Optimizing bags of artificial neural networks for the prediction of viability from sparse data

The prediction of organismal viability in response to exposure to different nanoparticles and conditions characterized at the molecular scale is challenging because several orders of magnitude must be bridged. A so-called bag of artificial neural networks has recently been shown to provide such a connection when trained through the use of relatively small databases. In the present work, we found that individual neural networks do not always converge under training. The use of bags and weighted means for consensus predictions provides a mechanism for effectively pruning the effects from the diverging networks without requiring prior conditioning. The optimized structure of these machines was also found to be effective in predicting the relative viability of nanoparticles that had not been used in developing the major findings of this work.