Integrative Data science

INTEGRATIVE DATA SCIENCE

Mass spectrometry generates large and complex multidimensional datasets: a single LC-MS/MS run can contain tens to hundreds of thousands of spectra, and a mass spectrometry imaging experiment often millions. Turning these data into biological knowledge depends on computational methods. For glycans, this is harder. Unlike proteins, glycans are not directly encoded in the genome, and glycans with the same composition can differ in how their building blocks are linked and branched. Hence, a substantial part of the spectra from glycomics and glycoproteomics experiments remains unidentified with standard search strategies. Antibodies, immunopeptides, and proteins with complex post-translational modifications (PTMs) pose a similar problem, because there is no reference template to search against.

Protein language models such as ESM and ProtBERT, trained on millions of unannotated sequences, have improved the prediction of protein structure and function. The same idea can be applied to mass spectra. Within BioBeyond_NL, we are developing MS foundation models that learn from publicly available mass spectrometry data and are then fine-tuned for a specific task, such as glycopeptide identification or de novo sequencing, using a limited amount of annotated data. Such models make predictions faster than database searches, so they can also be used during a measurement, for example to decide which ions to fragment next.

 ACCESS TO AI AND DATA INFRASTRUCTURE

Data science work in BioBeyond_NL takes place at all four partner sites, next to the facilities that produce the data. At Utrecht University, the AI Technology for Life group and the Biomolecular Mass Spectrometry and Proteomics group (prof. dr. Sanne Abeln, dr. Pavel Sinitcyn) develop the MS foundation models and machine-learning tools for PTM localization and de novo sequencing. This builds on earlier work on learning from limited annotated data and on proteomics software such as MaxQuant, MaxDIA, and Perseus. At Radboudumc, the group of dr. Hans Wessels develops software for real-time glycoproteomics data processing, including the Myriad workflow in GlycoScape and GlycoPaSER. These tools make results-dependent acquisition (rda-MS) possible: glycopeptides identified during the measurement determine the next fragmentation steps. In Leiden, dr. Magnus Palmblad (LUMC) brings expertise in bioinformatics and data management. LUMC's ion mobility measurements of glycans and glycan fragments will also feed a database of collision cross section (CCS) values, which helps to identify unknown structures. In Maastricht, dr. Benjamin Balluff (M4i) works on the analysis of mass spectrometry imaging data, including co-registration with microscopy and clustering of cells and tissue regions.

Together, these groups are building a platform, still under development, that gives researchers practical tools to identify and analyze peptides, proteins, glycans, and PTMs in mass spectrometry data. It has three parts:

• Specialized tools for glycomics and glycoproteomics identification, including real-time identification to guide data acquisition;
• AI models for faster MS data analysis, including PTM localization and de novo sequencing of antibodies and immunopeptides;
• Integrated workflows for spatial and multimodal multi-omics data, linking MS imaging with (glyco)proteomics, microscopy, and single-cell and spatial transcriptomics.

Computing runs on local GPU servers and storage within the consortium, supplemented by national SURF services such as the Snellius supercomputer. For each user project, the Data Science expert team first estimates the storage and compute needs together with the user and, where needed, arranges extra resources through SURF. We provide standardized workflows (e.g., Snakemake, Nextflow) and expert consultation, and we advise on FAIR data management and on depositing data in public repositories such as PRIDE/ProteomeXchange. Training is offered through hands-on workshops and together with the national BioSB research school and ELIXIR. AI models and data analysis methods developed within BioBeyond_NL are released under the MIT license by default. The goal is to make it easier for experimental researchers to analyze their own data.

Contact us here about your access inquiries.