Skip to content
Technical University of Munich

Advanced Practical Course Bioinformatics - 2026 - Advancing single-cell proteomics by foundational models

Unifying the state-of-the-art fragment ion intensity prediction models to improve experiment agnostic protein quantification

Topic

Accurate prediction of fragment ion intensities is a central component of modern proteomics workflows, directly impacting peptide identification, spectrum matching, and quantitative analyses in mass spectrometry (MS). Prosit, a GRU-based deep learning model, has set a benchmark in this domain by enabling highly accurate prediction of MS/MS fragment ion intensities. Building on this success, a Transformer-based Prosit architecture has recently been developed in our lab. 

Over the past years, multiple large-scale proteomics datasets have been collected and used to train independent Prosit models, each tailored to specific instrument types, ion types, or post-translational modifications (PTMs). While these specialized models achieve strong performance within their respective domains, they are limited in their ability to generalize across heterogeneous data sources. 

Furthermore, recent studies have shown that single-cell proteomics data differs from bulk data, exhibiting overall lower ion intensities and a reduced presence of annotated peaks. In a previous project, we demonstrated that incorporating the sum of y and b fragment ion intensities as an additional model input can improve Prosit’s prediction accuracy, particularly for low-intensity and single-cell–like spectra. 

This project aims to develop a unified, “all-in-one" Prosit Transformer model (Figure 1) trained on a diverse collection of proteomics datasets. The integrated training data will consist of: 

  • The PROSPECT PTMs dataset

  • Data containing an expanded collection of fragment ion types beyond classical b and y ions 

  • timsTOF-based data 

  • An external dataset containing the lactylation PTM acquired on two different instrument types 

  • Additional publicly available data 

By integrating these complementary data sources, the resulting model will be PTM-aware, capable of predicting fragment ion intensities across multiple instrument types, and flexible with respect to different ion types.  

The central challenge of this project lies in harmonizing heterogeneous datasets with varying fragmentation behaviors, ion definitions, and experimental biases. Addressing these challenges will involve careful data processing and standardization, model design choices, and evaluation strategies to ensure robust and generalizable performance. 

Figure 1

Aim

Development of a Unified Transformer Architecture 

  • Implement and refine a Transformer-based Prosit model capable of handling multi-modal inputs (modified sequences, collision energy, etc.) 

  • Design a flexible output head to accommodate an expanded vocabulary of fragment ions (e.g., a, b, c, x, y, z ions, and neutral losses). 

Data Harmonization and Standardization 

  • Integrate diverse datasets including PROSPECT PTM, timsTOF, MultiFrag, lactylation, and other datasets into a unified training pipeline. 

  • Develop robust preprocessing workflows to normalize intensities and standardize PTM encodings across different experimental setups. 

Performance Evaluation and Generalization 

  • Evaluate the "all-in-one" model against specialized state-of-the-art models. 

  • Assess the model's ability to generalize unseen PTMs and different mass spectrometry fragmentation methods (HCD, CID, ECD, ETciD, EID, UVPD). 

Impact on Downstream Quantification 

  • Investigate how improved prediction accuracy translates to better protein quantification and peptide-spectrum match (PSM) rescoring. 

General Schedule

Phase 1: Methods, Tools, Techniques

The first phase consists of a series of seminars and lectures in which you will learn the basics of various topics necessary for the project. There will be a mix of presentations by team members of our research group, practical sessions where applicable, and short presentations prepared by the participants:

Kickoff Seminar: Introduction to the course structure, organizational aspects, and overall project goals. Students will get an overview of the project and possible focus areas corresponding to the aims.

Proteomics Beyond Bulk: An introduction to general MS-based proteomics SCP technologies, experimental design, and computational approaches for peptide identification and quantification.

Hands-on Data Processing and deep Learning for Proteomics: Practical session on handling large-scale parquet proteomics datasets and encoding peptide sequences for DL models. Further, there will be an introduction to the DLOmix, a Python framework for Deep Learning in Proteomics, and PROSPECT (PROteometools SPECTrum compendium) large, annotated datasets. We will also dive deep into the evolution of Prosit, from GRUs to Transformers, and the importance of fragment ion intensity in rescoring.

Working with git as a team: This lecture will provide you with a project management system for working on larger coding projects. We will cover the concepts of issues, branches, pull requests, and the review process.

Additional Topics - In case you want to get deeper knowledge, we are open to holding an additional seminar on a topic of your choice. You decide.

Phase 2: Research project planning

In the second phase, we want you to prepare a detailed project plan. At the end of this phase, you will present your plan and discuss it with us. We will assist you during the planning of your project and provide you with feedback to ensure that you are able to bring your project to a successful end. Most importantly, you should discuss the following points: 

  • Requirement Analysis: Definition of research questions, pipeline extensions, and QC metrics to be implemented 

  • Organization: Milestones, task distribution, and time planning. Students are encouraged to use project management tools. Communication will be conducted via Slack 

Phase 3: Implementation and Research

This is the main phase of your project. According to your plan, you will implement, integrate, and test your work according to the plan. We will hold weekly progress meetings to discuss your progress.

  • Semester Work: Students are expected to work throughout the semester. On-site work is encouraged but not mandatory; virtual participation is possible 

  • Full-Time Block: Depending on progress, an optional intensive block of two to three weeks may be scheduled. The specific time and requirements will be discussed with you 

  • Submission: Deliverables include implemented code, documentation, benchmarking results, and a written report. Students will present their work in a final presentation 

Skills Gained

  • Expertise in Transformer-based deep learning architectures for biological sequences 

  • Experience in Large-scale data engineering and harmonization of heterogeneous scientific data 

  • In-depth knowledge of Mass Spectrometry principles and PTM biology 

  • Competence in benchmarking AI models against state-of-the-art biological baselines 

Organisation

Programming language: Python (PyTorch, TensorFlow, Pandas, NumPy) 

Must-have skills

  • Intermediate programming skills in Python
  • Solid understanding of Deep Learning
  • Experience with data science libraries

Good-to-have skills

  • Familiarity with MS proteomics data
  • Experience with GPU computing and model training at scale

Supervisors

Grading

  • Presentation [max. 30 minutes, whole team]
  • Report [~10 pages, including all sections]
  • Track record of indivual contributions must be supplied

Team size This project is designed for a team of 3-4 people.

Location OG-L 19, Maximus-von-Imhof-Forum 3, 85354 Freising

Submission

  • A Git repository with code and documentation

  • A comprehensive report detailing the project 

 

Material

All materials are made available in TUM Moodle.

Literature

Gessulat, S., Schmidt, T., Zolg, D. P., Samaras, P., Schnatbaum, K., Zerweck, J., Knaute, T., Rechenberger, J., Delanghe, B., Huhmer, A., Reimer, U., Ehrlich, H., Aiche, S., Kuster, B., & Wilhelm, M. (2019). Prosit: proteome-wide prediction of peptide tandem mass spectra by deep learning. Nature Methods, 16(6), 509–518. https://doi.org/10.1038/s41592-019-0426-7 

Gabriel, W., Zolg, D. P., Giurcoiu, V., Shouman, O., Prokofeva, P., Seefried, F., Bayer, F. P., Lautenbacher, L., Soleymaniniya, A., Schnatbaum, K., Zerweck, J., Knaute, T., Delanghe, B., Huhmer, A., Wenschuh, H., Reimer, U., Médard, G., Kuster, B., Wilhelm, M., . . . Wilhelm, M. (2025). Learning the Unseen: Data-Augmented Deep Learning for PTM Discovery with Prosit-PTM. bioRxiv (Cold Spring Harbor Laboratory). https://doi.org/10.1101/2025.11.07.687302 

Steen, H., & Mann, M. (2004). The abc’s (and xyz’s) of peptide sequencing. Nature Reviews Molecular Cell Biology, 5(9), 699–711. https://doi.org/10.1038/nrm1468 

Picciani, M., Gabriel, W., Giurcoiu, V., Shouman, O., Hamood, F., Lautenbacher, L., Jensen, C. B., Müller, J., Kalhor, M., Soleymaniniya, A., Kuster, B., Matthew, & Wilhelm, M. (2023). Oktoberfest: Open‐source spectral library generation and rescoring pipeline based on Prosit. PROTEOMICS, 24(8), e2300112. https://doi.org/10.1002/pmic.202300112 

Angelis, J., Schröder, E. A., Xiao, Z., Gabriel, W., & Wilhelm, M. (2025). Peptide Property Prediction for mass spectrometry using AI: An introduction to state of the art models. PROTEOMICS, 25(9–10), e202400398. https://doi.org/10.1002/pmic.202400398 

To top