Fig. 3: Overview of the rescoring pipeline and its results on ground truth datasets.

a Schematic illustration of the data-driven rescoring pipeline based on Prosit-XL as implemented in Oktoberfest. First, unfiltered results from supported XL-DBSEs (xiSEARCH or Scout) and mass spectrometry (MS) files (e.g., RAW) are required as input for rescoring. Oktoberfest performs spectrum annotation, normalized collision energy (NCE) calibration, and retrieves fragment ion intensity predictions from Prosit-XL to generate an extensive set of intensity-based features for each CSM provided by the XL-DBSE search results. Percolator is run at PSM level (rather than at CSM level). The final CSM score is obtained by taking the minimum percolator discriminant score of each PSM in a CSM and is submitted to xiFDR for FDR estimation on CSM-, peptide pair-, and PPI-level. b Vennbars show the number of identified CSMs and peptide pairs lost (orange), shared (blue), and gained (green), at an FDR of 1% on CSM- and peptide pair-levels when comparing results from xiSEARCH+Prosit-XL+xiFDR to xiSEARCH+xiFDR on a synthetic peptide dataset. Percentages inside the bars represent the actual FDRs, estimated by the ground truth synthetic peptide dataset. The analysis is based on both self- and between-link comparisons. Source data are provided in Supplementary Data 2. c Vennbars show the number of identified CSMs, peptide pairs, and PPIs lost (orange), shared (blue), and gained (green) at an FDR of 1% on CSM-, peptide pair-, and PPI-level when comparing results from Scout+Prosit-XL+xiFDR to Scout+xiFDR on a synthetic protein dataset. Percentages inside the bars represent the actual FDRs. The analysis is based on between-links only. Source data are provided in Supplementary Data 3.