Predicting multi-class responses to preoperative chemoradiotherapy in rectal cancer patients

Abstract

Background: Preoperative chemoradiotherapy (CRT) has become a widely used treatment for improving local control of disease and increasing survival rates of rectal cancer patients. We aimed to identify a set of genes that can be used to predict responses to CRT in patients with rectal cancer. Methods: Gene expression profiles of pre-therapeutic biopsy specimens obtained from 77 rectal cancer patients were analyzed using DNA microarrays. The response to CRT was determined using the Dworak tumor regression grade: grade 1 (minimal, MI), grade 2 (moderate, MO), grade 3 (near total, NT), or grade 4 (total, TO). Results: Top ranked genes for three different feature scores such as a p-value (pval), a rank product (rank), and a normalized product (norm) were selected to distinguish pre-defined groups such as complete responders (TO) from the MI, MO, and NT groups. Among five different classification algorithms, supporting vector machine (SVM) with the top 65 norm features performed at the highest accuracy for predicting MI using a 5-fold cross validation strategy. On the other hand, 98 pval features were selected for predicting TO by elastic net (EN). Finally we combined TO- and MI-finder models to build a three-class classification model and validated it using an independent dataset of rectal cancer mRNA expression. Conclusions: We identified MI- and TO-finders for predicting preoperative CRT responses, and validated these data using an independent public dataset. This stepwise prediction model requires further evaluation in clinical studies in order to develop personalized preoperative CRT in patients with rectal cancer.