Unlocking Bacterial Epitope Discovery: How Rescoring and PEAKS Deep Learning Boost Immunopeptidomics

Antimicrobial resistance is a serious and growing global health threat. Developing novel bacterial vaccines is a cost-effective strategy to prevent it. At the core of vaccine development is the identification of bacterial epitopes, which can be achieved through mass spectrometry-based immunopeptidomics. However, analyzing immunopeptidomics data presents unique challenges for any bioinformatics pipeline.

To address this, Prof. Impens and colleagues developed an optimized bioinformatics framework designed to improve bacterial epitope identification. Their workflow combines search results from multiple search engines, including PEAKS (version 12), and applies a rescoring step to enhance the final identifications.

This integration and rescoring strategy increased epitope identifications by approximately 20 percent in most cases, with one exception: PEAKS 12. This is expected since PEAKS 12 already includes a built-in rescoring feature known as Deep Learning Boost. As a result, there was minimal additional improvement observed for PEAKS search results.

Beyond identification and quantification, the authors also explored ambiguous peptide-spectrum matches (PSMs) reported by different search engines. One example in the paper highlights a two-amino-acid swap between PEAKS and Comet results. Because there are no clear fragment ions in that region, the difference is difficult to resolve using fragmentation alone. The authors propose that other features, such as spectral similarity, can help. PEAKS showed higher spectral correlation in this case, with a value of 0.98 compared to other tools.

Retention time (RT) prediction may also help resolve ambiguous cases. To test this, PEAKS users can export top candidates for the target MS2 scan from the de novo or DeepNovo result nodes and check whether those algorithms are already ranking candidates based on RT differences.

Tip:
In PEAKS database searches, you can enable rescoring by checking the “Deep Learning Boost” box in the search parameters. For DeepNovo-based workflows, rescoring is turned on by default.

📄 Read the full article here:
ACS Publications – Journal of Proteome Research