Osteosarcoma is the most common malignant bone tumor in children. After initial diagnosis is made with a biopsy, treatment consists of preoperative chemotherapy followed by definitive surgery and postoperative chemotherapy. The degree of tumor necrosis in response to preoperative chemotherapy is a reliable prognostic factor and is used to guide the choice of postoperative chemotherapy. Patients with tumors, which reveal > or = 90% necrosis (good responders), have a much better prognosis than those with < 90% necrosis (poor responders). Despite previous attempts to improve the outcome of poor responders by modifying the postoperative chemotherapy, their prognosis remains poor. Therefore, there is a need to predict at the time of diagnosis patients' response to preoperative chemotherapy. This will provide the basis for developing potentially effective therapy that can be given at the outset for those who are likely to have a poor response. Here, we report the analysis of 34 pediatric osteosarcoma samples by expression profiling. Using parametric two-sample t test, we identified 45 genes that discriminate between good and poor responders (P < 0.005) in 20 definitive surgery samples. A support vector machine classifier was built using these predictor genes and was tested for its ability to classify initial biopsy samples. Five of six initial biopsy samples that had corresponding definitive surgery samples in the training set were classified correctly (83%; confidence interval, 36%, 100%). When this classifier was used to predict eight independent initial biopsy samples, there was 100% accuracy (confidence interval, 63%, 100%). Many of the predictor genes are implicated in bone development, drug resistance, and tumorigenesis.