Computational MS-based proteomics: Latest developments and tools for data analysis

This presentation will provide a quick overview of CompOmics and their open-source deep learning tools and computational frameworks for mass spectrometry-based proteomics. The behavior predictors include MS²PIP for fragmentation spectra, DeepLC for retention times, and IM2Deep for ion mobility, all of which are used in MS²Rescore to significantly boost peptide identification rates across challenging workflows like immunopeptidomics and metaproteomics. Beyond identification, we show the methods used to maximize the value of public proteomics data through large-scale reprocessing of PRIDE repositories, powering resources such as Scop3P for phosphosite structural context, tissue and cell-type prediction models, and MoDPA for discovering co-regulated post-translational modifications. Complementing these efforts, tools like lesSDRF tackle the critical metadata gap in public datasets, while AI-assisted approaches are being explored to further automate metadata extraction. Overall, the presentation highlights how CompOmics combines cutting-edge machine learning with systematic data reuse to extract deeper biological insight from the growing wealth of publicly available proteomics data.

Mass spectrometry (MS)-based proteomics is a well-established strategy for
analyzing complex biological mixtures. It is routinely used to compare relative protein
quantities in bulk cell lysates, analyze post-translational modifications, identify
protein–protein interactions, disease-specific protein variants, and the function and
dysfunction of biological processes in general. This variety of applications is
accompanied by an increasing number of instruments and data acquisition strategies
that generate specific types of data, which in turn require tailored analysis
algorithms. For these reasons, numerous workflows exist that are dedicated to the
analysis of MS-based proteomics data, and new ones are continually being
developed. Information about how these new workflows perform is often incomplete
and biased, relying mainly on performance results reported by the developers.
Furthermore, their performance can differ substantially from previously published
workflows or older versions of the same workflow. This lack of continuous and
transparent benchmarking makes it challenging to evaluate how new or updated
tools compare with each other.
We propose ProteoBench, a single platform that brings together software developers
and software users to provide an ever-evolving comparison of state-of-the-art
proteomics data processing tools. ProteoBench is an open-source resource
structured in independent benchmark modules proposed and openly discussed by
the community. Each module is designed to cover a specific aspect of proteomics
data analysis performed on the same input data. The platform currently contains 211
public workflow outputs from 23 tools. It allows to compare results from tools and
parameter sets dedicated to DDA and DIA data with mixed-species samples (bulk
and very low-input runs mimicking single-cell data), as well as a newly developped
module dedicated to de novo search.
ProteoBench enables the community to evaluate data analysis workflows, develop
benchmarking modules dedicated to specific comparisons, and discuss the best
methods to compare software tools. It ensures that benchmarks evolve alongside
advances in proteomics data analysis workflows; guides researchers towards the
best-suited tool and parameters for their specific project and data according to their
needs; and developers can test their newly developed tools or workflows privately,
before adding them as public references. This community-driven effort will increase
transparency and reproducibility between MS data analysis workflows, as well as
facilitate the development and publication of software workflows in the field.

Manuscript on BiorXiv: https://doi.org/10.64898/2025.12.09.692895
Application: https://proteobench.cubimed.rub.de/
Documentation: https://proteobench.readthedocs.io

Choosing a peptide search engine has long been a source of variability in
proteomics workflows: different algorithms, different results, and limited
consensus. In this talk, I present a systematic benchmarking study spanning
seven widely used search engines, four datasets acquired on distinct mass
spectrometry platforms, and protein databases of varying size and
composition. We evaluated three rescoring strategies -Percolator, MS2Rescore,
and Oktoberfest, alongside standard target-decoy FDR estimation, and
assessed their impact on peptide identification rates, inter-engine agreement,
and FDR reliability.
Our findings show that prediction-based rescoring dramatically reduces the
identification gap between search engines, effectively harmonizing results that
would otherwise differ quite substantially. We further examine how database
composition shapes these outcomes – particularly in metaproteomic settings –
and evaluate FDR control through entrapment-based analyses. Practical
considerations including computational runtime and resource consumption
are given as personal references.
The key message of the talk is: the choice of search engine matters far less
than you might think once modern rescoring is applied, the peptides are
within reach. What remains essential is careful validation of FDR control at
every step.

The incorporation of ion mobility (IM) in MS-based proteomics and metabolomics
enables isomer separation, resolves structural heterogeneity, and improves
identification confidence. A key derivative from IM, the collisional cross-section
(CCS), serves as a structural descriptor aiding peptide sequence and small molecule
identification. However, accurate CCS prediction remains challenging: post-
translational modifications (PTMs) strongly impact peptide CCS, yet existing
models struggle to handle their diversity. Additionally, both peptide precursors and
small molecules can exhibit multimodal IM distributions due to gas-phase
multiconformational behavior. Here presented are two deep learning approaches
for CCS prediction that explicitly address these challenges across peptides and
small molecules, IM2Deep and GraphXSection.
IM2Deep predicts CCS as a function of molecular composition and modification,
accurately predicting CCS values for peptides carrying any PTM, even those unseen
during training. Furthermore, by training on a dataset containing experimentally
observed multiconformation peptide ions, IM2Deep improves prediction accuracy
for both single- and multiconformational peptides.
GraphXSection, on the other hand, is a graph neural network that learns structure-
CCS relationships directly form molecular graphs. The model generalizes across
diverse chemical classes and ion adducts, outperforming existing graph-based
approaches. Incorporation of charge localization information enables accurate CCS
prediction for distinct protomers of the same molecule. Transfer learning further
extends performance to chemically challenging classes such as per- and
polyfluoralkyl substances. Finally, we evaluate GraphXSection on peptide CCS
prediction, exploring a one-architecture-fits-all- approach for CCS prediction.

1. Declercq A. and Devreese R et al. J. Proteome Res. 2025, 24, 3, 1067–1076, https://doi.org/10.1021/acs.jproteome.4c00609
2. Devreese R. et al. Anal. Chem. 2025, 97, 28, 15113–15121, https://doi.org/10.1021/acs.analchem.5c01142

Fragmentation spectra provide detailed information for identifying precursor ions such as
peptidoforms and proteoforms. Depending on sequence length and fragmentation conditions, a
substantial proportion of the signal observed in MS/MS spectra can arise from internal fragment
ions, that is, fragments generated by two or more cleavage events. These ions represent
internal sequence regions and may contain valuable information about amino acid composition
and post-translational modifications. Incorporating internal fragment ions may therefore increase
confidence in proteoform identification, particularly in challenging cases such as the localization
of closely spaced PTMs.
Here, we investigate MS data to define the conditions under which internal fragment ions
improve sequence interpretation rather than contribute with additional spectral complexity. To
support this analysis, we developed Fragment Explorer, which enables annotation of internal
fragment ions across diverse MS data sets. The tool can be used to interrogate MS/MS spectra
for internal ion signals, reveal statistical trends of terminal versus internal ions, and support
peptidoform identification such as manual resolution of assignment ambiguities.

Micha Johannes Birklbauer [1], Louise Buur [1], Vladimir Gorshkov [2], Arthur Grimaud [2],
Caroline Lennartson [3], Lev Levitsky [2], Veit Schwämmle [2], Zoltan Udvardy [4]
[1] University of Applied Sciences Upper Austria, Hagenberg, Austria
[2] University of Southern Denmark. Odense, Denmark
[3] Novo Nordisk Foundation, Center for Protein Research, Copenhagen, Denmark
[4] Institute of Pharmacology & Structural Biology (CNRS-IPBS), Toulouse, France