Knockoff variant FDR

The problem

A PSM on a UniProt variant entry claims that the sample carries that variant. Most false claims are not random matches. They are an ordinary peptide that a common modification or an isotope error makes look like the variant:

Looks like a variant	Actually
N→D, Q→E	deamidation
G→A, D→E	methylation
a +1 Da substitution	a one-isotope mis-pick on the canonical peptide

Decoys model random matches only, so a decoy-based q-value never sees this error.

The idea

Give every variant claim a null that makes the same mistakes: a knockoff — the same substitution placed at another position of the same protein that holds the same residue, with matching context. A knockoff is almost certainly absent from the sample, but a modification or isotope error that mimics the real substitution mimics the knockoff just as well.

flowchart TD
    V["Real variant entry<br/>…EN<b>D</b>VK… (N→D at 7)"] --> K
    K["Up to 4 knockoff entries<br/>same N→D at other positions<br/>same residue class, same sequon status,<br/>similar peptide length, far from known variants"]
    K --> P["Search and Percolator<br/>treated exactly like real variants"]
    P --> S["Split by substitution class<br/>deamidation-like · methyl-like · oxidation-like · other"]
    S --> E["Knockoff PSMs above a score<br/>estimate false real PSMs above it"]
    E --> Q["Knockoff q-value<br/>reported q = max(global, class, knockoff)"]

A variant with no valid knockoff position has no null, so it is not accepted: measured against genotypes, such variants are false several times more often than tested ones.

What it changes

Checked against donor genotypes (PXD022787: four donors, one TMT10 set, 69 runs, ion-trap MS2). A variant PSM is false when no donor haplotype carries its variant.

Setting	Variant PSMs reported	True FDP
No knockoffs, class q ≤ 0.01	2,452	65%
No knockoffs, class q ≤ 0.05	3,886	71%
Knockoffs, --proteoform-fdr 0.05	201	2.5%
Knockoffs, --proteoform-fdr 0.10	307	6.0%

Only 0.55% of knockoffs turned out to be a real donor variant, so the null is clean. The price on everything else is small: PSMs at q ≤ 0.01 change by −0.43% on pooled Astral DDA and −0.30% on the TMT set, and search time is unchanged.

Prior tiers

Variant claims can come from several sources (sample, common, curated, catalogue). Splitting the estimate by tier stops a sample's own variants and catalogue variants from sharing one error estimate:

Limits

• About 30% of single-residue variants can be given a knockoff, so most variants are not reported.
• Multi-residue variants, indels and isoform-borne variants get no knockoffs.
• Calibrated on one genotyped dataset with low-resolution MS2; the deamidation- and oxidation-like classes rest on small counts.

Use it

andes --spectrum run1.mzML --spectrum run2.mzML --database uniprot_sprot_human.xml.gz \
  --uniprot-variants --uniprot-evidence curated --rescore --output-pin variants.pin

Parameters: --knockoffs, --proteoform-fdr, --uniprot-evidence. The full description is in Optional modes.