Differences from ProteoBench

apb-proteobench reproduces ProteoBench's quantification scores from APB2 results. The calculation is the same; the inputs can differ in two places: which protein column decides feature exclusion, and which abundance layer is scored. Both are listed here with their measured effect.

Feature exclusion uses FASTA matches

Scoring drops decoys, contaminant features and features of more than one species (min_count_multispec). APB decides species and FASTA contaminants from apb-fasta's varm["fasta_validation"]: every FASTA protein containing the feature's peptide gives fasta_matching_organisms and fasta_matches_contaminant. Decoys, and contaminants the software flags or adds itself such as MaxQuant's CON__, come from apb2's apb_Decoy and apb_Contaminant. ProteoBench decides them from each tool's column mapped to Proteins.

For DIA-NN, whose Protein.Ids lists every protein containing the peptide, both rules agree. Tools whose Proteins column is a protein group or a tool-specific list lose features that APB now recognises as shared with a contaminant or a second species.

Measured against the stored ProteoBench datapoints of 219 corpus submissions, nr_feature at cutoff 1:

• DIA-NN and FragPipe (DIA-NN quant): median difference 0.1–0.3 %; before this rule, 0.5–3.9 %
• Plasma DIA-NN: 2.9 % before, 0.34 % after
• MaxQuant, AlphaDIA, Spectronaut, PEAKS, Sage, AlphaPept, FragPipe (DDA), quantms: previously within 0.7 % and mostly exact, now up to 3.3 %
• MaxQuant: its 0.2 % of features without a FASTA match are its reversed decoys (REV__), which ProteoBench excludes as decoys too
• i2MassChroQ: 21–24 %, because APB2's i2MassChroQ rule keeps modifications in ProForma_peptide, so modified peptides match no protein

The scored layer is chosen by the caller

run and benchmark score one abundance layer: --layer NAME, by default X, the APB2 rule's primary layer. ProteoBench fixes the quantity column per module and tool, and its plasma module departs from its other DIA modules:

To reproduce ProteoBench's plasma datapoints from DIA-NN results, pass --layer Precursor_Quantity. With X, DIA-NN 2.x plasma results differ by more than 50 %, because DIA-NN's normalisation factors differ between the two conditions there.

Entrapment pairs come from the FASTA

ProteoBench labels each precursor and finds its partner in a separate pair file. APB derives both from the entrapment FASTA, which it already reads to verify peptides:

• Label: protein_fasta's is_entrapment flags the sp|<peptide>_p_target|… entries
• Unmodified pair: the FASTA lists every target directly before its entrapment, so the pair is the entry position halved
• Modified forms: the pair file pairs each one with the partner form carrying the same modifications on the same occurrence of each residue, e.g. the second Met oxidised; APB's pair key is the unmodified pair plus those sites
• Measured on the 2026 pair file: this key reproduces all 2,573,718 pairs, with no two forms sharing a key
• Guard: scoring stops when two adjacent FASTA entries are not a target followed by a same-length entrapment
• Difference: the pair file lists only the modified forms ProteoBench generated; APB pairs any Oxidation and Carbamidomethyl combination, which the pair-file version refused. None of the 12 entrapment submissions has such a form: all scored under that version.

Smaller differences

• Two-species (HY) modules write no nr_quantified_ECOLI; ProteoBench's plasma code always writes it, as 0
• Per-species keys exist only for the module's species; key names and formulas otherwise follow ProteoBench 0.18.0