Abstract P042: A genomic model to predict cellular radiosensitivity

Abstract

Abstract Purpose: This group has previously worked on Radiosensitivity Index (RSI),1,2 which has been clinically validated in 1919 patients across multiple cohorts.3 The Linear Quadratic (LQ) model is a widely accepted model used to prescribe radiation dose and fractionation. We hypothesize that clonogenic data can be fitted to the linear quadratic equation to predict α and β based on gene expression. Here we introduce a novel genomic model (gLQ) developed in cell lines to estimate the α and β parameters in cell lines. Methods: We analyzed public data on 494 irradiated clonogenic cell lines,4 considering Survival Fraction (SF) at known radiation dosages (2-10 Gy, i.e., SF2 through SF10) and fractions to calculate “ground truth” α and β using RAD-ADAPT software.5 Ground truth α and β were aligned to publically available Affymetrix U133 2.0 plus microarray and RNA seq data. Normalization was performed by Robust Multi-array average and RSEM respectively. 90% of the 494 cell line data (n=444) was used to train the model, while 10% (n=50) was left untouched. To narrow the data on 18,468 genes for each cell line to a smaller set of representative genes, we used a combination of clustering analysis, feature selection using Max-min Markov Blanket, and evaluating the functionality of selected features as they relate to radiosensitivity and cancer.6 Limiting the genes used as inputs in the machine-trained regression improved model training time. Machine learning, specifically Bayesian Ridge Regression, was utilized to identify relationships between calculated α and β and the representative gene expression profile for each cell line. 598 genes were selected for the α model, and 1198 genes were selected for the β model. These models were then locked down and tested on the 50 untouched validation cell lines. Results: Calculated ground truth α and β had a strong fit (R2=0.94). The resulting trained, locked-down model predicted α and β values for the remaining unseen 10% of the cell lines as a separate validation cohort with R2 values of 0.8809 and 0.8175, respectively. Conclusions: This suggests that genomic data can be used to effectively predict cellular radiosensitivity.