| Protein biomarkers are derived from mass spectrometry (MS)-based proteomic data, integrative genomic studies using microarrays, and literature mining. This often results in many tens to hundreds of candidates whose presence and concentration must be verified in blood, the biofluid of choice, prior to developing a clinical assay. Historically, verification was performed using antibodies; however, the required immunoassay-grade reagents necessary for sensitive and specific detection in blood only exist for a tiny fraction of the proteome. Thus, alternative technologies are required to bridge the gap between discovery and clinical assay development.;Targeted hypothesis-driven mass spectrometry is increasingly being used to verify changes in protein response due to disease or therapy. These methods require selecting "signature" peptides as detectable and quantifiable surrogates for each protein. Peptides that produce the highest ion-current response are ideal choices as they are more sensitively detected. Unfortunately, selection of these peptides, particularly in the absence of experimental data, is a challenging problem and represents a significant resource constraint in targeted assay development by mass spectrometry.;This thesis describes the development of the Enhanced Signature Peptide (ESP) predictor, a computational method to predict high-responding peptides from a given protein. I used an MS-based analysis of yeast to define a training set and took advantage of the theoretical properties of the Random Forest (RF) algorithm, an ensemble of individual classification trees, to model peptide response. On average, the ESP predictor achieves a success rate of 89% at selecting one or more high-responding peptides per protein. More importantly, we applied the ESP predictor to select de novo signature peptides and show that it correctly selected 74% of experimentally validated signature peptides (2 peptides per protein on average). Furthermore, the ESP predictor is a robust model that, unlike existing computational methods, performs well across all common electrospray ionization (ESI)-MS experimental types. The ESP predictor fills a critical gap, enabling selection of optimal candidate signature peptides to detect and quantify any protein of interest in the absence of MS experimental data. We anticipate the use of the ESP predictor will greatly improve the efficiency of biomarker verification. |
