See all

Confidence prediction of fine-tuned AI peptide property models for single-cell proteomics data analysis

Technical University of Munich (TUM)

Munich

Germany

2/16

Supervision

Mathias Wilhelm
TUM, 1st Supervisor
Viktoria Dorfer, 2nd Supervisor

Objectives

Implementation and evaluation of a method to predict the confidence peptide property prediction model on limited datasets, particularly addressing the unique characteristics of e.g. single-cell proteomics data. Further, integration of confidence prediction into peptide identification pipelines while maintaining accurate false discovery rate estimation on all levels.

Methodology

Establish a (semi-)automatic method to enable the estimation of the confidence of a model’s prediction on minimal data using e.g. conformal prediction, accounting for the unknown experimental parameters influencing peptide properties. The resulting confidence intervals should enable a more accurate data-driven rescoring.

Required skills

Computational mass spectrometry, deep learning, bioinformatics, peptide property prediction, confidence/conformal prediction

Expected results

An enhanced data analysis pipeline particularly, but not exclusively, for single cell proteomics data. The developed methods will enable deeper peptide and protein coverage with increased accuracy particularly for post-translational modifications

Planned Secondments

Host: FHOOE (V. Dorfer), Duration: 2 Months; When: Year 1, Goal: Adding confidence prediction to MS2 classification using NNs.

Host: BRUKER (D. Trede), Duration: 1 Month; When: Year 2, Goal: Testing confidence prediction on single cell timsTOF data.

Host: SANGER (B. Lehner), Duration: 1 Month; When: Year 3, Goal: Adding interpretability to confidence prediction.

Enrolment in doctoral programs

Graduate Center of Life Sciences

References

https://doi.org/10.1002/pmic.202400398
https://doi.org/10.1016/j.mcpro.2024.100798
https://github.com/wilhelm-lab/oktoberfest
https://github.com/wilhelm-lab/koina
https://github.com/wilhelm-lab/dlomix
https://huggingface.co/collections/Wilhelmlab/prospect-ptms-665db48431a7e844634660ba