Context: Confidence Estimation for Targeted Proteomics

Overview

Context produces PSM- and peptide-level q-values and PEPs for targeted mass-spec assays. Because the reference panel is usually too small to train a discriminant directly, Context trains the discriminant on the background PSMs from the same run (a large, statistically well-behaved background) and transfers the learned linear discriminant to score the reference panel. Two engines are supported, each paired with its own confidence-estimation back-end:

• percolator (default) — Percolator trains the discriminant, and pyIsoPEP (q2pep) then computes FDR + q-values from TDC scores and fits an I-spline q -> PEP, separately at the PSM and peptide levels.
• mprophet — a Python re-implementation of the mProphet-based rescoring and statistical validation components built into EncyclopeDIA for PRM. q-values are Benjamini–Hochberg-adjusted p-values from a Gaussian decoy null, and PEPs are Storey's local FDR (qvalue::lfdr, KDE on logit-p-values, π₀ floored at 0.05), exactly as EncyclopeDIA reports them. pyIsoPEP is not used on this pathway.

Context is distributed as both a container image (Podman-, Docker- and Apptainer-compatible) and a Python package exposing the same CLI.

Installation

PyPI

pip install context-ms

Context is a Python package. The mprophet engine has no external dependency beyond NumPy/pandas. The percolator engine prefers a local percolator executable on PATH and falls back to a container image via Podman or Docker (set with --container-cmd, default podman):

• Percolator container: ghcr.io/percolator/percolator:master

Container image

podman pull ghcr.io/shannon225/context:main
# or
apptainer build context.sif docker://ghcr.io/shannon225/context:main

Input & output

Input: two Percolator-compatible tsv files for background and reference PSMs/peptides. They must share an identical header (same columns, same order).

Output: for context run --background BG --reference REF --prefix P --outdir results/ --engine E:

results/
  weights/P.weights.txt           # trained weights
  P.rescored_features.tsv         # reference features rescored with the trained weights
  P.psm.reference.txt             # PSM-level reference targets
  P.peptide.reference.txt         # peptide-level reference targets

Weights file format:

• percolator: Percolator's native --weights output (3 lines per CV bin).
• mprophet: a two-column TSV with feature and weight, plus a final __bias__ row carrying the LDA constant.

Each output path can be overridden individually via --weights-out, --rescored-out, --psm-out, --peptide-out. Values may be plain file names (written inside --outdir, or inside --outdir/weights for the weights file) or absolute paths.

Command-line reference

context run -h

Input profiles (mprophet only)

• auto — if any column starts with var_ or main_var_, keep only those (OpenSWATH/pyprophet convention); otherwise fall back to encyclopedia.
• pin — keep every column the pin convention exposes as a feature.
• encyclopedia — drop the metadata columns EncyclopeDIA's MProphetFeatureReader excludes by name (pepLength, charge1..4, precursorMass, RTinMin, midTime, numberOfMatchingPeaksAboveThreshold, primary, TD).

--input-profile has no effect on the percolator engine.

Seed coefficients (mprophet only)

A JSON dictionary mapping feature column names to starting linear coefficients for the seed LDA. Names that don't appear in the input contribute 0; if all entries drop out, the seed model is disabled and inner iter 0 falls back to ranking by the single best feature. Pass none to disable the seed model unconditionally.

Confidence estimation

The two engines use different back-ends for q-values and PEPs:

• percolator — Percolator trains the discriminant, then pyIsoPEP computes q-values (from TDC counts) and PEPs (from an I-spline fit to q(score)). Called separately for the PSM-level and the peptide-level tables.

• mprophet — EncyclopeDIA's method throughout, at both the PSM level and (after per-sequence deduplication, highest score wins) the peptide level:

  • fit a Gaussian N(µ_d, σ_d) to the decoy scores;
  • p = 1 - Φ((score - µ_d) / σ_d) for both targets and decoys;
  • q = Benjamini–Hochberg-adjusted p-values;
  • PEP = Storey qvalue::lfdr (KDE on the logit-transformed p-values, Silverman bandwidth, monotone non-decreasing in p), with π₀ floored at 0.05.

  The same routine is used inside the mProphet training loop for the "passing targets" selection at each inner iteration and for the held-out evaluation that picks between the trained LDA and the seed model.

Examples

cd example

PyPI

# Percolator (default)
context run \
  --background background.tsv \
  --reference    reference.tsv \
  --prefix    run01 \
  --outdir    results_run01

# mProphet
context run \
  --background background.tsv \
  --reference    reference.tsv \
  --prefix    run01 \
  --outdir    results_run01 \
  --engine    mprophet

Container image

# Podman
podman run --rm -v "$PWD:/work" -w /work \
  ghcr.io/shannon225/context:main \
  run --background background.tsv --reference reference.tsv \
      --prefix run01 --outdir results_run01 --engine mprophet

# Apptainer
apptainer run --bind "$PWD:/work" --pwd /work context.sif \
  run --background background.tsv --reference reference.tsv \
      --prefix run01 --outdir results_run01

Links

• PyPI package: https://pypi.org/project/context-ms
• Container image: https://github.com/shannon225/Context/pkgs/container/context
• GitHub repository: https://github.com/shannon225/Context
• Percolator: http://percolator.ms
• EncyclopeDIA: https://bitbucket.org/searleb/encyclopedia
• pyIsoPEP: https://github.com/statisticalbiotechnology/smooth_q_to_pep