MCCX: Combining Monte Carlo, Convolutional Neural Networks and Boosting in Predicting Nanoparticle Viability

Abstract Modeling and prediction relying on learning from small experimental data sets is challenging, especially when the data set is insufficient, noisy or involves highly complex relationships. Leveraging Neural Networks (NNs), we developed predictive models capable of addressing such small data sets. We introduce MCCXcombining Monte Carlo (MC) sampling, Convolutional Neural Networks (CNNs), and Extreme Gradient-Boosting (XGB) as one such model in the context of learning and predicting the viability of organisms exposed to nanoparticles. Our approach can achieve higher accuracy when given smaller data sets. It gives competitive or lower prediction errors than the benchmark models considered here and substantially improves upon our previous k-bagging Artificial Neural Network (ANN) model. The strong performance of MCCX demonstrates that complex nonlinear neural network approaches can maximize the use of every data point and feature as we have verified for the prediction of viability from a small data set of nanoparticles, exposure conditions and organisms. We thus posit that MCCX enables the efficient learning from small-scale experimental data sets for predicting properties within the descriptor space represented by the available data.