MicroRNA-based Biomarkers predicting Long-Term Toxicity to Prostate SBRT

Abstract

While toxicity rates following stereotactic body radiotherapy (SBRT) for localized prostate cancer do not appear to be greater than those from conventionally fractionated radiotherapy (CF-RT), a percentage of patients do experience significant genitourinary (GU) and gastrointestinal (GI) toxicity after either approach. Evidence suggests that radiation-associated toxicity can be predicted by patient-specific germ-line biomarkers. We hypothesized that we could apply microRNA-based germ-line biomarkers that would predict toxicity to prostate SBRT, and that they would be distinct from microRNA-based biomarkers predicting toxicity to CF-RT (submitted), affording the opportunity for guided treatment selection. Patients enrolled on two prospective phase II studies (NCT01059513 and NCT02296229) who received SBRT to the prostate only (8 Gy x 5) were included (n=93). DNA was extracted from blood, FFPE tissue or saliva, and a panel of microRNA-based germ-line mutations were evaluated. Machine learning techniques were used to simultaneously identify prognostic features and perform classification of the biomarkers. For each of the methods, 10-fold cross validation was used to assess its performance and generality. Four classifiers were studied: logistic regression with lasso regularization (LASSO-LR), deep classification tree (DT), random forest (RF) and boosted trees (BT), with corresponding hyper-parameters being regularization weights (LASSO-LR), split and node level (DT), number of trees (RF), and learning rate/tree level/splits (BT). We then developed a treatment support system to predict toxicity risk to SBRT versus CF-RT. Of the 93 SBRT patients, 16 (17.2%) had Grade ≥2 late (LT) GU and GI toxicity, but only 3 (3%) had acute Grade ≥2 toxicity, thus we focused on LT toxicity. Using DT we were able to predict LT GI and GU toxicity with a sensitivity of >96%, specificity >88%, AUC >93%, accuracy >95%, PPV >99%, NPV> 78%, and an F1 score of >97%. There was little overlap between biomarkers predicting LT GI verses GU toxicity for SBRT, and little overlap between biomarkers predicting LT GI or GU toxicity to SBRT versus CF-RT. Based on a patient's biomarker profile we performed toxicity prediction for each regimen (SBRT versus CF-RT) in parallel and assessed the likelihood of toxicity associated with each option. We could stratify patients into 3 subgroups: a subcohort whose toxicity outcome (either positive or negative) is indifferent to the treatment regimen (∼85%), a cohort where SBRT is safer than CF-RT (∼5%), and a cohort where CF-RT is safer than SBRT (∼10%). We have identified a microRNA-based germ-line biomarker signature that is able to predict LT GI and GU toxicity to SBRT, which is different than that predicting LT toxicity to CF-RT, and developed a classifier that can stratify patients based on the risk for toxicity, possibly guiding treatment selection. Ongoing work is being done to further validate these findings prospectively.