Deep Learning for de novo peptide sequencing

Project Description

The three major classes of molecules of life are DNA, RNA, and proteins. Unlike for DNA and RNA, there is, to date, no accurate, high-throughput sequencing technology for proteins. The closest technology is tandem mass spectrometry, which yields mass spectra of protein fragments called peptides. Highly accurate de novo peptide sequencing (DNPS), i.e., determining peptide amino acid sequences solely from tandem mass spectra, will make proteomics amenable for applications including genotyping, cancer surveillance, pathogen surveillance, immuno-oncology, metagenomics, and paleogenomics.

In this project, we developed Koina, an open, containerized, web-accessible service that standardizes and accelerates access to peptide property prediction models. On top of Koina, we created Oktoberfest, a search-engine-agnostic Python package that generates spectral libraries and rescores peptide-spectrum matches. Leveraging these foundations, we introduced Spectralis, a DNPS method for tandem mass spectrometry. It features a convolutional neural network layer connecting peaks in spectra spaced by amino acid masses, proposes classifications of fragment ion series, and provides a peptide-spectrum confidence score. Together, these components improve peptide recall over initial predictions and make state-of-the-art DNPS accessible and scalable.

Results

• Developed Koina, an open-source containerized, decentralized, and online-accessible high-performance prediction service.
• Used FragPipe as an example to demonstrate Koina integration with existing proteomics software tools.
• Compiled Oktoberfest, an open-source Python package of our spectral library generation and rescoring pipeline.
• Demonstrated its ability to improve rescoring analyses on two distinct use cases.
• Oktoberfest is freely available on GitHub.
• Introduced Spectralis, a de novo peptide sequencing method for tandem mass spectrometry.
• Spectralis achieved 40% sensitivity at 90% precision, nearly doubling the state-of-the-art.
• Application to unidentified spectra confirmed its superiority and showcased its applicability to variant calling.

Follow-up

PostDoc Joel Lapin joined Computational Mass Spectrometry and is working on alternatives for encoding tandem mass spectra (Lapin et al. 2025) collaboratively with researchers from Applied Systems Biology at KTH (Royal Institute of Technology in Stockholm).

Participation in the “OpenMS Summer 2024 Fellowship” resulted in hosting a PhD student from Adrem Data Lab at the University of Antwerp. This led to a collaboration on a novel benchmark platform for DNPS tools (Pominova et al. 2026).

Another goal of our labs will be the development of tools for DNPS of post-translationally modified peptides, including support of chimeric spectra (multiple peptides). To this end, we are generalizing the underlying algorithms to predict modified residues (Klaproth-Andrade et al. 2026), including the transformer Casanovo (Straub et al. 2025). For validation, we will further improve Prosit (Gabriel et al. 2025), a foundation for Spectralis.