Protein probabilities in shotgun proteomics: evaluating different estimation methods using a semi-random sampling model

The calculation of protein probabilities is one of the most intractable problems in large-scale proteomic research. Current available estimating methods, for example, ProteinProphet, PROT_PROBE, Poisson model and two-peptide hits, employ different models trying to resolve this problem. Until now, no efficient method is used for comparative evaluation of the above methods in large-scale datasets. In order to evaluate these various methods, we developed a semi-random sampling model to simulate large-scale proteomic data. In this model, the identified peptides were sampled from the designed proteins and their cross-correlation scores were simulated according to the results from reverse database searching. The simulated result of 18 control proteins was consistent with the experimental one, demonstrating the efficiency of our model. According to the simulated results of human liver sample, ProteinProphet returned slightly higher probabilities and lower specificity than real cases. PROT_PROBE was a more efficient method with higher specificity. Predicted results from a Poisson model roughly coincide with real datasets, and the method of two-peptide hits seems solid but imprecise. However, the probabilities of identified proteins are strongly correlated with several experimental factors including spectra number, database size and protein abundance distribution.