Bio Proteomics Protein Inference

Version Compatibility

Reference examples tested with: pyOpenMS 3.1+, pandas 2.2+

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures
• R: packageVersion('<pkg>') then ?function_name to verify parameters

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

The pyOpenMS protein-inference class names have varied across releases. Confirm the exact spelling at the installed version with help(pyopenms.EpifanyAlgorithm) and help(pyopenms.BasicProteinInferenceAlgorithm) before relying on the reference code.

Protein Inference -- A Chosen Explanation of Peptide Evidence, Reported as Groups

"Tell me which proteins are present from my identified peptides" -> Assign the observed peptides to a minimal or probability-weighted set of proteins, reported as groups of indistinguishable proteins with a leading accession -- because bottom-up MS measures peptides, and the protein set behind them is inferred, not observed.

• Python: pyopenms.BasicProteinInferenceAlgorithm().run(peptide_ids, protein_ids) for parsimony grouping
• Python: pyopenms.EpifanyAlgorithm (TOPP tool Epifany) for Bayesian belief-propagation inference
• CLI: ProteinProphet (TPP) for EM-based probabilistic inference; Philosopher filter for FragPipe FDR

Scope: this skill OWNS peptide-to-protein grouping, the indistinguishable/subsumable distinction, the leading-protein convention, inference-method choice, and protein/protein-group FDR. PSM-level and peptide-level FDR plus the search engines that produce the peptide list -> peptide-identification. The quantitative fallout of razor vs unique peptides on protein abundance -> quantification. OUT OF SCOPE: resolving splice isoforms, single-AA variants, or PTM-defined proteoforms (bottom-up groups cannot separate them; that is top-down / proteoform work).

The Single Most Important Modern Insight -- Protein Inference Is Underdetermined, So the Honest Unit Is a Group, Not a List

1. The protein set is not uniquely recoverable from peptides, so a protein group -- not a flat protein list -- is the only honest reporting unit. Many peptides are shared across paralogs, gene families, and isoforms, so distinct protein sets can explain the same peptide evidence equally well. The inference picks ONE explanation under an assumption (parsimony, or a probability model); proteins that the observed peptides cannot tell apart (indistinguishable) MUST be reported as one group with a designated leading protein. A flat list double-counts indistinguishable proteins and breaks target/decoy symmetry at the protein level, silently corrupting FDR.

2. Protein FDR is its own estimation problem that INFLATES on large data; the fix is PICKED FDR, not the PSM formula reused. Controlling PSM-FDR at 1% does not give 1% protein-FDR. A deep run has many false PSMs in absolute terms, and each can nucleate a one-hit-wonder false protein; because true proteins accumulate many peptides while false proteins are hit once, the naive protein-FDR balloons to 10-30% on deep datasets. Savitski 2015 picked-protein FDR pairs each target protein with its decoy and keeps only the higher-scoring of the pair before counting, removing the target/decoy asymmetry; The & Kall 2016 extends this to the group level (picked-group FDR), which is required because parsimony grouping is anticonservative otherwise.

3. The two-peptide rule is wrong -- it increases protein FDR and discards real proteins. Requiring >=2 peptides per protein (Gupta & Pevzner 2009, "A strike against the two-peptide rule") removes MORE target proteins than decoy proteins, so it raises protein-level FDR rather than lowering it, while throwing away legitimate low-abundance single-peptide IDs. Replace the blanket rule with: control protein-level (picked) FDR, then judge single-peptide IDs by their score, not their peptide count.

Vocabulary the Rest of This Depends On

• Shared (degenerate) peptide: maps to >1 protein in the searched database. Cannot, alone, distinguish which protein is present.
• Unique peptide: maps to exactly one protein -- the only direct evidence for a specific protein. "Unique" is DATABASE-RELATIVE: a peptide unique against SwissProt may be shared against TrEMBL+isoforms+contaminants. Always document the exact database (isoforms, contaminants, decoys included).
• Indistinguishable proteins: explained by the SAME set of observed peptides -> one group, never two confident IDs.
• Subset / subsumable protein: its observed peptides are a subset of another protein's -> parsimony drops it (the larger protein explains everything it would).
• Leading / representative protein: the group's reported accession. Convention: most peptides, then highest score, then SwissProt canonical over TrEMBL. Downstream tables key on this accession but must retain group membership -- "protein P12345" usually means "the group led by P12345".
• Protein group vs proteoform: a group is an inference artifact (proteins lumped because peptides cannot separate them); a proteoform is a real molecular species (one gene product with a specific sequence + PTM + cleavage state). Bottom-up groups DO NOT resolve proteoforms -- claiming "isoform X present" from a shared-peptide group is overreach.

Tool Taxonomy

Decision Tree by Scenario

Default when uncertain: run parsimony grouping (BasicProteinInferenceAlgorithm with annotate_indistinguishable_groups), report protein GROUPS with a leading accession, and control protein-GROUP FDR with picked-group FDR at 1%. Do NOT impose a two-peptide rule.

Group Proteins by Parsimony with pyOpenMS

Goal: Turn an FDR-filtered peptide identification list into protein groups with a leading protein, resolving shared-peptide ambiguity.

Approach: Load the idXML from peptide identification, run the parsimony algorithm with indistinguishable-group annotation on, then read the inferred groups off the protein identification run.

from pyopenms import IdXMLFile, BasicProteinInferenceAlgorithm

protein_ids = []
peptide_ids = []
# protein_ids is FIRST in both load() and store() for IdXMLFile
IdXMLFile().load('peptides_1pct_fdr.idXML', protein_ids, peptide_ids)

inference = BasicProteinInferenceAlgorithm()
params = inference.getParameters()
# annotate_indistinguishable_groups reports indistinguishable proteins as ONE group
params.setValue('annotate_indistinguishable_groups', 'true')
inference.setParameters(params)
inference.run(peptide_ids, protein_ids)

# indistinguishable groups live on the protein identification run
for prot_id in protein_ids:
    for group in prot_id.getIndistinguishableProteins():
        leading = group.accessions[0]  # convention: highest-evidence accession first
        print(leading, group.probability, list(group.accessions))

Bayesian Inference + Group FDR with EPIFANY

Goal: Assign calibrated protein/group posteriors and control protein-group FDR with a probability model rather than greedy parsimony.

Approach: EPIFANY consumes idXML whose PSMs already carry posterior error probabilities (from Percolator or IDPosteriorErrorProbability), then propagates belief over the peptide-protein graph. The TOPP tool is reliably named Epifany; the pyOpenMS class spelling has varied across releases, so introspect first.

import pyopenms
from pyopenms import IdXMLFile

# CONFIRM the class name at the installed version before use:
#   help(pyopenms.EpifanyAlgorithm)
algo_cls = getattr(pyopenms, 'EpifanyAlgorithm')

protein_ids = []
peptide_ids = []
IdXMLFile().load('peptides_with_pep.idXML', protein_ids, peptide_ids)

algo = algo_cls()
# EPIFANY expects PSM posteriors as input; greedy_group_resolution controls
# whether shared peptides are razor-resolved after inference
algo.inferPosteriorProbabilities(protein_ids, peptide_ids, False)

for prot_id in protein_ids:
    for group in prot_id.getIndistinguishableProteins():
        print(group.accessions[0], group.probability)