DEPT. OF EPIDEMIOLOGY
& BIOSTATISTICS

Computational/Statistical biology with emphasis on data mining of high-throughput data or "omics" data. Applications in genomic and proteomics data. Statistical methods in high dimensional settings (p >> n problem), model selection, bump hunting, statistical computing.

Conventional statistical techniques and methods literally fall apart or are inappropriate at best when dealing with modern large datasets where the number of variables greatly exceeds the number of observations (so-called p >> n problem). It is a hard problem with several statistical issues causing potential risks of severe errors and model unfitting. Particular challenges are the high dimensionality of the data, the massive parallel inference due to the multiplicity of the variables at play, the inherent noise due to the employed technologies, and finally the correlation between variables.

There is often more structure and heterogeneity in the data than we initially thought of or assumed, and it is of great interest and a challenge to unveil these structures in high dimensional and noisy settings as is always the case with omics data. Currently, I am finishing a work in collaboration with Dr. J.S. Rao on a search strategy of multiple modes in high dimensional data also known as bump hunting: . The algorithm has been applied with success on real data collected from a large ongoing colon cancer study at Case to classify patients and predict the outcomes from various colon cancer metastatic stages.

One can cast a variety of bioinformatics question into that of a model selection problem. Bayesian variable selection algorithm like shrunken Bayesian ANOVA and model selection algorithm like the "Fence" (recently developed by Dr. J.S. Rao and colleagues) have proven extremely efficient in reducing the complexity and dimensionality of the data with often dramatic improvement in model fitting and prediction error (as compared to existing methods). The so-called "Fence" method which was initially designed for model selection in mixed effects models provides a state-of-the art strategy for tackling these challenging problems. My interest is in developing and applying model selection algorithms and Fence variations, appropriate to specific settings in bioinformatics such as Differential Expression, Data Integration, Pathway Analysis/Gene Set Enrichment Analysis, Association Studies for simple and complex traits and Quantitative Trait Loci (QTL) mapping, and Protein interaction network.