UC San Diego

Tandem mass spectrometry (MS2) is widely used for peptide and protein identification. One of the most fundamental problems for peptide identification in MS2 is to score peptide annotations against the spectrum which is produced by the peptide. In this thesis, a Bayesian network model is proposed for scoring peptides from Q-TOF mass spectrometers. The research is based on the Bayesian network probabilistic methodology used by InsPecT software, which exploits a hybrid strategy of both database search and de novo algorithms for peptide identification. Initially we focused on the connections of InsPecT scoring model without any changes of nodes. We attempted to determine the connections between nodes for Q-TOF by their dependencies. In order to prove that we need the complete set of nodes as the original InsPecT scoring model, we reduced the number of nodes and surprisingly caused significant improvement in peptide identification performance. The 18-node model was reduced to 10-node models for both charge 2 and charge 3 ions, and we obtained the percentage gain in spectra identification 37.51% for charge 2 and 57.68% for charge 3 ions compared to the InsPecT software 2006.10.20 version. The simplified model also leads to computation time reduction. Currently InsPecT does not perform as well as Mascot on Q-TOF data. Reason for that may be that InsPecT was originally trained for LTQ data and in this thesis we only focused our improvement on the InsPecT scoring stage. Deficiencies may occur in the initial tagging and final calculation of the score. Further research may do an exhaustive combination of fragment ions to derive a set of most discriminative and informative ions.