Spring 2008 Meeting

Held on March 13, 2008 at the Wyndham Glenview Suites.

The Program consisted of three presentations:

  1. A new method to identify significant endpoints in a closed test setting
    Carlos Vallarino, Ph.D., Takeda
  2. Challenges in statistical modeling of chromatin sequence for Eukaryotic species
    Ji-Ping Wang, Ph.D., Northwestern University
  3. Important Statistical considerations in biomarker discovery and clinical evaluations  
    Viswanath Devanarayan, Ph.D., Abbott

A new method to identify significant endpoints in a closed test setting by Carlos Vallarino, Ph.D., Takeda

Biographical Background
When testing several hypotheses simultaneously, the challenge is to combine them into a multiple test procedure that controls the familywise error rate, or overall α. One useful approach is the closure method. We develop a powerful statistical test for use with the closure method for testing co-primary endpoints in clinical trials, when there is a common effect direction and the test statistics are normally distributed. We show that the simple sum test is maximin, then alter its rejection region to make it consonant, i.e., to guarantee that rejection of the intersection hypothesis implies that at least one endpoint is significant. This new test has greater power and retains the maximin property. Our application to PROactive, a CV-outcome trial of patients with Type 2 DM and CV disease history, shows how efficacy for one key endpoint could have been established. Taking the primary and secondary endpoints as co-primary, in a closed test design, both the simple sum test and the consonant sum test yield p-values below the allocated  α.

Dr. Carlos Vallarino has been working in the pharmaceutical industry for the past 7 years. Following 3 years in Phase II/III clinical trials at Eli Lilly and Pharmacia, he joined Takeda's Outcomes Research group in 2003, where the emphasis is on the analysis of Phase IV outcomes from claims databases. He recently transferred to the Epidemiology group, where he has expanded his responsibilities to include the statistical analysis of safety data, particularly early signal detection. Dr. Vallarino holds a Bachelor's degree in Mathematics, Economics and Statistics from the State University of New York at Buffalo, Master's and Ph.D. degrees in Statistics from the University of California at Berkeley.

Challenges in statistical modeling of chromatin sequence for Eukaryotic species by Ji-Ping Wang, Ph.D., Northwestern University

Biographical Background
Dr Ji-Ping Wang earned his Ph.D. degree in statistics from Penn State University in 2003. He has been an assistant professor in statistics at Northwestern University since then. His current research focuses on two areas:  (1) mixture model, computing algorithms and applications; (2) bioinformatics and computational biology. His research topics include penalized NPMLE and algorithm for species richness estimation, Expressed Sequence Tag (EST) data analysis, nucleosome sequence alignment and prediction and human braining mapping. He has been publishing in journals including JASA, Nature, Bioinformatics, BMC Bioinformatics, NAR, PLoS

As evidence aggregates showing that DNA sequence itself is the determinant factor in nucleosome positioning in Eukaryotic cells, statistical modeling of chromatin sequence remains exceedingly challenging.  The challenges arise as a consequence of the fact that the signals intrinsic in nucleosome sequences are very weak; in addition knowledge on the linker DNAs, which are interwoven with nucleosome motifs in chromatin fiber, is very limited. In particular the linker length, which determines the orientation of adjacent nucleosomes, is of fundamental importance in understanding chromatin structure. We investigate the linker length distribution of two Eukaryotic species including yeast and human using two novel methods: a Fourier analysis of dinucleotide frequency in the extended region of nucleosome core particles and a duration Hidden Markov Model (DHMM) for dinucleosome sequences. Both methods conclude that the linker length distribution is not uniform but periodic.  The DHMM method further shows that the linker length prefers a form 10n + d0, where d0 =  5 bp for yeast as opposed to 10 bp for  human overwhelmingly.

Important Statistical considerations in biomarker discovery and clinical evaluations by Viswanath Devanarayan, Ph.D., Abbott

Biographical Background
Dr. Viswanath Devanarayan received his Ph.D. in Statistics from NC State, 1996.  He has 12 years of experience in the pharmaceutical industry, spanning Lilly, Merck and Abbott.  His primary areas of focus included drug discovery applications, assay methodologies, clinical pharmacology, experimental medicine, and biomarker discovery and evaluations in both pre-clinical and clinical applications.  In all of these topics, he has given numerous invited presentations at various international meetings.  As part of some committees within PhRMA, AAPS, AACR, SBS, etc., he has coauthored with relevant subject-matter experts from other companies, FDA, academia and NIH, in important position papers related to immunogenicity, high throughput screening, biomarker method validation and clinical biomarker qualification.

The quality of statistical thinking and methods used is a major determinant in the successful discovery and evaluation of biomarkers.  Data normalization and transformation methods can greatly impact the analysis.  False discovery rates (q-values) and miss rates are critical for setting meaningful thresholds for identification.  While it is important to analyze each marker individually (e.g., ANOVA), it is also important to keep in mind that a marker useless on its own may be great in a composite.  The use of multivariate modeling methods that account for the interactions, similarity and diversity of the markers is essential for the identification of composite biomarkers.  Examples of such methods include generalized additive models, shrunken centroids, random forests, kNN clustering, etc.  The predictive utility of these composite biomarkers should be assessed carefully via appropriate internal cross-validation methods, and further tested in independent cohorts (external validation) to expand and qualify their use in early drug development.  In addition, the statistical evaluation of validation data from biomarker analytical platforms can have a major impact on the utility of such platforms.  This short presentation will provide some background on biomarker research and an understanding of all these statistical considerations with illustrations and graphs.