A machine learning examination of nanomaterial safety

Nanotechnology is an emerging technologies with enormous potential for innovative applications. The introduction of nanoparticles (NPs) offers significant societal benefits and economic opportunities while posing major challenges in research and regulatory bodies regarding their safety. NPs display high heterogeneity concerning their physicochemical and quantum-mechanical properties and as such, their toxicological affect, narrowing their risk assessment to an ad hoc testing process. Traditional, toxicological risk assessment relies heavily on costly, ethically disputed animal testing One alternative to test the hazard of NPs is in silico techniques. Given that risk assessment as a subject of academic research is multidisciplinary by character, this thesis provides a multidimensional research in the premises of the challenge triangle of nanoscience, toxicology and machine learning. Over the last decades, various types of Machine Learning (ML) tools have been developed for predicting toxicological effects of nanoforms. In this thesis, I initially document the work that has been carried out, systematically. We investigate in details and bookmark ML methodologies used to predict toxicological outcomes and provide a review of the sequenced steps involved in implementing a model. Additionally, this thesis records the data used in published studies that predict endpoints and maps the pathways followed, involving biological features in relation to NPs exposure, their physicochemical characteristics and the most commonly predicted outcomes. The results, derived from published research of the last decade, are summarized visually, providing prior-based data mining paradigms to be readily used by the nanotoxicology community in computational studies. A bridging physicochemical properties of NPs, experimental exposure conditions and in vitro characteristics with biological effects of NPs on a molecular cellular level from transcriptomics studies, is demonstrated. The bridging is achieved by developing and implementing Bayesian Networks with or without data preprocessing. Early stage nanotoxicity measurements represent a challenge, not least when attempting to predict adverse outcomes and modeling is critical to understanding the biological effects of exposure to NPs. In this thesis, categories of ML classifiers are compared to investigate their performance in predicting NPs in vitro toxicity. Physicochemical properties, toxicological and quantum-mechanical attributes and experimental conditions were used as input variables to predict the toxicity of NPs based on cell viability. Voting, an ensemble meta-classifier was used to combine base models to optimize the classification prediction of toxicity. To facilitate inter-comparison, a Copeland Index was applied that ranks the classifiers according to their performance and suggested the optimal classifier. In summary, this Thesis explores past work in the field, systematically capturing information regarding the data used in computational tools (Chapter 2). It demonstrates methodologies and the state-of-the-art approaches (Chapter 3) and creates an original Bayesian tool that can predict multiple toxicological outcomes in a molecular level from transcriptomics outcomes (Chapter 4). Finally, it develops and demonstrates a clever and compact methodology for researchers to compare and choose the optimal classifiers in their unique case of data (Chapter 5)