From literature to predictive modeling: Insights and machine learning applications from in vitro comet assays related to the genotoxicity of titanium dioxide nanomaterials

Abstract

The genotoxicity of titanium dioxide nanomaterials (TiO₂ NMs) remains a debated topic in the scientific community. In this study, we applied the read-across concept based on machine learning (ML) algorithms to predict the genotoxic potential of TiO₂ NMs. Key objectives included: (i) compiling a systematic dataset capturing DNA damage percentage from in vitro comet assays, (ii) creating a homogenized dataset integrating physicochemical properties, exposure conditions, and experimental details, (iii) training ML models for prediction, (iv) evaluating model performance, and (v) identifying the features that contribute the most to predictive accuracy. The dataset was divided into three parts: the Entire dataset (all features), the Physicochemical dataset, and the Experimental design dataset. Extra Trees Regressor and XGB Regressor demonstrated high predictive accuracy, achieving R² values of 0.906 and 0.788 for the P-chem and Experimental dataset, respectively. Exposure concentration, cold lysis conditions, and electrophoresis parameters emerged as key predictors of DNA damage, alongside contributions from NM properties. These findings highlight the intricate interplay between NM properties and experimental conditions in genotoxicity assessments. By providing a FAIR dataset, this study facilitates future research, allowing for the integration of additional variables and quality criteria to enhance the modeling approach. This work reinforces the value of nano-informatics in nanosafety and serves as a footing for advancing data-driven hazard assessment methodologies, positioning ML-enabled read-across strategies as a valuable tool for regulatory nanosafety framework.