Using Machine Learning Techniques for Asserting Cellular Damage Induced by High-LET Particle Radiation

Abstract

This is a study concerning the use of Machine Learning (ML) techniques to ascertain the impacts of particle ionizing radiation (IR) on cell survival and DNA damage. Current empirical models do not always take into account intrinsic complexities and stochastic effects of the interactions of IR and cell populations. Furthermore, these models often lack in biophysical interpretations of the irradiation outcomes. The linear quadratic (LQ) model is a common way to associate the biological response of a cell population with the radiation dose. The parameters of the LQ model are used to extrapolate the relation between the dosage and the survival fraction of a cell population. The goal was to create a ML-based model that predicts the α and β parameters of the well known and established LQ model, along with the key metrics of DNA damage induction. The main target of this effort was, on the one hand, the development of a computational framework that will be able to assess key radiobiophysical quantities, and on the other hand, to provide meaningful interpretations of the outputs. Based on our results, as some metrics of the adaptability and training efficiency, our ML models exhibited 0.18 median error (relative root mean squared error (RRMSE)) in the prediction of the α parameter and errors of less than 0.01 for various DNA damage quantities; the prediction for β exhibited a rather large error of 0.75. Our study is based on experimental data from a publicly available dataset of irradiation studies. All types of complex DNA damage (all clusters), and the number of double-stranded breaks (DSBs), which are widely accepted to be closely related to cell survival and the detrimental biological effects of IR, were calculated using the fast Monte Carlo Damage Simulation software (MCDS). We critically discussed the varying importance of physical parameters such as charge and linear energy transfer (LET); we also discussed the uncertainties of our predictions and future directions, and the dynamics of our approach.