Abstract
Tandem mass-spectrometry has become the method of choice for high-throughput, quantitative analysis in proteomics. However, since the link between the peptides and the proteins they originate from is typically broken, identification of the analyzed peptides relies on matching of the fragmentation spectra (MS2) to theoretical spectra of possible candidate peptides, often filtered for precursor ion mass. To this end, peptide-spectrum matching algorithms score the concordance between the experimental and the theoretical spectra of candidate peptides, by evaluating the number (or proportion) of theoretically possible fragment ions observed in the experimental spectra, without any discrimination. However, the assumption that each theoretical fragment is just as likely to be observed is inaccurate. On the contrary, MS2 spectra often have few dominant fragments.
We propose a novel prediction algorithm based on a hidden Markov model, which allow for the training process to be carried out very efficiently. Using millions of MS/MS spectra generated in our lab, we found an overall good reproducibility across different fragmentation spectra, given the precursor peptide and charge state. This result implies that there is indeed a pattern to fragmentation that justifies using machine learning methods. Furthermore, the overall agreement between spectra of the same peptide at the same charge state serves as an upper limit on how well prediction algorithms can be expected to perform.
We have investigated the performance of a third order HMM model, trained on several million MS2 spectra, in various ways. Compared to a mock model, in which the fragment ions and their intensities are shuffled, we see a clear difference in prediction accuracy using our model. This result indicates that our model can pick up meaningful patterns, i.e. we can indeed learn the fragmentation process. Secondly, looking at the variability of the prediction performance by varying the train/test data split, in a K-fold cross validation scheme, we observed an overall robust model that performs well independent of the specific peptides that are present in the training data.
Last but not least, we propose that the real value of this model is as a pre-processing step in the peptide identification process, by discerning fragment ions that are unlikely to be intense for a given candidate peptide, rather than using the actual predicted intensities. As such, probabilistic measures of concordance between experimental and theoretical spectra, would leverage better statistics.
