False discovery rate control enables trustworthy deep-learning-based de novo peptide sequencing

Abstract

Deep-learning-based de novo peptide sequencing identifies peptides beyond predefined databases, but confidence thresholds do not consistently indicate false discovery rates (FDRs). We develop Counterpart Calibration Theory (CCT), which uses counterpart counts to estimate incorrect target predictions and defines conditions for conservative FDR estimation and threshold selection. We implement CCT in {pi}-NovoQC at the peptide-spectrum match and peptide levels, integrating it with {pi}-xNovo sequencing models, protein inference and quantification. Evaluations across sequencing models and datasets examine FDR estimation and correct-identification retention. Isotope-labelled data provide complementary evidence for identifications recovered from database-unidentified spectra. In the Deep Human Proteome dataset, the workflow recovers low-abundance peptide evidence that extends protein-group inference and quantification. These results support the incorporation of de novo identifications into proteomic analyses at specified FDR thresholds.