Peptidoform Scoring And Filtering

peptidoform-scoring-and-filtering

Summary

Post-hoc validation and ranking of de novo peptide sequence predictions by applying mass tolerance constraints, minimum length thresholds, and per-residue confidence scoring. This skill ensures predicted peptidoforms meet chemical plausibility and quality criteria before downstream interpretation.

When to use

After a transformer-based de novo sequencing model (such as Casanovo) generates candidate peptide sequences from MS/MS spectra, before exporting results or using them in database matching or visualization workflows. Essential when working with data-dependent acquisition bottom-up tandem mass spectrometry datasets where sequence predictions require post-hoc validation against precursor mass and confidence thresholds.

When NOT to use

• Input peptides are already identified via database matching and do not need de novo filtering.
• Precursor m/z or charge state information is missing from the dataset.
• Working with closed-search or peptide library matching workflows where mass tolerance is enforced upstream.

Inputs

• MS/MS spectra (MGF or mzML/mzXML format)
• de novo peptide predictions with per-residue logits and precursor m/z
• configuration parameters: min_peptide_len, precursor_mass_tol (ppm or Da)

Outputs

• filtered peptidoform predictions (mzTab format)
• ranked candidate sequences with amino acid scores and peptide scores
• sequence and modification annotations in ProForma 2.0 notation

How to apply

Apply three complementary filters in sequence: (1) reject predictions with sequence length below min_peptide_len threshold to eliminate implausibly short identifications; (2) penalize or exclude predictions whose calculated precursor mass deviates beyond precursor_mass_tol (specified in ppm or Daltons) from the experimental precursor m/z, correcting for charge state; (3) compute amino acid scores as the product of per-residue confidence logits from the model's decoder output, then aggregate into a peptide-level score for ranking. Retain only predictions passing all filters, then export with sequence, modifications (ProForma 2.0 notation), amino acid scores, peptide score, precursor m/z, charge state, and scan metadata to mzTab format for downstream analysis.

Related tools

• Casanovo (transformer-based de novo sequencing model that generates candidate peptide sequences and per-residue confidence logits as input to scoring and filtering) — https://github.com/Noble-Lab/casanovo
• DepthCharge (spectrum encoder used upstream in Casanovo pipeline to generate spectral representations fed to the transformer)
• PyTorch (deep learning framework underlying Casanovo model inference and logit computation)
• limelight-import-casanovo (post-filtering converter that transforms filtered mzTab outputs to Limelight XML for visualization and quality assessment) — https://github.com/yeastrc/limelight-import-casanovo
• PDV (graphical viewer for inspecting and validating filtered Casanovo predictions and annotated spectra) — https://github.com/wenbostar/PDV

Evaluation signals

• All retained peptides meet min_peptide_len threshold (no sequences shorter than configured cutoff).
• Precursor mass deviation for all retained peptides is within ±precursor_mass_tol across all charge states (verifiable by recalculating theoretical mass from sequence).
• Per-residue logit products are non-negative and bounded; peptide-level scores monotonically reflect confidence of highest-confidence residues.
• mzTab export contains all required fields: sequence, ProForma 2.0 modifications, amino acid scores, peptide score, precursor m/z, charge state, scan metadata.
• Filtered prediction count is ≤ unfiltered count; no predictions incorrectly retained or rejected due to parameter misconfiguration.

Limitations

• Precursor mass tolerance filtering assumes accurate precursor m/z assignment; errors in charge state determination upstream will propagate as false negatives.
• Per-residue logits reflect model confidence, not ground-truth correctness; high logit scores do not guarantee sequence accuracy, especially for post-translational modifications or non-standard amino acids.
• ProForma 2.0 notation support depends on upstream model training; modifications not explicitly modeled during training may be missed or mislocalized.
• Filtering thresholds (min_peptide_len, precursor_mass_tol) are dataset- and instrument-specific; values must be tuned for each LC-MS platform and sample type.

Evidence

• [other] Apply post-hoc filters: reject predictions failing the minimum peptide length (min_peptide_len), penalize predictions outside precursor mass tolerance (precursor_mass_tol in ppm or Da), and assign amino acid scores as the product of per-residue confidence logits.: "Apply post-hoc filters: reject predictions failing the minimum peptide length (min_peptide_len), penalize predictions outside precursor mass tolerance (precursor_mass_tol in ppm or Da), and assign"
• [other] Export predictions to mzTab format, including sequence, modifications in ProForma 2.0 notation, amino acid scores, peptide score, precursor m/z, charge state, and scan metadata.: "Export predictions to mzTab format, including sequence, modifications in ProForma 2.0 notation, amino acid scores, peptide score, precursor m/z, charge state, and scan metadata."
• [readme] Casanovo can be used to find unexpected peptide sequences in any data-dependent acquisition, bottom-up tandem mass spectrometry dataset, and is particularly useful for immunopeptidomics, metaproteomics, paleoproteomics, venomics: "find unexpected peptide sequences in any data-dependent acquisition, bottom-up tandem mass spectrometry dataset"
• [methods] A converter for generating Limelight XML from Casanovo results is available, see the converter documentation: "A converter for generating Limelight XML from Casanovo results is available"