The Good, the Bad and the Ugly: when statistics tells you to throw away peptide IDs

Importance of the Topic

In bottom-up proteomics, reliably distinguishing true peptide-spectrum matches (PSMs) from false hits is critical for data quality and downstream biological interpretation. Standard target-decoy approaches control global false discovery rates (FDR) but may discard many valid identifications arising from low-quality or chimeric spectra. Improved validation methods can increase proteome coverage and confidence, particularly in complex samples such as HLA peptidomes or proteogenomic analyses.

Objectives and Study Overview

This study compares classical single-score target-decoy workflows with semi-supervised machine learning (Percolator) to identify the optimal strategy for PSM validation. The team evaluates identification rates at the PSM, peptide, and protein levels and visualizes representative spectra to illustrate strengths and limitations of each method.

Materials and Methods

The test sample was a 200 ng HeLa protein digest analyzed on a 60 min gradient (8–30 % B) with a Thermo Scientific™ Orbitrap Exploris™ 480 operated in data-dependent acquisition. Data processing used Thermo Scientific™ Proteome Discoverer™ 2.5 with classical target-decoy scorers and Percolator in both concatenated and separated modes.

Instrumental Setup

• LC: EASY-Spray PepMap RSLC C18, 25 cm column, gradient to 90 % acetonitrile.
• MS: Orbitrap Exploris 480 with 1 s cycle time for precursor selection.
• Software: Proteome Discoverer 2.5, Percolator for semi-supervised rescoring, MSPepSearch for spectral library matching.

Main Results and Discussion

Percolator outperformed classical target-decoy validation, yielding higher numbers of validated PSMs, peptides, and protein groups at 1 % FDR. The concatenated mode provided slightly better separation of targets versus decoys than separated searches. Nevertheless, many ‘ugly’ spectra—those with high interference or marginal fragment series—remain unassigned by all methods. Examples reveal:

• Spectra with high isolation interference (> 90 %) rescued by spectral library scoring but missed by database search rescoring.
• PSMs containing complete fragment series but low engine scores still discarded, indicating room for improved feature integration.

Benefits and Practical Applications

Adopting semi-supervised machine learning improves sensitivity and expands proteome coverage, which is particularly valuable for:

• Low-abundance peptides in immunopeptidomics or proteogenomics.
• Complex metaproteomic samples with broad search spaces.
• Quality control workflows where high confidence assignment is critical.

Future Trends and Applications

Ongoing developments aim to integrate predicted fragment intensities and retention times into rescoring algorithms. Upcoming Proteome Discoverer updates will feature prediction-assisted rescoring modules, and advances in deep-learning fragmentation models promise enhanced discrimination of true versus false matches. These innovations will further reduce the number of overlooked PSMs in challenging datasets.

Conclusion

Semi-supervised machine learning via Percolator represents a significant improvement over classical single-score target-decoy FDR control, increasing validated identifications while maintaining error rates. However, many valid PSMs remain hidden in low-quality spectra. Integrating additional predictive features and library information will be key to capturing these ‘good’ matches.

References

1. Käll L., Canterbury J.D., Weston J., Noble W.S., MacCoss M.J. Semi-supervised learning for peptide identification from shotgun proteomics datasets. Nat. Methods 2007, 4, 923–925.