QSAR Models for Non-Effective Levels of Chemical Substances Using the CORAL Software
TL;DR
A comparative analysis shows that incorporating the CII significantly improves model performance: the average coefficient of determination (R2) for the validation set increased, underscoring the advantage of this approach, and a similar improvement occurs on the calibration set.
Abstract
A new predictive model is proposed for determining the non-effective levels (NELs) of chemical substances across seven different training and validation splits. The model utilizes optimal molecular descriptors calculated from SMILES (Simplified Molecular Input-Line Entry System) strings, with correlation weights optimized via the Monte Carlo method. To improve predictive accuracy, we implemented this randomized optimization process using a Las Vegas algorithm integrated with the Correlation Intensity Index (CII) as a novel predictive criterion. A comparative analysis shows that incorporating the CII significantly improves model performance: the average coefficient of determination (R2) for the validation set increased from 0.57 ± 0.07 (without CII) to 0.69 ± 0.07 (with CII), underscoring the advantage of this approach, and a similar improvement occurs on the calibration set. This shows that the adopted methodology is useful, which allows us to develop more general models.