Chaotic neural network algorithm with competitive learning integrated with partial Least Square models for the prediction of the toxicity of fragrances in sanitizers and disinfectants

This study addresses the need for accurate structural data regarding the toxicity of fragrances in sanitizers anddisinfectants. We compare the predictive and descriptive (model stability) potential of multiple linear regression(MLR) and partial least squares (PLS) models optimized through variable selection (VS). A novel hybrid chaoticneural network algorithm with competitive learning (CCLNNA)–PLS modeling strategy can offer specific opti-mization with satisfactory results, even for a limited dataset. While also exploring the preliminary comparativeanalysis, the goal is to introduce an adapted novel CCLNNA optimization strategy for VS, inspired by neuralnetworks, along with exploring the influence of the percentage of significant descriptors in the optimizationfunction to enhance the final model's capabilities.We analyzed an available dataset of 24 molecules, incorporating ADMET and PaDEL descriptors as predictorvariables, to explore the relationship between the response/target variable (pLC50) and the meticulously opti-mized set of descriptors. The suitability of the selected PLS models (cross- and external-validated accuracycombined with percentage of significant descriptors at a level equal to or >80 %) underscores the importance of expanding the dataset to amplify the validation protocols, thus enhancing future model reliability and envi-ronmental impact.