Purpose: Patients presenting with locally advanced rectal cancer currently receive preoperative radiotherapy with or without chemotherapy. Although pathologic complete response is achieved for approximately 10% to 30% of patients, a proportion of patients derive no benefit from this therapy while being exposed to toxic side effects of treatment. Therefore, there is a strong need to identify patients who are unlikely to benefit from neoadjuvant therapy to help direct them toward alternate and ultimately more successful treatment options. Experimental Design: In this study, we obtained expression profiles from pretreatment biopsies for 51 rectal cancer patients. All patients underwent preoperative chemoradiotherapy, followed by resection of the tumor 6 to 8 weeks posttreatment. Gene expression and response to treatment were correlated, and a supervised learning algorithm was used to generate an original predictive classifier and validate previously published classifiers. Results: Novel predictive classifiers based on Mandard's tumor regression grade, metabolic response, TNM (tumor node metastasis) downstaging, and normal tissue expression profiles were generated. Because there were only 7 patients who had minimal treatment response (>80% residual tumor), expression profiles were used to predict good tumor response and outcome. These classifiers peaked at 82% sensitivity and 89% specificity; however, classifiers with the highest sensitivity had poor specificity, and vice versa. Validation of predictive classifiers from previously published reports was attempted using this cohort; however, sensitivity and specificity ranged from 21% to 70%. Conclusions: These results show that the clinical utility of microarrays in predictive medicine is not yet within reach for rectal cancer and alternatives to microarrays should be considered for predictive studies in rectal adenocarcinoma.