Feature-Driven Nano-QSTR Modeling for Nanoparticle Cytotoxicity Prediction using Ensemble Machine Learning

Abstract

The rapid growth of engineered nanoparticles in medicine, electronics, and industry has increased the need for reliable toxicity assessment. In response, this study develops nano-QSTR regression models using Extreme Gradient Boosting (XGBoost) and Random Forest (RF), with Ridge regression included as a baseline model, to predict nanoparticle cytotoxicity measured as percent cell viability. The models incorporate features related to nanoparticle properties, biological characteristics, exposure conditions, synthesis methods, and surface coatings. A dataset of 4,856 samples was used, with 3,887 for training and 969 for testing. During repeated cross-validation (rcv) on the training data, XGBoost achieved R2rcv=81.84%±1.79, RMSErcv=12.73%±0.60, and MAErcv=8.15%±0.29, and RF reached R2rcv=76.42%±2.46, RMSErcv=14.79%±0.90, and MAErcv=10.01%±0.40. Whereas, Ridge performed substantially worse (R2rcv=33.74%±2.53, RMSErcv=24.33%±0.61 , MAErcv=18.42%±0.45). External validation confirmed these trends: XGBoost achieved R2test=80.25%, RMSEtest=13.09%, and MAEtest=9.01%, RF reached R2test=75.59%, RMSEtest=14.87%, MAEtest=10.32%, and Ridge obtained R2test=33.02%, RMSEtest=24.13%, MAEtest=18.27%. Feature importance analysis from the ensemble models identified primarily nanoparticle concentration, size, diameter, zeta potential, exposure time, number of cells, and synthesis method as key drivers of cytotoxicity. These results demonstrate that ensemble models can reliably predict cytotoxicity and highlight factors critical for early safety assessment.