2008 Rutgers Biostatistics Day.

A conference to celebrate Biostatistics Research

Institute of Biostatistics
Department of Statistics & Biostatistics,
Rutgers University

Time: Friday, April 25, 2008, 9:00AM
Place: Fiber Optics Auditorium, Busch Campus, Rutgers.

Abstracts


Data Mining for Drug-Adverse Event Associations
William DuMouchel Lincoln Technologies

Abstract

Methods for searching for drug-adverse event associations will be reviewed. Three types of data are considered: spontaneous reports, collections of clinical trials, and longitudinal medical records. Bayesian shrinkage methods that help adjust for the post-hoc selection of "unexpected but interesting" associations will be discussed, as well as the application of logistic regression and related methods to such data.


Developing Genomic Biomarkers for Early Drug Safety Screens
Nandini Raghavan, Johnson & Johnson Pharmaceutical Research and Development

Abstract

Drug safety is an area of critical importance to pharmaceutical companies in their attempt to cut down the costs of developing new drugs. Drug toxicity issues, which usually do not come to light until late in the development process, are often the cause of drug failure in late-stage development. This is especially true for carcinogenicity, since studies typically take a long time to complete. By improving the screening process for new drugs and accelerating the prediction of toxicity issues to early in the development process, companies not only reduce the downstream risk of compound failure due to safety concerns; but can also prioritize their lead optimization more efficiently. Attempts are currently underway to develop quicker, better and earlier safety screens for drugs by utilizing new high throughput "omics" technologies, like microarrays. In this talk we discuss some of the statistical considerations that arise in designing microarray-based experiments for predicting toxicity. We describe a procedure for classifying the toxicity of compounds using gene expression data and present a case-study for identifying non-genotoxic carcinogenicity using 24 hour experiments. We also discuss some of the complexities involved in developing gene-expression based biomarkers, or gene signatures, for toxicity.


Measuring Signaling Activity in Cancer Cells from Microarray Data.
Michael Ochs, Johns Hopkins University.

Abstract

Cellular signaling plays a critical role in carcinogenesis and is an extremely complex process that is still being elucidated. Nevertheless, understanding changes in signaling activity, especially as more therapeutics specifically target signaling proteins in cancer and other diseases, is critical to the development of personalized medicine and the development of novel therapies. In general, it is difficult to directly measure signaling protein states (e.g., phosphorylation) in vivo, however it has become routine to obtain global mRNA profiles with microarrays. We have developed techniques for isolating overlapping mRNA signatures using Bayesian Markov chain Monte Carlo, and we have extended these methods to direct estimation of transcription factor activity and the linking of this activity to changes in cell signaling in cancer cells during treatment. We demonstrate this approach using imatinib mesylate (Gleevec) treatment of a gastrointestinal tumor cell line.

Statistical challenges in Utilizing Gene Expression and Genetics in Biomarker Research
Peggy Wong, Peter Hu, Xiang Yu Merk Inc.

Abstract

The first statistical challenge when translating genetic or gene expression data involves reducing the high dimensional genomic data to a meaningful signature, pathway or composite score. The second statistical challenge is to attempt to translate the information found in the preclinical model into a clinical proof of concept study and that involves assumptions made to power that gene signature, pathway or composite score. Actual clinical examples will be utilized to highlight challenges being faced by the analysis teams to find and then validate genomic biomarkers for use in drug development.


Simultaneous Model Selection via Rate-Distortion Theory, with Applications to Clustering and Significance Analysis of Gene Expression Data
Rebecka Jornsten, Department of Statistics, Rutgers University

Abstract

The analysis of gene expression data present many challenges that can be formulated as model selection problems. In model-based clustering, we group genes that exhibit similar expression profiles across experimental conditions. To allow for direct and objective inference of the clustering outcome, we need to determine a sparse representation of each cluster; between which experimental conditions does the cluster expression profile truly differ? Model selection in clustering is combinatorial in the number of clusters and the number of experimental conditions, and thus presents a computationally challenging task. We introduce a simultaneous approach to subset model selection, which draws on results from rate-distortion theory. The rate-distortion formulation allows us to turn the combinatorial model selection into a fast and simple line search. Furthermore, by considering each gene as its own cluster, the simultaneous selection framework extends to significance analysis of differential expression. We can thus determine not only if a gene is differentially expressed, but also which are the disciminatory experimental conditions. These days, data often have a complex structure, and the clustering techniques we apply should reflect this. We introduce multi-level mixture models to address this issue. The multi-level framework can incorporate multiple distance metrics into clustering simultaneously, and be used to analyze multi-factor experiments. Multi-level mixture models extend model selection in clustering to between-cluster comparisons, and can constitute a substantial savings of model parameters, allowing for more clusters to be detected than with standard clustering techniques.

Interim treatment selection using normal approximation in clinical trials.
Zhenming Shun, Sanofi-Adventis

Abstract

We consider a study starting with two treatment groups and a control group. During the planned interim Analysis, the inferior treatment group will be dropped after the interim analysis, and only the winning treatment and the control will continue to the end of the study. This 'Two-Stage Winner Design' is based on the concepts of multiple comparison, adaptive design, and winner selection. In a study with such a design, there is less multiplicity, but more adaptability if the interim selection is performed at an early stage. If the interim selection is performed close to the end of the study, the situation becomes the conventional multiple comparison where Dunnett's method may be applied. The unconditional distribution of the final test statistic from the 'winner' treatment is no longer normal, the exact distribution of which is provided in this paper, but numerical integration is needed for its calculation. To avoid complex computations, we propose a normal approximation approach to calculate the type I error, the power, the point estimate, and the confidence intervals. Due to the well understood and attractive properties of the normal distribution, the 'Winner Design' can be easily planned and adequately executed, which is demonstrated by an example. We also provide detailed discussion on how the proposed design should be practically implemented by optimizing the timing of the interim look and the probability of winner selection.

On Adaptive Extensions of Group Sequential Trials
Qin Liu,Johnson & Johnson PRD, Keaven Anderson, Merck Inc

Abstract

In group sequential trials, it is important to obtain adequate data to assess overall benefits and risks for patients using the medical product. To achieve this goal, we provide a general, formal framework for adaptively extending a group sequential trial to stop at any interim analysis time, often after a significance boundary for a clinical endpoint is crossed. For statistical inference, we propose to order the sample space by a class of well-ordered group sequential tests. On this basis, we develop a unified sequential statistical inference approach that is applicable to both interim monitoring and final analysis. We also show that the new ordering provides the foundation for the repeated confidence intervals procedure of Jennison and Turnbull (1989).
Please direct all questions regarding this conference webpage to collins@stat.rutgers.edu
Last updated: April 24, 2008, 12:59 pm

© 2008 Rutgers, The State University of New Jersey. All rights reserved.