Inference of geometry and dependence in moeling complex disease phenotypes
Dr. Sayan Mukherjee
November 5, 2008
Lally 102, 4:00 p.m. to 5:00 p.m.
Refreshments at 3:30 p.m.
Tumor progression is a complex (disease) phenotype. The challenge in modeling tumorigenesis is heterogeneity with respect to phenotype, stages of the disease, and genotype or gene expression variation. This is particularly challenging in the case of high-dimensional data.
I will discuss machine learning tools or Bayesian statistical models
that allow for the inference of which genes or sets of genes
vary across stages of tumors and which are specific to certian
stages. I will also discuss the inference of gene and pathway networks.
The modeling tools are (Bayesian) hierarchical or multi-task regression models as well as a generalization inverse regression based on inference of gradients on manifolds. Statistical properties of this
approach will be developed in the context of simultaneous dimension
reduction and regression as well as graphical models. Consistency of dimension reduction as well as inference of the graphical model will be stated. An interesting observation is that the rate of convergence of both estimates depends on the dimension of the underlying manifold
on which the marginal distribution is assumed to be concentrated, the result does not depend on the number of non-zero entries of the conditional independence matrix but on the rank of this matrix.
(Joint work with: Qiang Wu, Elena Edelman, and Justin Guinney.)
Hosted by: Dr. Sanmay Das (x2782)
Last updated: October 17, 2008