PNNL

Abstract

Current proteomics techniques, such as mass spectrometry, focus on protein identification, usually ignoring most types of modifications beyond post-translational modifications, with the assumption that only a small number of peptides have to be matched to a protein for a positive identification. However, not all proteins are being identified with current techniques and improved methods to locate points of mutation are becoming a necessity. In the case when single-nucleotide polymorphisms (SNPs) are observed, brute force is the most common method to locate them, quickly becoming computationally unattractive as the size of the database associated with the model organism grows. We have developed a Bayesian model for SNPs, BSNP, incorporating evolutionary information at both the nucleotide and amino acid levels. Formulating SNPs as a Bayesian inference problem allows probabilities of interest to be easily obtained, for example the probability of a specific SNP or specific type of mutation over a gene or entire genome. Three SNP databases were observed in the evaluation of the BSNP model; the first SNP database is a disease specific gene in human, hemoglobin, the second is also a disease specific gene in human, p53, and the third is a more general SNP database for multiple genes in mouse. We validate that the BSNP model assigns higher posterior probabilities to the SNPs defined in all three separate databases than can be attributed to chance under specific evolutionary information, for example the amino acid model described by Majewski and Ott in conjunction with either the four-parameter nucleotide model by Bulmer or seven-parameter nucleotide model by Majewski and Ott.