Computational approaches for serum marker discovery

Parag Mallick
Institute for Systems Biology

Whole cell, genome-wide analysis of gene expression, protein expression, protein state (e.g. glycosylation, phosphorylation) and related differential analyses in differentially perturbed cells have been widely applied to study biological processes and disease states. Such analysis has also been attempted on blood serum to detect and identify “fingerprints” or prognostic and diagnostic markers; blood serum is highly accessible and contains enormous information about physiologic state. A technology that can perform early detection of cancer can have a significant impact on cancer mortality when treated with existing cancer therapies. A key problem with the proteomic analysis of serum and many other body fluids is the highly skew composition of blood serum, which is dominated by a few highly abundant proteins; albumin alone represents over 50% of total serum protein content. Noting that many clinical biomarkers and therapeutic proteins, such as Her2/neu, human chorionic gonadotropin, alpha-fetoprotein, PSA and CA125 are glycosylated, a technique for the enrichment of serum glyco-proteins was developed.  In addition to experimental challenges, numerous computational challenges exist to quantitative proteomics of enriched LC-MS samples. We can describe the problem of identifying fingerprints to determine if a serum sample came from a healthy patient or from a cancer patient as a pattern analysis or discrimination problem. Having framed the problem in a pattern framework, we have begun developing and testing different algorithms and software tools to treat the three common subproblems of transduction, feature extraction and classification.