Outcome prediction in male germ cell tumors based on gene expression signatures

2632

Greater than 90% of newly diagnosed cases of male germ cell tumors (GCTs) will be cured by cisplatin based therapies. Although markers such as high serum levels of HCG and AFP identify patients at high risk for treatment failure, we sought to identify gene expression signatures that could more accurately predict outcome in patients with GCTs. Expression profiling was performed on a panel of 101 GCT specimens using Affymetrix U133A+B microarrays. Of these 101 tumors, there were 86 non-seminomatous GCTs, including 73 patients who had received chemotherapy and had follow up of sufficient length for outcome analysis. These patients were divided into two categories: a good outcome (sensitive) category consisting of those who had complete response and no evidence of disease for at least two years, and a poor outcome (resistant) category consisting of patients who died of disease. Prediction Analysis for Microarrays (PAM) was unable to separate tumors into good and poor prognosis groups based on gene expression when the entire panel of tumors was examined. Instead, we stratified tumors using an expression based classifier that we have recently developed into embryonal carcinomas (EC), teratomas (T), and Yolk Sac (YS) groups for outcome prediction. Choriocarcinoma samples were excluded since there were insufficient samples for outcome analysis. PAM was able to identify small sets of genes that accurately predicted outcome in each histologic subtype. For 31 EC tumors, PAM identified a set of 45 transcripts with a classification rate of 82%. PAM identified a set of 32 transcripts that correctly classified 80% of the 15 T samples, while there were 37 transcripts that correctly classified 86% of the 21 patients with YS tumors. Interestingly, several sensitive patients who were classified as resistant by the gene classifier had either metachronous tumors develop or had late relapse of their disease. The predictive gene sets showed significant overlap between EC and T, with a total of 6 transcripts in common. In contrast, there was no overlap in the YS predictive gene set and the gene sets for EC and T. Kaplan-Meier analysis revealed significant differences in the survival curves of the predicted good and poor outcome groups for all three tumor types. In summary, we have identified gene expression signatures which are able to predict with a high degree of accuracy patients with EC, T, or YS tumors who are likely to succumb to disease.