Site claims and modification identity

The problem

"This peptide is phosphorylated" and "this serine is phosphorylated" are different claims, and "+42.047" can be acetylation or trimethylation. A site list needs an error estimate for the site, and an honest answer to which modification.

Knockoff sites

For annotated modification sites (UniProt, PEFF \ModRes), andes adds knockoff sites in the same spirit as knockoff variants: other residues of the same protein with the same context. A PSM that puts the modification on an annotated site supports that claim; one that puts it on a knockoff site is a false claim. Within each modification × residue × prior tier, the knockoff PSMs estimate how many annotated-site claims are false, giving a site q-value.

Occupancy and identity

Every accepted site gets two probabilities, computed over all rows of a spectrum so the explanations compete with each other:

flowchart LR
    S["One spectrum<br/>all candidate rows<br/>targets and decoys"] --> L
    L["Per-row logit<br/>rescored score<br/>precursor isotope offset<br/>fragment ions unique to this form<br/>diagnostic ions and neutral losses"] --> P["Posterior over<br/>the spectrum's rows"]
    P --> O["occupancy<br/>share of the posterior that<br/>modifies this residue"]
    P --> I["identity<br/>given a modification there,<br/>the share of this one"]

identity_alternatives lists the competing masses, for example +42.0470:0.29 beside acetyl at 0.71. How much a diagnostic ion or neutral loss says is learned from the run rather than fixed.

The output is <stem>.sites.tsv, one row per accepted site, with columns including site_q, confidence, occupancy, identity and identity_alternatives.

Where it works

Site claims are experimental and most useful with freely placed modifications. With phospho restricted to annotated sites, many knockoff residues are really phosphorylated, so only 119 sites passed in the benchmark; the limits are in the full reference.