Machine Learning Classification of Nanoparticle Toxicity for Enhanced Cancer Treatment

The nanoparticles (NPs) used in recent advancements in medicine have emerged as agents for cancer treatment. However, countering their toxicity remains a considerable obstacle to the field's pursuit of clinical success. This study investigates the performance of five machine learning (ML) algorithms, Random Forest, Bagging Classifier, Tuned Bagging Classifier, Decision Tree, and Tuned Decision Tree, in accurately classifying nanoparticles as toxic or non-toxic based on ten physicochemical parameters. Extensive preprocessing of the dataset of 810 nanoparticles was performed, including class balancing, outlier mitigation, and strict inclusion criteria based on the completeness and relevance of the physicochemical data for clinical applications of nanoparticles. Class balancing was achieved by under-sampling the majority class. This not only allowed the minority class to be better represented during model training, but also reduced model training time. Outlier mitigation and strict inclusion of relevant data enhanced the study's predictive robustness. The study reports the Random Forest classifier as the best performer among the five ML algorithms, with an F1 measure of 97.1%.