|Date: Friday, November 06, 2015
Location: 1084 East Hall (4:10 PM to 5:00 PM)
Title: Sparse Proteomics Analysis: Toward a Mathematical Theory for Feature Selection from Forward Models
Abstract: Tumor diseases, such as cancer, rank among the most frequent causes of death in Western countries. The clinical research of the last decades has shown that the pathological mechanisms of many diseases are manifested on the level of protein activities. In order to improve the clinical treatment options and early diagnostics, it is therefore necessary to better understand protein structures and their interactions. The related research field of proteomics focuses on analyzing the so-called proteome, which denotes the entire set of proteins of a human individual at a certain point of time. Unfortunately, proteomics-data, e.g., produced by mass spectrometry, is usually extremely high-dimensional. Therefore, it is a very difficult task to extract a disease fingerprint, which is a small set of proteins allowing for an appropriate classification of a patient's health status.
In the first part of this talk, we will see that the assumption of sparsity can help us to cope with this challenge. In this context, the method of Sparse Proteomics Analysis (SPA) will be introduced, which is a combination of various generic algorithmic steps enabling us to build sparse and reliable classifiers. The second part of the talk is then devoted to a theoretical foundation of SPA. Relying on a simple linear forward model for the data, we will see that very recent results from high-dimensional estimation theory can be used to prove rigorous recovery guarantees.
Speaker: Martin Genzel
Institution: TU Berlin
Event Organizer: Jeremy Hoskins and Derek Wood