Supplementary Figure 2: Hierarchical Bayes model description.
From: TagGraph reveals vast protein modification landscapes from large tandem mass spectrometry datasets
a) Overview of hierarchical Bayes model with Expectation Maximization fitting. First, attributes (b, below) are assigned to each peptide-spectrum match from the entire dataset. Initial, naïve likelihood parameters are applied noting the likelihood that peptide-spectrum matches are correct or incorrect given each attribute (P(A|+), P(A|-)), respectively. Second, all attributes are combined to estimate the relative likelihood that each peptide-spectrum match is correct given all available data (P(+|D0)). Third, the parameters used to estimate all likelihood models are refined based on the correct (P(D|+)) and incorrect (P(D|-)) distributions learned from the previous iteration. b) Bayes model used for fitting correct (+) and incorrect (-) peptide-spectrum match distributions. Gray arrows indicate dependencies between model attributes and the distribution being trained. Blue arrows indicate dependencies between model attributes. Attributes in magenta ovals specifically pertain to sequence modifications. Further details are provided in Supplementary Note 4. c) Example distributions for several model attributes derived from the A375 dataset (Fig. 1). Likelihood distributions were iteratively refined across multiple measurement dimensions using expectation-maximization (EM).
