Modelling enzyme inhibition toxicity of ionic liquid from molecular structure via convolutional neural network model

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• Quantitative structure-activity/property relationship (QSAR/QSPR) models are crucial for predicting chemical properties.
• Deep learning (DL) offers advanced methods for handling complex data relationships in QSAR/QSPR.
• Ionic liquids (ILs) require accurate toxicity assessment for safe application.

Purpose of the Study:

• To develop and evaluate advanced DL models for predicting acetylcholinesterase inhibitory toxicity of ILs.
• To compare the performance of CNN-based feature extraction against traditional feature engineering (FE) methods.
• To establish an intelligent tool for guiding the design and screening of low-toxicity ILs.

Main Methods:

• Development of a Convolutional Neural Network (CNN) model for automatic feature extraction from IL structures.
• Integration of CNN features with Support Vector Machine (SVM), Random Forest (RF), and Multilayer Perceptron (MLP) models.
• Optimization of model hyperparameters using grid search and 10-fold cross-validation.
• Comparison of CNN-based models against FE-based models (FE-SVM, FE-RF, FE-MLP).

Main Results:

• CNN-based models (CNN-SVM, CNN-RF, CNN-MLP) demonstrated substantially improved regression results compared to FE-based models.
• All six evaluated models exhibited good prediction accuracy.
• The CNN-integrated models outperformed traditional FE approaches in predicting IL toxicity.
• Optimized models showed improved performance over existing literature models.

Conclusions:

• CNNs are effective for self-learning and extracting relevant features from IL structures for toxicity prediction.
• DL-based QSAR/QSPR models, particularly those employing CNNs, offer superior predictive power.
• The developed models serve as a valuable computational tool for designing and screening safer ionic liquids.