ristretto

Lean, fast PSM rescoring for proteomics, purpose-built for use in MS²Rescore.

About

ristretto is a small, dependency-light reimplementation of the Percolator semi-supervised identification rescoring algorithm. Like Mokapot it leverages the scikit-learn library for machine learning operations. However, as ristretto was written for leanness, it is not as flexible as Mokapot.

Given a table of peptide-spectrum matches (PSMs) with rescoring features, a target/decoy label, and a spectrum identifier, it trains a classifier to separate correct from incorrect matches, then reports scores, q-values, and posterior error probabilities (PEPs) at the PSM, peptide, and protein level.

Design goals:

• Lean. Three runtime dependencies: numpy, scikit-learn, pandas.
• Fast. Hyperparameters are tuned once rather than every iteration, and hot paths are vectorized.
• Simple API. One function in, one result object out. A pandas DataFrame is the only input; column roles are given by name.

Two models are available:

• svm (default): a linear SVM trained with the semi-supervised iterative procedure of Käll et al. 2007, matching the Percolator/mokapot default.
• lda: single-pass Fisher linear discriminant analysis, in the style of Sage. Faster, no iteration.

Installation

pip install ristretto

Quickstart

import pandas as pd
from ristretto import rescore

# One row per PSM: feature columns + a decoy flag + a spectrum identifier.
features = pd.read_parquet("rescoring_features.parquet")

result = rescore(
    features,
    is_decoy_col="is_decoy",
    spectrum_id_col="spectrum_id",
    peptidoform_col="peptidoform",  # required
    peptide_col="sequence",         # optional: stripped seq, peptide-level rollup
    protein_col="proteins",         # optional: protein-level rollup
    decoy_pattern="rev_",           # optional: enables picked-protein competition
    model="svm",
    n_jobs=8,
)

# PSM table: identifier columns + score, qvalue, pep
psms = result.psms
confident = psms[(~psms["is_decoy"]) & (psms["qvalue"] <= 0.01)]
print(f"{len(confident)} PSMs at 1% FDR")

# Higher aggregation levels (columns: <key>, score, qvalue, pep, is_decoy, n_psms)
result.peptidoforms   # always
result.peptides       # if peptide_col
result.proteins       # if protein_col

# Rejoin features to scored PSMs via the preserved index
features.loc[result.psms.index]

# Per-fold learned feature weights (feature x fold)
result.feature_weights.to_csv("weights.tsv", sep="\t")

Input

rescore takes a single DataFrame with one row per PSM. Required:

• feature columns: all numeric columns by default, or pass feature_cols explicitly. Must be finite (no NaN/inf).
• is_decoy_col: boolean, True for decoy PSMs.
• spectrum_id_col: groups PSMs by spectrum for cross-validation and competition.
• peptidoform_col: a unique peptidoform key. A peptidoform rollup is always returned, grouping by this column verbatim (no parsing, same as peptide_col/ protein_col). By convention, peptidoform-level FDR is charge-independent, so this should typically hold sequence + modifications only, without charge.

Optional peptide_col (stripped sequence) and protein_col enable the peptide- and protein-level rollups.

How it works

1. Spectrum-grouped cross-validation. PSMs are split into n_folds folds by spectrum, so all PSMs of a spectrum stay in one fold. Each fold is scored by a model trained on the others, preventing data leakage.
2. Model training. For svm, the refinement loop (Käll et al. 2007): label targets below train_fdr as positives and all decoys as negatives, fit, rescore, repeat until the positive set stabilizes. The SVM class weight is tuned once (first iteration) and reused. If no trained model beats the best single feature, that feature is used instead. For lda, a single fit on all targets vs decoys.
3. Spectrum competition. By default (multi_rank_rescoring=False) only the best-scoring PSM per spectrum is kept, matching Percolator/mokapot/Sage. Set to True to keep and score all PSMs (all search-engine ranks).
4. FDR estimation. q-values by target-decoy competition; PEPs by a histogram-based binomial model with isotonic smoothing.
5. Rollup. Peptidoform and peptide levels use classic target-decoy competition (best PSM per group). Protein level uses picked-protein competition when decoy_pattern is given (target and decoy accessions paired by stripping the tag, higher score wins), or classic competition otherwise.

Result

rescore returns a RescoreResult with:

• psms: the identifier columns (spectrum_id_col, is_decoy_col, peptidoform_col, and peptide_col/protein_col if given) plus score, qvalue, pep. Features and other input columns are left out; the input row order and index are preserved, so they can be rejoined with features.loc[result.psms.index].
• peptidoforms: peptidoform-level table (always).
• peptides: peptide-level table, or None if no peptide_col.
• proteins: protein-level table, or None if no protein_col.
• pi0: estimated fraction of incorrect targets.
• n_iterations: refinement iterations per fold.
• feature_weights: per-fold learned feature weights.

Related projects

• MS²Rescore: the rescoring pipeline ristretto is built to plug into.
• MS²PIP: fragment intensity prediction, a source of rescoring features.
• DeepLC: retention time prediction.
• psm_utils: PSM parsing and handling.
• mokapot: the reference semi-supervised rescorer this library is modeled on.
• Percolator: the original semi-supervised target-decoy rescoring method.

References

• Käll, L., Canterbury, J. D., Weston, J., Noble, W. S., & MacCoss, M. J. (2007). Semi-supervised learning for peptide identification from shotgun proteomics datasets. Nature Methods. doi:10.1038/nmeth1113
• Käll, L., Storey, J. D., MacCoss, M. J., & Noble, W. S. (2008). Posterior error probabilities and false discovery rates: two sides of the same coin. Journal of Proteome Research. doi:10.1021/pr700739d
• Savitski, M. M., Wilhelm, M., Hahne, H., Kuster, B., & Bantscheff, M. (2015). A scalable approach for protein false discovery rate estimation in large proteomic data sets. Molecular & Cellular Proteomics. doi:10.1074/mcp.M114.046995
• Fondrie, W. E., & Noble, W. S. (2021). mokapot: fast and flexible semisupervised learning for peptide detection. Journal of Proteome Research. doi:10.1021/acs.jproteome.0c01010