Arwa Bin Raies is currently a Ph.D. Candidate in Computer Science at King Abdullah University of Science and Technology. She joined the Computational Bioscience Research Center under the supervision of Professor Vladimir Bajic, and she was in charge of projects related to text mining of biomedical literature to extract associations between methylated genes and disease in various species, and development of biological databases. Currently, she works on In Silico Toxicology to develop computational models to predict chemicals toxicity. Additionally, Arwa worked as a Teaching Assistant of several Computer Science courses including Programming Languages, High Performance Computing and Operating Systems.
________________________________________________________________________________________________________
Predicting adverse effects of toxic compounds computationally can contribute to overcoming some shortcomings of existing safety assessments. In this study, we generated and evaluated the performance of about 20,000 predictive models using combinations of many computational algorithms to classify compounds to 17 in vivo toxicity endpoints using in vitro data and mechanism of action. A challenging aspect of this study is dealing with incompletely labeled datasets (i.e., some toxicity endpoints are unknown for some compounds), which is a common problem in many toxicity datasets. This limitation can inhibit algorithms' ability to identify correlations between toxicity endpoints and may reduce their performance. Recently, some predictive algorithms were applied to such toxicity datasets. However, in these studies different datasets, pre-processing steps or performance evaluation metrics were used. Such inconsistent assessment conditions render proper comparison and reproducibility of models performance infeasible. Therefore, the goal of this study is to provide a systematic evaluation of many predictive models to identify causes of variability in predictive performance across endpoints and compounds. The results of this thorough analysis illustrated some models ability to predict toxicity with accuracy better than that of the baseline approach. The best performing model was able to predict the toxicity of compounds with accuracy values 89.96% and 90.59% for internal evaluation (cross-validation on the training set) and external evaluation (blind testing set), respectively, while the baseline approach achieved accuracy values 72.20% and 68.43%, respectively. However, some models showed performance disparities between internal and external evaluations or exhibited a considerable predictive variability across endpoints and compounds.