Predictive models of cytotoxicity as mediated by exposure to chemicals or drugs

Predicting cytotoxicity is a challenging task because of the complex biological mechanisms behind it. Cytotoxicity due to toxin - biologically produced poison - is known to play a substantial role in a disease process. Two objectives in this research are to derive robust general predictive cytotoxicity models to minimize unnecessary toxicity. The first objective is to build accurate predictive statistical models for cytotoxicity data based on lymphoblastoid cell lines obtained from in vitro studies. This could be an important step for accomplishing a goal in biomedecial/biophamarceutical research, by obtaining the best medical outcomes by minimizing toxicity in regard to a person's genetic profile. The second objective is to build predictive models to predict population-level cytotoxicity for unknown compounds based on chemical structural features. These two objectives were accomplished by a proposed variable selection process, the random forests, and the least absolute shrinkage and selection operator method. We achieved an excellent prediction result with the random forests algorithm using SNP markers from the proposed approach, having the smallest root mean squared error among the teams which participated in the DREAM Toxicogenetics Challenge. Since chemical compounds for drugs have great influence on human health, the predictive statistical models for these objectives could be helpful to government agencies in relevant decision-making.