Nanotoxicity Prediction Based on Half Maximal Effective Concentration (EC50) Values Using Elastic Net Regression and k-Nearest Neighbors

Abstract

This study focuses on predicting the toxicity of nanoparticles using their half maximal effective concentration (EC50) values through Elastic Net Regression (ENET) and k-Nearest Neighbors (kNN) models. A dataset of 5,926 samples with 10 descriptors was used to develop Nano - Quantitative structure - activity relationship (Nano-QSAR) models. The kNN model with  demonstrated the best performance, achieving an  of 91.66%, RMSE of 0.79, and MAE of 0.34 on the training set, and an  of 91.58%, RMSE of 0.79, and MAE of 0.32 on the test set, indicating strong predictive ability and good generalization to new data. The ENET model, however, showed substantially lower accuracy with performance measures below 60%, suggesting it was not suitable for this dataset. Key descriptors associated with electronegativity and atomic properties were identified as major contributors to toxicity. These findings highlight the potential of kNN as an effective approach for predicting nanoparticle toxicity and supporting the development of safer nanomaterials.