UCLA Radiation Oncology Los Angeles, CA, United States
J. B. Weidhaas1, K. McGreevy2, K. A. Goodman3, X. Qi1, A. Raldow1, and D. Telesca2; 1Department of Radiation Oncology, University of California, Los Angeles, Los Angeles, CA, 2Department of Biostatistics, Fielding School of Public Health, University of California, Los Angeles, Los Angeles, CA, 3Department of Radiation Oncology, Icahn School of Medicine at Mount Sinai, New York, NY
Purpose/Objective(s):In the treatment of locally advanced rectal cancer (LARC) there are several approaches considered standard of care, including chemoradiotherapy (CRT) followed by surgery with or without adjuvant chemotherapy, chemotherapy without radiation followed by surgery, and CRT with consolidative or induction chemotherapy (total neoadjuvant treatment, TNT) followed by surgery or non-operative management for those who achieve a clinical complete response (CR). However, there has not previously been a molecular based assay to predict which patients are most likely to achieve a pathologic CR (pCR) to CRT, to help with treatment selection. Here we applied recently identified germline biomarkers that disrupt microRNA-signaling, called miRSNPs, to determine if they were associated with a pCR to CRT. Materials/
Methods: The study included 90 LARC patients receiving neoadjuvant CRT before surgery, with 75 receiving CRT and 15 receiving TNT treatment. Patient age ranged from 33 to 87 years (median 58 years). 54 patients were male and 36 were female. pCR was defined as an absence of viable tumor cells in the resection specimen (score = 0). In this study, 18 (20%) people had a pCR, and 26, 33, and 13 people had residual tumor burden of 1, 2, and 3, respectively. miRSNPs were included as potential covariates if they were marginally associated with complete response via Fisher’s Exact Test or Jonckheere-Terpstra Test using the union SURE screening procedure at a p-value threshold of < 0.3. For both tests p-values were obtained using Monte Carlo procedures, providing robust calculations in the case of small cell counts. Results: 34 out of 116 miRSNPs were marginally associated with pCR via Fisher’s Exact and/or Jonckheere-Terpstra tests at p-value < 0.3. These 34 SNPs along with clinical variables including age, tumor stage, nodal status and tumor KRAS mutation status were included for building a preliminary pCR signature. Several classification models were trained using 10-fold Cross Validation (CV) with upsampling of cases. The constructed miRSNP/clinical pCR signature, an Elastic Net (EN) model, had a nested Leave One Out Cross Validation (LOOCV) AUC of 0.722. This surpassed pCR performance of the best clinical model, an EN model with a nested LOOCV AUC of 0.639. Out of the variables included in the model, the only meaningful clinical variable was KRAS tumor mutational status. In total, there were 8 important miRSNPs that improved the model AUC. The three most important were in ABL1, MDM2, and FOXP3. Conclusion: We have identified a miRSNP signature predicting a pCR to CRT in LARC. This signature will be further validated in a larger cohort and also evaluated as a predictor of toxicity to CRT. Table 1. LOOCV Performance Metrics for pCR Models PPV= positive predictive value; NPV = negative predictive value; AUC = area under the curve
