Abstract
De novo peptide sequencing is crucial for identifying novel proteins, yet its broader application is constrained by the lack of a robust quality control system. In response, we developed a transformer-based model, π-xNovo, that accurately predicts peptides. By analyzing the model's attention matrix, we elucidated the contribution of spectral peaks to amino acid predictions, thus making de novo sequencing results explainable. Leveraging these insights, we designed a quality control system, π-xNovo-QC, which distinguishes peptide predictions with an accuracy exceeding 80% and a sensitivity above 90%. Applying this system to a large-scale deep human proteome dataset resulted in the identification of 1,931,761 additional peptides, marking a 137% increase over traditional database search results. These newly identified peptides with high confidence facilitated a 17.9% increase in protein identification, a 23.59% increase in the detection of single amino acid polymorphism events, and a 20.02% increase in exon-skipping splicing events. The deployment of this explainable AI system holds significant potential for expanding the application of de novo peptide sequencing, particularly in exploring the darker matter of the entire proteome universe.