Learning to Rank Peptide-Spectrum Matches Using Genetic Programming

The analysis of tandem mass spectrometry (MS/MS) proteomics data relies on automated methods that assign peptides to observed MS/MS spectra. Typically these methods return a list of candidate peptide-spectrum matches (PSMs), ranked according to a scoring function. Normally the highest-scoring candidate peptide is considered as the best match for each spectrum. However, these best matches do not necessary always indicate the true matches. Identifying a full-length correct peptide by peptide identification tools is crucial, and we do not want to assign a spectrum to the peptide which is not expressed in the given biological sample. Therefore in this paper, we present a new approach to improving the previous ordering/ranking of the PSMs, aiming at bringing the correct PSM for spectrum ahead of all the incorrect ones for the same spectrum. We develop a new method called GP-PSM-rank, which employs genetic programming (GP) to learn a ranking function by combining different feature functions that measure the quality of PSMs from different perspectives. We compare GP-PSM-rank with SVM-rank. The results show that GP-PSM-rank outperforms SVM-rank in terms of the number of identified peptides which are true matches. On a validation dataset with 120 spectra, the proposed method is used as the post processing step on the results of peptide identifications by two de novo sequencing algorithms. GP-PSM-rank improves the results of both de novo methods in terms of identifying the true matches.