Fig. 2
From: Improving MetFrag with statistical learning of fragment annotations
The training phase. The training consists of two major phases. For each phase a subset of the known reference MS/MS spectra is used. In the first phase MetFrag generates a list of assignments of m/z fragment peaks to fragment-structures for the given MS/MS spectra and their correct candidates. These assignments are generated by the in silico fragmentation of the correct candidate and the mapping of the generated fragment-structures to the m/z fragment peaks in the training spectrum. This assignments list (\(\mathcal {D}_{train}\)) is used in the second training phase along with the second subset of the reference spectra. Here, for each MS/MS spectrum the correct candidate is ranked with a candidate list using the consensus candidate score integrating besides the fragmenter (\(S^{c}_{MetFrag}\)) the two new statistical scoring terms (\(S^{c}_{Peak}, S^{c}_{Loss}\)). The number of correct Top1 rankings is used to optimize pseudo count and scoring weight parameters. The first training phase is used in analogy for the generation of the list containing assignments of m/z fragment losses to fragment-structures (\(\mathcal {D}^{L}_{train}\))