A Statistical Framework for the Development of Prediction and Clustering Models in the Hazard Assessment of Nanomaterials
Ying Pei · 2018
of Prediction and Clustering Models in the Hazard Assessment of Nanomaterials Ying Pei Compared to conventional materials or chemicals, it remains challenging to fully assess the hazards of various nanomaterials (NMs) stemming from their physicochemical properties.Tremendously large variety of NMs calls for high-throughput screening methods, and the ultimate goal of nanotoxicology is to develop prediction models, also referred to as QNAR (quantitative nanostructure-activity relationships), which relate the adverse bioactivity effects of NMs to their physicochemical properties.Such models enable the prediction of a new NM's toxicity without performing additional biological experiments, which leads to substantial savings in time and money.For the efficient development of prediction models, a statistical framework is provided and demonstrated in this dissertation.There are four stages of analysis and modeling in this framework: variable selection, design of experiments, quantitative modeling and shape clustering.Variable selection is first performed on existing nanotoxicology data to identify the important predictors, most of which are materials' physicochemical properties, for NMs' toxicity.Then design of experiments is carried out in the space of identified predictors for efficient data collection.Third, stochastic kriging with qualitative factors (SKQ) is employed to model the relationship between predictors and toxicity responses for the development of prediction models.Lastly, shape clustering methods are adapted to cluster NMs based on their toxicity profiles.This framework has been illustrated by a simulation case derived from a nanotoxicology database including 25 in-vivo studies for 1899 rodent animals.