Separating the Wheat from the Chaff: Prediction-Assisted Rescoring of Peptidic Fragment Ion Spectra

Significance of the Topic

Accurate peptide identification in proteomics is critical for biomarker discovery, drug development and understanding biological processes. Traditional database search engines rely primarily on matching fragment masses, often neglecting the intensity information of MS/MS spectra. Integrating intensity-based predictions via deep learning can enhance the confidence of peptide-spectrum matches (PSMs), reduce false identifications and push the limits of complex applications such as immunopeptidomics and metaproteomics.

Study Objectives and Overview

The main goal of this work was to evaluate a novel rescoring algorithm that incorporates predicted fragment ion intensities into existing search workflows. Using a beta version of Thermo Scientific™ Proteome Discoverer™ 2.5 with SequestHT and a Prosit-derived rescoring node by MSAID, the study aimed to measure improvements in identification rates across diverse datasets: a HeLa digest, an immunopeptidomics (HLA) set and a human stool metaproteome.

Methodology and Instruments Used

The workflow combined:

• Proteome Discoverer 2.5 software with SequestHT for initial PSM proposals.
• A Prosit-based rescoring node predicting fragment intensities at peptide-specific collision energies.
• Percolator for semi-supervised machine learning classification using both conventional and intensity-based scoring features.

No new instrumentation was required beyond standard high-resolution tandem mass spectrometers. Data sources included Thermo Scientific™ Pierce™ HeLa digest, ProteomeXchange immunopeptidomics and metaproteomics raw files.

Main Results and Discussion

Rescoring with predicted intensities consistently improved identification across all levels:

• HeLa dataset: +10% PSMs, +8% peptides, +4% proteins.
• Metaproteomics (stool microbiome): +13% PSMs, +11% peptides, +10% proteins.
• Immunopeptidomics (HLA Class I): +59% PSMs, +55% peptides, +34% proteins.

Additionally, the rescoring approach allowed usage of stricter false discovery rate (FDR) thresholds (e.g. 0.1%) with minimal impact on identification counts, thereby increasing overall confidence. Receiver operating characteristic-like comparisons of target versus decoy scores demonstrated markedly improved separation after rescoring.

Benefits and Practical Applications

Incorporating intensity-based features into PSM scoring offers:

• Enhanced depth of proteome coverage by recovering spectra previously deemed low-quality.
• Greater robustness in high-complexity samples such as host-microbiome mixtures and immunopeptidomes.
• The ability to adopt more stringent statistical thresholds without sacrificing sensitivity, crucial for clinical or regulatory environments.

Future Trends and Opportunities

Developments to watch include:

• Real-time prediction and rescoring during data acquisition for adaptive instrument control.
• Extension of intensity-based rescoring to non-tryptic and cross-linked peptide workflows.
• Integration with multi-omic platforms and single-cell proteomics to maximize data yield.
• Community-driven models that refine predictions for post-translational modifications and novel fragmentation chemistries.

Conclusion

Deep-learning-driven rescoring of fragment ion spectra significantly enhances peptide and protein identification rates and confidence levels across diverse proteomic applications. By leveraging predicted intensities, this approach addresses inherent limitations of mass-based scoring, enabling more stringent FDR control and unlocking new potential in challenging workflows.

Reference

1. Gessulat S, Schmidt T, et al. Prosit: proteome-wide prediction of peptide tandem mass spectra by deep learning. Nat Methods. 2019;16(6):509-518.
2. Käll L, et al. Semi-supervised learning for peptide identification from shotgun proteomics datasets. Nat Methods. 2007;4:923-925.
3. Chong C, et al. Integrated proteogenomic deep sequencing and analytics accurately identify non-canonical peptides in tumor immunopeptidomes. Nat Commun. 2020;11(1):1293.
4. Rechenberger J, et al. Challenges in Clinical Metaproteomics Highlighted by the Analysis of Acute Leukemia Patients with Gut Colonization by Multidrug-Resistant Enterobacteriaceae. Proteomes. 2019;7(1).
5. Li J, et al. An integrated catalog of reference genes in the human gut microbiome. Nat Biotechnol. 2014;32:834-844.