Compact Transformers for Fast and Accurate De Novo Peptide Sequencing Across Diverse Organisms and Instruments
Title: Compact Transformers for Fast and Accurate De Novo Peptide Sequencing Across Diverse Organisms and Instruments
DNr: Berzelius-2026-355
Project Type: LiU Berzelius
Principal Investigator: Andre Mateus <andre.mateus@umu.se>
Affiliation: Umeå universitet
Duration: 2026-11-13 – 2027-06-01
Classification: 10203
Homepage: https://mateuslab.com/
Keywords:

Abstract

Mass spectrometry-based proteomics typically requires a protein database constructed specifically for that sample. However, in the case of samples that contain unknown organisms due to contamination or that originate from the environment, this might not be possible. De novo peptide sequencing predicts sequences directly from spectra, bypassing this issue, but the speed and sensitivity of current methods limit its widespread use. During our previous Berzelius allocation, we developed compact models for data-dependent and data-independent acquisition (DDA and DIA). Our DDA model reaches 69.4% peptide recall on the community standard nine-species development benchmark, compared with 59.9–67.6% for five comparators, while achieving 20-200 times higher throughput. Our DIA model recovers over seven times as many proteome-matching peptides as Cascadia, the current state-of-the-art alternative, using the same published test data. We will build on these results by comparing model architectures, training datasets, sampling strategies, ablation studies and measuring training variance across different seeds. The best model variants will then be trained large-scale and evaluated on a diverse set of projects and organisms to make sure that the models generalise well. Prepared datasets, checkpoints and an optimised implementation is already available. The work will support a methods publication and an open-source release with a command-line tool and GUI, integrated into NovoTax, which was developed during our previous Berzelius allocation. This will enable larger proteomics reanalyses, including our work on Bacteroidota O-glycosylation, and make de novo sequencing more practical on standard hardware.