This PhD project aims to establish a (semi-)automatic framework to estimate the confidence of AI model predictions on minimal data. The approach will leverage techniques such as conformal prediction, explicitly considering unknown experimental parameters that influence peptide properties. The resulting confidence intervals will support more accurate data-driven rescoring in proteomics analyses.

Methodology

• Develop methods for model confidence estimation using conformal prediction and related uncertainty quantification techniques.
• Integrate confidence estimation into peptide property prediction workflows.
• Evaluate the methods on diverse datasets, with a focus on single-cell proteomics.

Expected Results

The project will deliver an enhanced data analysis pipeline that improves peptide and protein identification coverage and accuracy—particularly for post-translational modifications. While optimized for single-cell proteomics, the developed methods will be broadly applicable to complex proteomics datasets.

See more about the project at https://www.protaiomics.eu/project/dc2-tum/ and the overall grant here https://www.protaiomics.eu/research/