Transfer learning peptide properties of cross-linked peptides
[call for this position closed]
The objectives of this subproject are the development of an easy to use framework for transfer learning of Prosit prediction models, to allow the model generation for peptide classes where insufficient training data is available, and test this on cross-linked peptides with the aim to increase sensitivity and specificity of identifying such peptides by integrating peptide properties such as fragment intensities, retention time and MS-accessibility.
Transfer learning is a powerful mechanism by which prior models can be re-trained to be able to predict properties for classes where insufficient data is available to train a model from scratch. This is particularly the case for cross-linked peptides where insufficient training is available. However, specifically, MS-cleavable cross-linkers can be viewed as modifications where existing prediction models can be retrained. Earlier research has shown that this is possible for regular modifications and that the integration of additional properties such as fragment intensities and retention time can significantly increase the confidence in the correct identification of peptides from spectra.
cross-linking data analysis
fragment ion intensity prediction
Fully functional pipeline which can rescue peptide spectrum matches of cross-linked peptides by the integration and scoring of additional peptide properties such as fragment intensities.
Host: ANAXOMICS (J. Farrés), Duration: 1 Month; When: Year 1, Goal: Network analysis.
Host: VIB (S. Degroeve), Duration: 1 Month; When: Year 2; Goal: Exchange scoring algorithms.
Host: THERMO (B. Delanghe), Duration: 1 Month; When: Year 3; Goal: Integration into ProteomeDiscoverer.
Enrolment in doctoral programs
ESRs located at TUM will be enrolled at TUM Graduate Center Weihenstephan (https://www.gzw.wzw.tum.de).
1. 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.-C., Aiche, S., Kuster, B., & Wilhelm, M. (2019b). 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
2. O’Reilly, F. J., & Rappsilber, J. (2018). Cross-linking mass spectrometry: methods and applications in structural, molecular and systems biology. Nature Structural & Molecular Biology, 25(11), 1000–1008. https://doi.org/10.1038/s41594-018-0147-0