π-MNovo improves de novo peptide sequencing through microbial-domain adaptation and evidence-guided candidate selection

Abstract

The high taxonomic and strain-level diversity of microbial communities makes it difficult for reference databases to fully represent the protein sequences present in metaproteomic samples, limiting database-dependent peptide identification. De novo peptide sequencing can recover peptide sequences directly from tandem mass spectra without relying on reference databases, providing complementary peptide evidence for metaproteomics. However, most existing de novo sequencing models were developed largely from non-microbial proteomic data and lack specific adaptation to microbial spectra. Here, we established a dedicated microbial spectral resource comprising more than 10 million annotated high-quality tandem mass spectra from 72 cultured microbial isolates, with peptide digests from each isolate separated into five high-pH reversed-phase fractions to increase the opportunity for deeper and more diverse peptide sampling. Using this resource, we developed {pi}-MNovo, which combines adaptation to microbial spectra with evidence-guided candidate selection to improve full-length peptide sequencing precision. On a seven-species external benchmark, {pi}-MNovo achieved 66.30% complete-peptide recall under the same residue-mass-based complete-peptide matching criterion applied to all models, exceeding four published models by 6.9-21.7% in relative recall, with consistent gains across species and peptide-length groups. In two independent metaproteomic datasets, {pi}-MNovo increased reference-supported taxonomic coverage, while synthetic-community analysis recapitulated the designed abundance ranking. From pFind-unidentified spectra, {pi}-MNovo recovered 1,587 pFind-unreported, reference-matched peptides that passed spectrum-evidence filtering. These results show that {pi}-MNovo expands recoverable peptide evidence and enhances the utility of de novo sequencing for metaproteomic analysis.