Identifying Biomarkers to Predict Radiation Exposure

Long structure of the DNA double helix in depth of view.

Overview

Following a mass casualty nuclear event, the ability to precisely determine an individual’s radiation exposure (biodose) is critical for the accurate identification of critically exposed individuals for treatment. To support the development of an accurate biodose tool, Gryphon Scientific developed a supervised machine learning pipeline to identify the most important biomarkers (miRNA and mRNA signals) for predicting radiation exposure. This work was performed in collaboration with the Experimental Therapeutics Section at the National institutes of Health (NIH) National Cancer Institute (NCI).

Methods

Gryphon analyzed data from a set of experiments in mice measuring changes in mRNA and miRNA expression following different levels of radiation exposure using microarrays. The experimental design included multiple radiation doses and timepoints post-exposure, yielding microarray data featuring 3,000 mRNA and 24,000 miRNA probe values. The microarray datasets included a high number of features (miRNA/mRNA signals) but a low number of observations per radiation exposure category (mice in each experimental group), presenting analytic challenges. We implemented an ensemble of state-of-the-art feature selection algorithms to identify the miRNA and mRNA signals that change the most (increase or decrease) with different levels of radiation exposure, representing the first time some of these algorithms had been applied to microarray data. We selected the top 20-30 signals that were able to differentiate between different levels of radiation exposure with the highest level of predictive accuracy in a predictive classification model. Models were developed with and without time point data (knowing how long post-exposure the blood samples were taken) to better understand the signal change over time following radiation exposure. With these signals, Gryphon developed and validated a predictive classifier of radiation exposure using support vector machine and random decision forest algorithms.

 Results

The results were analyzed for their biological relevance to radiation injury and were used to advance the NCI’s goal towards understanding biological pathways that are most affected by radiation exposure. This work will inform the design of future animal radiation exposure response experiments and is an important step in the development of a biodosimeter following a mass casualty radiation event. Feature selection algorithms were used to identify the 23 miRNAs and 22 mRNAs with the greatest expression changes in response to radiation exposure in a time and/or dose dependent manner. Classification models based on the miRNA or mRNA signatures were developed and validated using a test dataset; accuracy in the predictive models was estimated at 56% and 93%, respectively.

 Resources

The paper, “Microarray analysis of miRNA expression profiles following whole body irradiation in a mouse model” was published in the Biomarkers Journal in May, 2018.