Comparative analysis of predictive models used to identify drivers of nanomaterial toxicity

Understanding the inherent and conditional factors associated with nanomaterial toxicity is critical to the development nanotechnologies that pose minimal threats to humans and the environment over the life cycle of the nanomaterial.  We compared the results of several different models built to predict toxicity from the open-source data on nanomaterial toxicity to embryonic zebrafish (Danio rerio) found in the Nanomaterial-Biological Interactions (NBI) knowledgebase at Oregon State University.  Model comparisons included the ABMiner predictive models, MATLAB clustering analysis and the use of Self-Organizing Map (SOM) based consensus clustering conducted on the data in the NBI knowledgebase (nbi.oregonstate.edu). Overall results suggest that exposure concentration and outermost surface chemistry (and thus surface charge) both should be considered in conjunction with the core composition of nanomaterials when trying to develop predictive models for developing zebrafish.  Core composition was found to be a significant contributor to the ABMinor predictive model. MATLAB clustering grouped materials into two clusters with outermost surface chemistry being the primary determinant.  The SOM modeling identified five significant clusters (clustering index = 0.89); while no core materials occurred in all 5 clusters, 4 material types (based on core composition) occurred in 4/5 clusters and 15 material types occurred in 3/5 clusters. In addition, over half of the materials appeared in multiple clusters depending on the dose applied.  Thus, classification of nanomaterials by simple descriptors such as core composition may not be sufficient for predicting nanomaterial toxicity or managing nanomaterial risks.