Faculty Bibliography

MOTIVATION/BACKGROUND:We propose a Bayesian ensemble method for survival prediction in high-dimensional gene expression data. We specify a fully Bayesian hierarchical approach based on an ensemble 'sum-of-trees' model and illustrate our method using three popular survival models. Our non-parametric method incorporates both additive and interaction effects between genes, which results in high predictive accuracy compared with other methods. In addition, our method provides model-free variable selection of important prognostic markers based on controlling the false discovery rates; thus providing a unified procedure to select relevant genes and predict survivor functions. RESULTS:We assess the performance of our method several simulated and real microarray datasets. We show that our method selects genes potentially related to the development of the disease as well as yields predictive performance that is very competitive to many other existing methods. AVAILABILITY/BACKGROUND:http://works.bepress.com/veera/1/.

Primary brain tumors are a heterogeneous group of malignancies with highly variable outcomes, and diagnosis is largely based on the histological appearance of the tumors. However, the diversity of primary brain tumors has made prognostic determinations based purely on clinicopathologic variables difficult. There is an increasing body of data suggesting a significant amount of molecular diversity accounts for the heterogeneity of clinical observations, such as response to treatment and time to progression. The last decade has witnessed an explosive advance in our knowledge of the molecular genetics of brain tumors, due in large part to the availability of high-throughput profiling techniques and to the completion of the human genome sequencing project. The large amount of data generated by these efforts has enabled the identification of prognostic and predictive factors and helping to identify pathways which are driving tumor growth. Identification of biomarkers will enable better patient stratification and individualization of treatment.

The diffuse gliomas are a heterogeneous group of malignancies with highly variable outcomes and diagnosis is largely based on the histological appearance of the tumors. Tumor classification according to cello type and grade provides some prognostic information. However, the diversity of gliomas, within tumor type and grade categories, has made prognostic determinations based purely on clinicopathologic variables difficult. There is an increasing body of data suggesting a significant amount of molecular diversity accounts for the heterogeneity of clinical observations, such as response to treatment and time to progression. The last decade has witnessed an explosive advance in our knowledge of the molecular genetics of brain tumors, due in large part to the availability of high-throughput profiling techniques, including new sequencing methodologies as well as multidimensional profiling by the Cancer Genome Atlas project. The large amount of data generated by these efforts has enabled the identification of prognostic and predictive factors and helping to identify pathways that are driving tumor growth. Identification of biomarkers, especially when coupled to clinical trials of newer targeted therapies, will enable better patient stratification and individualization of treatment.

We have profiled promoter DNA methylation alterations in 272 glioblastoma tumors in the context of The Cancer Genome Atlas (TCGA). We found that a distinct subset of samples displays concerted hypermethylation at a large number of loci, indicating the existence of a glioma-CpG island methylator phenotype (G-CIMP). We validated G-CIMP in a set of non-TCGA glioblastomas and low-grade gliomas. G-CIMP tumors belong to the proneural subgroup, are more prevalent among lower-grade gliomas, display distinct copy-number alterations, and are tightly associated with IDH1 somatic mutations. Patients with G-CIMP tumors are younger at the time of diagnosis and experience significantly improved outcome. These findings identify G-CIMP as a distinct subset of human gliomas on molecular and clinical grounds.

Hypermethylation of the O(6)-methylguanine-DNA-methyltransferase (MGMT) gene has been shown to be associated with improved outcome in glioblastoma (GBM) and may be a predictive marker of sensitivity to alkylating agents. However, the predictive utility of this marker has not been rigorously tested with regard to sensitivity to other therapies, namely radiation. To address this issue, we assessed MGMT methylation status in a cohort of patients with GBM who underwent radiation treatment but did not receive chemotherapy as a component of adjuvant treatment. Formalin-fixed, paraffin-embedded tumor samples from 225 patients with newly diagnosed GBM were analyzed via methylation-specific, quantitative real-time polymerase chain reaction following bisulfite treatment on isolated DNA to assess MGMT promoter methylation status. In patients who received radiotherapy alone following resection, methylation of the MGMT promoter correlated with an improved response to radiotherapy. Unmethylated tumors were twice as likely to progress during radiation treatment. The median time interval between resection and tumor progression of unmethylated tumors was also nearly half that of methylated tumors. Promoter methylation was also found to confer improved overall survival in patients who did not receive adjuvant alkylating chemotherapy. Multivariable analysis demonstrated that methylation status was independent of age, Karnofsky performance score, and extent of resection as a predictor of time to progression and overall survival. Our data suggest that MGMT promoter methylation appears to be a predictive biomarker of radiation response. Since this biomarker has also been shown to predict response to alkylating agents, perhaps MGMT promoter methylation represents a general, favorable prognostic factor in GBM.

The inference of transcriptional networks that regulate transitions into physiological or pathological cellular states remains a central challenge in systems biology. A mesenchymal phenotype is the hallmark of tumour aggressiveness in human malignant glioma, but the regulatory programs responsible for implementing the associated molecular signature are largely unknown. Here we show that reverse-engineering and an unbiased interrogation of a glioma-specific regulatory network reveal the transcriptional module that activates expression of mesenchymal genes in malignant glioma. Two transcription factors (C/EBPbeta and STAT3) emerge as synergistic initiators and master regulators of mesenchymal transformation. Ectopic co-expression of C/EBPbeta and STAT3 reprograms neural stem cells along the aberrant mesenchymal lineage, whereas elimination of the two factors in glioma cells leads to collapse of the mesenchymal signature and reduces tumour aggressiveness. In human glioma, expression of C/EBPbeta and STAT3 correlates with mesenchymal differentiation and predicts poor clinical outcome. These results show that the activation of a small regulatory module is necessary and sufficient to initiate and maintain an aberrant phenotypic state in cancer cells.

Only a subset of patients with newly diagnosed glioblastoma (GBM) exhibit a response to standard therapy. To date, a biomarker panel with predictive power to distinguish treatment sensitive from treatment refractory GBM tumors does not exist. An analysis was performed using GBM microarray data from 4 independent data sets. An examination of the genes consistently associated with patient outcome, revealed a consensus 38-gene survival set. Worse outcome was associated with increased expression of genes associated with mesenchymal differentiation and angiogenesis. Application to formalin fixed-paraffin embedded (FFPE) samples using real-time reverse-transcriptase polymerase chain reaction assays resulted in a 9-gene subset which appeared robust in these samples. This 9-gene set was then validated in an additional independent sample set. Multivariate analysis confirmed that the 9-gene set was an independent predictor of outcome after adjusting for clinical factors and methylation of the methyl-guanine methyltransferase promoter. The 9-gene profile was also positively associated with markers of glioma stem-like cells, including CD133 and nestin. In sum, a multigene predictor of outcome in glioblastoma was identified which appears applicable to routinely processed FFPE samples. The profile has potential clinical application both for optimization of therapy in GBM and for the identification of novel therapies targeting tumors refractory to standard therapy.

The analysis of high-throughput data sets, such as microarray data, often requires that individual variables (genes, for example) be grouped into clusters of variables with highly correlated values across all samples. Gene shaving is an established method for generating such clusters, but is overly sensitive to the input data: changing just one sample can determine whether or not an entire cluster is found. This paper describes a clustering method based on the bootstrap aggregation of gene shaving clusters, which overcomes this and other problems, and applies the new method to a large gene expression microarray dataset from brain tumour samples.