Evaluation and Visualization of Radiogenomic Modeling Frameworks for the Prediction of Normal Tissue Toxicities

Abstract

We explore techniques for the evaluation and visualization of radiogenomic data-driven models in an effort to investigate the integration of genetic variations (single nucleotide polymorphisms [SNPs] and copy number variations [CNVs]) with dosimetric and clinical variables in modeling radiation-induced rectal bleeding (RB).

One hundred (N=100) patients who underwent curative hypofractionated radiotherapy (66 Gy in 22 fractions) between 2002-2010 were retrospectively genotyped for SNPs and CNVs in six genes: XRCC1, XRCC3, VEGFa, TGF≥1, ERCC2 and SOD2. A logistic regression modeling approach was used to assess the risk of severe RB (Grade..3) using dosimetric, clinical and biological variables. Statistical resampling based on cross-validation was used to evaluate model predictive power and generalizability to unseen data. NTCP-colorwashed principle component analysis (PCA) and vector biplots were used to visualize the quality of model fit.

Biological variable XRCC1 CNV showed good overall fit to RB outcome data (p<0.001). When added to the logistic regression modeling, XRCC1 CNV improved classification performance over standard dosimetric models by 33.5%. No clinical variables were found to adequately fit the data.

As a proof-of-concept, we demonstrated that the combination of genetic and dosimetric variables could provide significant improvement in NTCP prediction using data-driven approaches. Moreover, we have shown that visualization techniques aid in interpreting multivariate model predictions.