between the predicted values ​​and the actual values ​​of the test set, and plot the residual histogram (to verify the normal distribution) (see Figure 11 ) and residual-predicted scatter plot (to verify residual independence) (see Figure 12 The residual mean is required to be close to 0 and the standard deviation is less than 0.5.

[0045] (5) Application Domain (AD) Analysis: AD is defined based on the leverage value (Hat value) and standardized residuals: samples with a leverage value > 3p / n (p is the number of features, n is the sample size) or a standardized residual > 3 are considered to exceed AD. For new samples, calculate their leverage value and standardized residuals. If they exceed AD, they are marked as "low reliability" and the user is prompted to conduct experimental verification. For AD analysis of MLP, XGBoost, RF, and ensemble models, see [link to relevant documentation]. Figure 13 (a), 13(b)-16(a), 16(b).

[0046] Then, toxicity prediction can be performed, relying on the feature extraction method, dimensionality reduction model, and trained ensemble model built in the previous steps. Ensuring the prediction process is consistent with the training process allows the trained model to be applied to new samples, outputting toxicity prediction results. Specifically, the prediction process can be implemented through the following steps: (1) Input acquisition: Get the SMILES string of the ionic liquid to be predicted (must be a standard CanonicalSMILES with no syntax errors).

[0047] (2) Feature extraction: Calculate molecular descriptors (RDKit / Mordred / PaDEL), fingerprint features (MACCS / ECFP) and ChemBERTa embeddings according to the method in step S2, and ensure that the feature types are consistent with the training set.

[0048] (3) Feature preprocessing: According to the low variance feature filtering rules in step S3, redundant features are removed, and the high-dimensional features are reduced using the PCA transformer saved during the training phase to ensure that the principal component dimension is consistent with the training set.

[0049] (4) Model prediction: Input the preprocessed features into the trained ensemble model and output the log_EC50 predicted value.

[0050] (5) Output results: The predicted log_EC50 value, model reliability (whether it is within AD) and key impact features (the first 5 features based on the SHAP value) are returned simultaneously.

[0051] Finally, the trained basic machine learning model and the selected best ensemble model can be deployed. Here, based on the models and processes from the previous steps, system deployment enables the technology to be implemented, serving as the final application carrier of the entire solution. This transforms the technical solution into a convenient and usable tool, enabling real-time and mobile toxicity prediction and improving the practicality of the technology. In specific applications, model deployment can be achieved through the following steps: (1) Backend system construction: Develop RESTful API based on Flask framework: encapsulate the feature extraction, preprocessing and model prediction process into an interface (input is the string SMILES, output is the prediction result), and support HTTP / HTTPS requests.

[0052] (2) Model deployment: Use joblib to serialize the trained ensemble model and PCA converter, deploy it on a cloud server (such as Alibaba Cloud ECS), and configure 4 cores and 8G memory to support parallel computing (response time <1 second).

[0053] (3) Front-end mini-program development: Develop WeChat mini-program: including 3 core pages - input page (supports manual input of SMILES string or uploading by scanning code), results page (displays EC50 predicted value, reliability assessment, key feature impact diagram), and history page (saves user's past queries).

[0054] (4) Interaction design: input validation (automatically detect the correctness of SMILES format), loading animation (prompt the user to wait for prediction), result visualization (use bar charts to show the influence of features).

[0055] Therefore, the system interaction flow in this embodiment is as follows: the user enters SMILES in the mini program → the request is sent to the cloud API → the cloud automatically performs feature extraction → preprocessing → model prediction → generates results including SHAP analysis → returns to the mini program for display, with the entire process taking less than 5 seconds.

[0056] Therefore, as can be seen from the description of this embodiment, compared with the existing multi-model integrated prediction method for the toxicity of ionic liquids to human breast cancer cells, this embodiment has the corresponding technical effects in the following aspects: 1. Significantly Improved Efficiency and Economy: The prediction method in this embodiment completely eliminates the bottlenecks of traditional experimental methods (such as the MTT method). Through automated code implementation of multi-feature fusion, model training, and prediction processes, the toxicity assessment of a large number of ionic liquids can be completed in a very short time. Compared to time-consuming and labor-intensive wet experiments, this method significantly reduces experimental and time costs, providing a fast and economical solution for the safety screening and application decisions of ionic liquids.

[0057] 2. Improved Prediction Accuracy and Reliability: A comprehensive approach utilizing multiple molecular characterization techniques (RDKit, MACCS, ECFP, ChemBERTa embedding, Mordred, PaDEL) for feature extraction and fusion overcomes the limitations of traditional QSAR models that rely on manually selected descriptors, enabling a more comprehensive capture of the complex structural information of ionic liquids. Furthermore, a systematic model training and optimization process (including hyperparameter optimization using GridSearchCV) significantly improves the performance of the prediction model. Experimental results demonstrate that the constructed model outperforms existing methods in key metrics (such as R² value and RMSE) on the test set, proving its advantages in prediction accuracy and reliability.

[0058] 3. Innovative Multi-Model Integration and Architecture Optimization Enhance Generalization Ability: Addressing the insufficient generalization ability of existing single-model or simple combination models, this embodiment designs and implements a multi-model ensemble prediction system based on a weighted average fusion strategy. The specific integration method is as follows: First, three high-performing base models—Random Forest, XGBoost, and Deep Neural Network (MLP)—are independently trained and optimized. Then, an ensemble system incorporating four fusion strategies is constructed: equal-weighted average ensemble, and three weighted ensemble methods based on the inverse proportionality of prediction errors (MAE, MSE) or the direct proportionality of the coefficient of determination (R²) of each model on the training set. The performance of each ensemble strategy is evaluated on an independent test set, and the best-performing strategy is selected as the final model. Experimental results show that the MSE-based weighted ensemble strategy is the most effective, more effectively assigning higher weights to high-performance models and improving overall prediction accuracy and stability. At the model architecture level, key designs are specifically implemented for the MLP model: a pipeline architecture including FeatureSelector (custom feature column selector), StandardScaler (normalization), and MLPRegressor (multilayer perceptron) is constructed. This architecture ensures input consistency through FeatureSelector, enhances training stability through StandardScaler, and effectively optimizes the performance and generalization ability of the MLP by combining early stopping and hyperparameter grid search (hidden layer sizes, activation, alpha, etc.). These differentiated models (RF / XGBoost based on trees, MLP based on neural networks) are complementary in feature learning and prediction mechanisms. The ensemble system comprehensively leverages the advantages of each model, significantly improving its adaptability and generalization ability to different types of ionic liquids and complex, variable data distributions.

[0059] 4. Enhanced Model Interpretability: This embodiment not only provides high-performance predictions but also deeply integrates model interpretation tools (SHAP value analysis). After model training, the importance of features is calculated and visualized by calling the SHAP library. This systematic interpretability analysis process can clearly reveal key molecular structural factors affecting the toxicity of ionic liquids (such as the role of specific atoms and substructures), far superior to the shortcomings of existing methods in terms of interpretive depth. This provides direct and valuable molecular-level insights for understanding the toxicity mechanism of ionic liquids and rationally designing novel low-toxicity ionic liquids. Application Domain Analysis and Reliability Assurance: The integrated Applicability Domain (AD) method effectively determines whether new samples are within the reliable prediction range of the model. For samples outside the application domain, the system can proactively warn users to treat the prediction results with caution and suggest experimental verification or model adjustment. By continuously analyzing the performance differences of the model inside and outside the application domain, this embodiment supports iterative optimization of the model's applicability and predictive ability, significantly improving the reliability and stability of the model in practical applications.

[0060] 5. Constructing an Intelligent and Real-Time Prediction Platform: To facilitate the convenient application of technological achievements, this embodiment constructs an intelligent prediction system based on cloud services and mini-programs. Technically, the backend deploys the trained optimal ensemble model on a cloud server (using the Flask / Django framework to provide a RESTful API), enabling stable model calls and efficient computation. The frontend develops a WeChat mini-program as the interactive interface. Users input the string "SMILES" for the ionic liquid via their mobile phones, and the system automatically completes feature extraction, toxicity prediction, and interpretability analysis (such as the contribution of key atoms or groups). This design embodies several core advantages: Eliminating environmental dependence—users do not need to install RDKit, Python, or machine learning libraries; all operations can be completed solely through the mini-program; Real-time response—relying on cloud computing power, the prediction process is completed within seconds, improving efficiency by over 1000 times compared to traditional experimental methods (which take several days); Mobile convenience—supporting researchers to conduct toxicity assessments anytime, anywhere in laboratories, conferences, and other scenarios; Full-process automation—from structural analysis and feature calculation to result generation, there is no human intervention throughout the entire process, significantly reducing operational errors and improving prediction reliability and repeatability.

[0061] Example 2 Based on Example 1, this example provides a multi-model integrated prediction system for the toxicity of ionic liquids to human breast cancer cells, including: The data collection and preprocessing unit is used to collect experimental data containing the toxicity of ionic liquids to human breast cancer cells and to perform preprocessing operations on them. The feature extraction and fusion unit is used to convert the molecular structure information of ionic liquids into quantifiable features, capture key information related to molecular structure and toxicity through multi-source feature fusion, and splice the quantifiable features to form the original feature matrix. The feature reduction unit is used to reduce the dimensionality of the original features in the original feature matrix to form the input feature matrix. The model training and ensemble unit is used to train basic machine learning models such as random forest, XGBoost and MLP using the input feature matrix as input data and to perform model parameter tuning. It also integrates the trained basic machine learning models based on a weighted average fusion strategy and selects the best ensemble model. The model evaluation and prediction unit is used to evaluate the trained basic machine learning model and apply the best ensemble model to predict the cytotoxicity of ionic liquids on human breast cancer cells in new samples. The model deployment unit is used to deploy the trained base machine learning model and the selected best ensemble model.

[0062] As can be seen from the description of Embodiment 1, the application scenario and implementation principle of this embodiment are the same as those of Embodiment 1, so they will not be repeated here.

[0063] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-model ensemble prediction method for ionic liquid cytotoxicity on human breast cancer cells, characterized by, The method comprises the following steps: Collecting experimental data on the toxicity of ionic liquids to human breast cancer cells and performing preprocessing operations thereon; Converting the molecular structure information of the ionic liquids into quantifiable features, capturing key information related to the molecular structure and toxicity through multi-source feature fusion, splicing the quantifiable features to form an original feature matrix; Performing feature dimension reduction processing on the original features in the original feature matrix to form an input feature matrix; Training random forest, XGBoost and MLP basic machine learning models using the input feature matrix as input data, performing model parameter tuning, and selecting the best ensemble model based on a weighted average fusion strategy; Evaluating the trained basic machine learning models and applying the best ensemble model to predict the toxicity of ionic liquids to human breast cancer cells in new samples; Model deployment of the trained basic machine learning models and the selected best ensemble model.

2. The method according to claim 1, wherein the method is characterized by, In the step of collecting experimental data on the toxicity of ionic liquids to human breast cancer cells, the collected experimental data covers at least 200 different types of ionic liquids, and the SMILES string of the molecular structure and the corresponding toxicity indicator EC50 value of each ionic liquid are recorded simultaneously; The preprocessing operations on the collected experimental data include: Outlier processing: identify outliers in the EC50 values through the box plot method, remove samples outside the range of [Q1-1.5*IQR, Q3+1.5*IQR], and confirm whether they are operation errors based on the original experimental records. If it is confirmed to be an outlier, it is removed; Missing value processing: if the SMILES string is missing, the sample is directly removed. If the EC50 value is missing and the sample size is <5% of the total samples, use the KNN algorithm to fill in the EC50 value based on similar molecular structures; Data conversion: log10_EC50 is used to convert the non-linearly distributed toxicity indicator into approximately normally distributed data.

3. The method according to claim 2, wherein the method is characterized by, The step of converting the molecular structure information of the ionic liquids into quantifiable features includes calculating molecular descriptors, generating fingerprint features, and ChemBERTa embedding; The molecular descriptors include physical and chemical descriptors, 2D descriptors, and combined descriptors. The calculation results of all descriptors are numerical values. If the calculation result of a descriptor is not a numerical value, it is filled with 0; The fingerprint features include MACCS fingerprints and ECFP fingerprints; ChemBERTa embedding refers to converting the SMILES string into a 768-dimensional dense vector based on the pre-trained ChemBERTa model, using Hugging Face's AutoTokenizer to tokenize the SMILES, setting the maximum length to 128 during tokenization, and extracting the mean value of the last layer of hidden states as the molecular embedding. The key information related to molecular structure and toxicity is captured by multi-source feature fusion, and the quantifiable features are spliced to form an original feature matrix. The original feature matrix is formed by splicing the calculated molecular descriptors, generated fingerprint features, and ChemBERTa embeddings according to the sample dimension. In the original feature matrix, each row corresponds to a sample, and each column corresponds to a feature.

4. The method according to claim 3, wherein the method is characterized by, The original features in the original feature matrix are subjected to feature dimension reduction processing to form an input feature matrix, including the following steps: Calculate the variance of all molecular descriptors and fingerprint features, and remove the molecular descriptors and / or fingerprint features with a variance below a threshold, and retain the molecular descriptors and / or fingerprint features with a high variance. The molecular descriptors, fingerprint features, and ChemBERTa embeddings are subjected to PCA dimension reduction respectively, and the parameter setting is set to retain 95% of the variance.

5. The multi-model ensemble prediction method of ionic liquid cytotoxicity to human breast cancer cells according to claim 1, wherein, The input feature matrix is used as input data to train random forest, XGBoost, and MLP-based machine learning models and perform model tuning, including the following steps: The feature-label samples after feature dimension reduction processing are divided into a training set and a test set in a ratio of 8:2, and the test set remains independent and does not participate in any training or tuning process. Based on the training set, random forest, XGBoost, and MLP-based machine learning models are constructed and trained, and a preset hyperparameter combination space is set. Grid search combined with 5-fold cross-validation is used to traverse the preset hyperparameter combination space, and the average R² of each parameter combination on all validation folds is used as the model selection standard to select the parameter combination with the highest R² as the optimal configuration.

6. The multi-model ensemble prediction method of ionic liquid cytotoxicity to human breast cancer cells according to claim 5, wherein, In the random forest-based machine learning model, the number of trees n_estimators=300, the maximum depth max_depth=10, and max_features is set to log2. In the XGBoost-based machine learning model, n_estimators is increased to 500 while learning_rate is reduced to 0.05, and the double sampling mechanism of subsample and colsample_bytree is introduced, both set to 0.7, and the regularization combination of reg_alpha and reg_lambda is used. In the MLP-based machine learning model, the network structure is adjusted from the default single hidden layer to double hidden layer, the activation function is changed from ReLU to tanh, the regularization coefficient alpha is increased from 0.0001 to 0.01, the batch_size is reduced to 16, and the max_iter is increased to 1000.

7. The method according to claim 5, wherein the method is characterized by, The weighted average fusion strategy includes: Equal integration: the weights of each base machine learning model are equal, and the weight calculation formula is: ; MAE-weighted ensemble: The weights are inversely proportional to the MAE of each base machine learning model on the training set. In this case, the weight calculation formula is: , Representing the A machine learning model MAE value; MSE-weighted ensemble: The weights are inversely proportional to the MSE of each base machine learning model on the training set. In this case, the weight calculation formula is: , Representing the A machine learning model MSE value; R2 weighted ensemble: the weight is proportional to the R2 of each model on the training set, and the weight calculation formula is: , R2 value representing the first machine learning model; representing a first machine learning model; and weights representing a second machine learning model. The formula for the final prediction is: , represent the final predicted values, , represent the weights and predicted values of the RF model, respectively, , represent the weights and predicted values of the XGBoost model, respectively, , represent the weights and predicted values of the MLP model, respectively. When selecting the best ensemble model, the performance of each weighted average fusion strategy is evaluated on the independent test set, with RMSE as the main evaluation indicator, and the weighted average fusion strategy corresponding to the best ensemble model is selected as the selected best ensemble model.

8. The method of claim 1, wherein the method is characterized by, The trained base machine learning model is evaluated, including: Model performance evaluation: Calculate evaluation metrics on the test set, including mean squared error (MSE), root mean squared error (RMSE = sqrt(MSE)), mean absolute error (MAE), coefficient of determination (R²), and Pearson correlation coefficient (r); Feature importance analysis: For random forest and XGBoost models, extract the built-in feature importance scores, sort them, and select the top 20 important features to analyze their corresponding molecular structure significance; SHAP value analysis: Use the KernelExplainer of the SHAP library to interpret the fusion model. With the test set samples as background data, calculate the SHAP value of each feature on the prediction result, and display the positive / negative impact of the feature on toxicity through the SHAP summary plot; Residual analysis: Calculate the residual between the predicted value and the true value of the test set, draw a residual histogram and a residual-predicted value scatter plot, and require the residual mean to be close to 0 and the standard deviation to be <0.5; Application domain analysis: Define the application domain (AD) based on the leverage value and standardized residual, where samples with leverage value > 3p / n or standardized residual > 3 are considered outside the AD, p is the number of features, and n is the sample size. For new samples, calculate their leverage value and standardized residual. If they exceed the AD, mark them as "low reliability" and prompt the user to verify them through experiments.

9. The method of claim 1, wherein the method is characterized by, The model deployment of the trained base machine learning model and the selected best ensemble model includes the following steps: Backend system construction: Develop RESTful API based on Flask framework, i.e., encapsulate feature extraction, preprocessing, and model prediction processes into interfaces, input SMILES string, output prediction result, and support HTTP / HTTPS requests; Model deployment: Serialize the trained ensemble model and PCA converter using joblib, deploy them on a cloud server with 4 cores and 8 GB of memory to support parallel computing; Front-end applet development: Contains three core pages: input page, result page, and historical record page; Interaction design: Includes input validation, loading animation, and result visualization.

10. A multi-model ensemble prediction system for ionic liquid toxicity on human breast cancer cells, applied to the multi-model ensemble prediction method for ionic liquid toxicity on human breast cancer cells according to any one of claims 1 to 9, characterized in that, It includes: Data collection and preprocessing unit, used to collect experimental data containing the toxicity of ionic liquids to human breast cancer cells, and to perform preprocessing operations on them; Feature extraction and fusion unit, used to convert molecular structure information of ionic liquids into quantifiable features, and to capture key information related to molecular structure and toxicity through multi-source feature fusion. The quantifiable features are spliced to form the original feature matrix; Feature dimension reduction unit, used to perform feature dimension reduction processing on the original features in the original feature matrix to form the input feature matrix; Model training and integration unit, used to train random forest, XGBoost, and MLP base machine learning models with the input feature matrix as input data and to perform model tuning. The trained base machine learning models are integrated based on the weighted average fusion strategy, and the best ensemble model is selected; A model evaluation and prediction unit is configured to evaluate the trained base machine learning model and apply the best ensemble model to prediction of the ionic liquid cytotoxicity to human breast cancer cells for new samples; A model deployment unit is configured to deploy the trained base machine learning model and the selected best ensemble model.

Images