A complete search of combinatorial peptide library greatly benefited from probabilistic incorporation of prior knowledge

The core of peptide detection in tandem mass spectrometry lies in associating fragment spectra with promising peptide candidates. We examined such detection in a synthetic combinatorial peptide library using four scoring metrics, against all theoretical peptides, and with a varying level of probabilistic prior knowledge—analyzing more than a trillion peptide-spectrum matches in total. Even after adjusting for peptide-length scoring bias, most MS/MS spectra had multiple at-least-as-good candidates as the correct peptide, showing that the highest spectral match was not a guarantee of correctness. As a remedy, we probabilistically integrated prior knowledge about expected cleavage behavior and expected peptide sequences into peptide scoring, reaching and even overcoming the performance of state-of-the-art de novo sequencing algorithms. Overall, we found that even partial and weak beliefs considerably improved peptide detection and are, in principle, generally applicable to any detection approach. Detection of peptides in a complete search thus often resulted in multiple admissible candidates near the maximal score, and the use of probabilistic prior knowledge substantially improved their discrimination. [Display omitted] •Peptide-spectrum metrics often assigned a near-maximal score to the correct peptides.•Scoring metrics tend to be biased towards longer peptides.•Cleavage-derived prior model substantially improved performance of complete search.•The use of prior probabilities leveraged problems with large search spaces.•A scoring metric utilizing a priori fragment distribution improved peptide detection.