Department of Biostatistics
Quantitative Issues in Cancer Research Working Seminar
2014 - 2015
ABSTRACT: To monitor quality of care in the US, the Centers for Medicare and Medicaid Services (CMS) currently reports, among other measures, hospital-specific 30-day readmission rates, estimated on the basis of a logistic-Normal GLMM. The focus of these efforts is on health conditions with low mortality, including pneumonia and heart failure. Expanding these efforts to include a broad range of increasingly prevalent 'advanced' health conditions, such as Alzheimer's disease and cancer, is problematic because the current CMS approach ignores death as a truncating event. A more appropriate analysis would be to frame quality of care assessments within the semi-competing risk framework although, to our knowledge, no statistical methods for clustered semi-competing risks data have been developed. We propose a novel semi-parametric hierarchical model for clustered semi-competing data based on an illness-death model. Estimation and inference is within the Bayesian paradigm, which facilitates the use of hospital-specific shrinkage targets and flexible random effects distributions. An efficient computational algorithm is developed, based on the Metropolis-Hastings-Green algorithm. The proposed framework is then applied to data on all individuals diagnosed with pancreatic cancer between 2005-2008 from Medicare Part A.
ABSTRACT: High-throughput sequencing data (DNA-Seq, RNA-Seq, etc.) can be used for estimating biological properties of interest (copy number of a genomic region, RNA abundance, etc.) but is not free from technical bias. Most troubling, the effect of technical bias can vary from sample to sample, and across experimental batches. I will review a number of examples of such known technical biases, and discuss solutions for recovering accurate statistical inference.
ABSTRACT: Oteosarcoma is a cancer with significant clinical heterogeneity. Genome-wide, "omics" studies could be very useful in deciphering the poorly understood molecular context of this tumor and may aid in identifying robust molecular markers for disease course and outcome. Using microRNA analyses, we have discovered highly prognostic profiles and also have identified a specific genomic locus as potentially very important in disease outcome. Furthermore, our data support the notion that the microRNA profiles are not simply prognostic but may be surrogates for a distinct molecular osteosarcoma subtype, perhaps driven by deregulated biology of the emerging critical genomic locus. Studies are planned to 1) externally validate the value of microRNAs, and develop clinically useful prognostic tools 2) study the "sequenome" of these microRNAs and 3) study the methylation patterns that could be partly responsible for modulating distinct subtypes and the clinical course of osteosarcoma.
ABSTRACT: Cancers arise through a process of somatic evolution. This evolutionary process can result in substantial clonal heterogeneity that has an impact on the tumor phenotype. The mechanisms responsible for the coexistence of distinct clonal lineages and the biological consequences of this coexistence remain poorly understood. Based on in vivo data from a mouse xenograft model, we investigate the influence of clonal heterogeneity on tumor properties, and mathematically model competitive expansion of individual clones. We find that tumor growth can be driven by a minor cell subpopulation. This minor population of cells enhances the proliferation of all cells within a tumor by overcoming environmental constraints. Yet, this driving cell population can be outcompeted by faster proliferating competitors. We describe how that non-cell autonomous driving of tumor growth supports clonal interference, stabilizes clonal heterogeneity and enables inter-clonal interactions, which can lead to new phenotypic tumor traits. When treatment is administered, heterogeneity can be reduced, also reducing evolutionary and metastatic potential. Another form of heterogeneity come about when we consider different cellular differentiation stages. If there is time, we will also present and discuss consequences of hierarchical tissue organization for tumor growth and treatment.
Amanda Randles, "Using Massively Parallel Simulation to Study Human Disease"
Lin Liu, "Stochastic Dynamics of Reprogramming to Induced Pluripotent Stem Cells"
ABSTRACT: None Given
ABSTRACT: Next-generation sequencing technologies and the increasing insight on the molecular origin of cancer types have boosted the development of new therapeutic agents that target specific molecular pathways and may have activity across cancer types. As a consequence, oncology trials have become more complex and changed the focus from single agent trials, which test treatment efficacy at the population level, to multi-agent trials with many sub-populations. To screen through a large number of therapies quickly and at the same time maintain high statistical power at the sub-population level, there is a need for novel statistical methods. This talk focuses on Bayesian methodology for designing early stage clinical trials. I will discuss the design and analysis of Phase II trials using Bayesian decision theory, hierarchical modeling and response-adaptive treatment assignment. Application of these statistical tools to multi-agent-multi-stage trials, biomarker-subgroup trials and surrogate-outcome models will be presented. It will be shown that Bayesian methodology can help save resources and reduce the time to drug discovery compared to standard balanced randomization trial designs.
ABSTRACT: Although our knowledge regarding genetic susceptibility is expanding exponentially through laboratory science, several critical factors need to be addressed by the scientific community to fully capitalize on these discoveries to transform mainstream medical care. In an increasingly fast-paced and complex medical world, health care providers need the help of population scientists to help integrate the meaning of genetic discoveries into clear and actionable endpoints that they can use in the care of their patients. I will discuss our prior work in prediction models for inherited colon cancer and discuss several critical challenges that need to be tackled in order to rationally provide and interpret genomic data for patient care. The broad goals of our work are to study the impact and potential role of germline genomics in terms of helping understand phenotypes of cancer by studying patients who have undergone testing of multiple cancer susceptibility genes via next-generation sequencing, integrate and assimilate the vast and ever increasing array of genetic markers into data-driven predictions of cancer risk for patients for a multitude of cancers over their lifetime, and harness this understanding into user-friendly electronic tools that can be used by patients and physicians via electronic health records to apply the genetic information into mainstream medical care for cancer risk assessment, and ultimately prevention.
ABSTRACT: RNA has diverse sets of regulatory functions besides being a messenger between DNA and protein. Recent analysis of RNA repertoire has identified a large numbers of non-coding transcripts. One of which, long intergenic non-coding RNA (lincRNA) with transcripts longer than 200 nucleotides, are located between the protein coding genes and do not overlap exons of either protein-coding or other non-lincRNA genes. lincRNAs have been considered to provide regulatory functions, however, their precise role in cellular biology remains unclear. I will review genomics tools to study lincRNAs for RNAseq data and I will discuss our focus on integrating lincRNAs with other genomic regulatory elements.
ABSTRACT: This talk presents a new computational method for linear mixed models having two variances, one for residuals and one for random effects. Researchers are often interested in either the restricted (residual) likelihood or the joint posterior density or their logarithms. Both functions can be multimodal and computations that rely on either a general purpose optimization algorithm or MCMC can fail to find regions where the target function is high. We present an alternative.
Letting f stand for the log of either the likelihood or the posterior, we show how to find a box B in the parameter space such that:
- all local and global maxima of f lie within B;
- f is guaranteed to be small outside of B; and
- f can be evaluated to within a prespecified tolerance epsilon everywhere within B.
Taken together these conditions imply that the parameter space can be divided into two parts: B, where we know f as accurately as we wish, and B-complement, where f is small enough to be safely ignored. We show how to find B and evaluate f.
|Back to SPH Biostatistics||
Maintained by the
Last Update: May 12, 2015