Pretreatment lymphoma biopsy samples were studied according to a protocol approved by the NCI Institutional Review Board. Lymphoma biopsies were reviewed by a panel of hematopathologists and were found to be DLBCLs morphologically. A “training” set of cases consisted of 36 biopsy specimens from 35 patients for whom the diagnosis of PMBL was considered. These patients all had mediastinal masses of at least 6 cm at presentation. These samples were profiled for gene expression using Lymphochip DNA microarrays comprised of 15,133 cDNA elements as described (3 (link)), and the data are available at http://llmpp.nih.gov/PMBL. A “validation” set of 274 lymphoma samples was previously profiled using Lymphochip DNA microarrays comprised of 12,196 cDNA elements (4 (link)); data for these samples were obtained from http://llmpp.nih.gov/DLBCL. All patients were treated with anthracycline-containing multiagent chemotherapy protocols with some patients additionally receiving radiation therapy.
The Bayesian statistical procedure used to create the gene expression-based PMBL predictor has been described (5 ). In the training set of cases, a Bayesian PMBL predictor was constructed from the 46 genes shown in Fig. 2 A. Since cases in the validation set were profiled on Lymphochip microarrays that lacked some of these genes, we constructed another Bayesian PMBL predictor using the 26 discriminating genes that were represented on these microarrays. After demonstrating that this predictor performed identically to the 46-gene predictor on the training set (not shown), it was then used to classify cases in the validation set of cases without reoptimization of the model parameters (Fig. 2 B).
Survival probabilities were estimated using the Kaplan-Meier method. P values for survival differences were evaluated using a log-rank test. P values for differences in age group and gender were generated using a chi-squared test. P values for differences in age as a continuous variable were computed using an ANOVA test.