Toxicheck: In-Silico Nano-QSAR Toxicity Classification using Hybrid Machine Learning Algorithms

Abstract

The health hazards and risks of nanoparticles (NPs) and engineered nanomaterials (ENMs) are linked to their physicochemical features. Due to their minute structure, they can cause intracellular and genetic damage, and harm the environment by forming toxic mixtures with other compounds. Thus, it is essential to assess them first before they are mass-produced for public use. Traditionally, nanomaterial toxicity involves in-vivo and in-vitro approaches, but in recent years, machine learning (ML) algorithms have also emerged as predictive tools through in-silico means. This approach provides a faster, cheaper, and safer way to assess the toxicological profile of a nanomaterial. This study aims to investigate the applicability and efficiency of using hybrid algorithms in nanomaterial toxicity classification. They are formed by combining Genetic Algorithm (GA) with different base classifiers, namely Logistic Regression (LR), Artificial Neural Network (ANN), and Random Forest (RF). Generally, the hybrid algorithm-based models perform better than their base classifier counterparts, with an increase in scores of up to 19%. Using MCC as the main metric, results show that GA-RF with SMOTE is the best-performing model with an MCC score of 0.34. Building upon this model, this study developed a web application that lets the user input information about a nanomaterial and the cell-based assay that will be exposed for a certain amount of time. It predicts the cell viability of the assay to produce a toxicity classification for the nanomaterial.