MainHigh-throughput proteomics heavily relies on mass spectrometry to detect and quantify peptides, which are then used to infer protein presence and abundance1,2. Peptides are identified from acquired fragmentation (MS2) spectra, in which each peak has an associated mass-to-charge ratio (m/z) and intensity value, most of which are related to the amino acid sequence of the fragmented precursor peptide. A peptide search engine then uses the predictable fragmentation behavior of peptides to score peptide-spectrum matches (PSMs) for the recorded MS2 spectra3,4,5,6. However, even with state-of-the-art instruments and identification algorithms, the exact peptide sequence cannot always be unambiguously determined7.Three main factors contribute to this challenge of identification ambiguity. First, some peptides generate b-ions and y-ions of (nearly) identical mass and are thus difficult to distinguish by mass spectrometry. Well-known examples are isobaric amino acids such as leucine and isoleucine, pairs that become isobaric through post-translational modifications (PTMs), such as aspartic acid and deamidated asparagine, and nearly isobaric pairs such as oxidized methionine and phenylalanine. Resolving these requires spectral evidence beyond b-ion and y-ion m/z values. Second, ambiguity increases when no fragment ion is observed at a cleavage site in either the b-ion or the y-ion series, that is, a missing complementary ion pair. This leaves the site unannotated and increases the number of isobaric sequences that can explain the observed spectrum well7, and often results from suboptimal fragmentation. Third, additional fragment peaks due to co-fragmentation or experimental noise can create spurious amino acid ladders, again increasing the number of likely peptide candidates.The most common approach to mitigate sequence ambiguity is to dramatically limit the search space to only the very small subset of peptide sequences that are expected in the sample8. This is achieved through sequence database searching, where observed spectra are matched only against sequences from the predefined protein database9,10,11,12,13. One key constraint of this approach is the inability to identify proteins or isoforms that are not part of the database. Proteogenomics addresses this by including sequence variants derived from genomic and transcriptomic data14, and metaproteomics workflows often rely on large, across-species catalogs or entire reference databases such as UniProtKB or NCBI RefSeq15. Expanding the database, however, increases the number of candidate peptide sequences and therefore the potential for identification ambiguity16,17, making it a less efficient filter. Machine learning-based rescoring approaches, such as MS2Rescore, mitigate this18 by considering not only b- and y-ion m/z information but also fragment intensity patterns19,20, retention time21 and collisional cross section22,23,24, integrating these into a scoring model trained to separate ‘likely true’ from ‘known false’, that is, decoy matches25,26. Even so, they remain reliant on databases and are thus unable to identify novel peptides or unexpected variants.De novo sequencing is a promising alternative, as it does not rely on sequence databases and can thus, in theory, identify any peptide from a fragmentation spectrum. However, because de novo sequencing essentially removes all constraints on the search space, the ambiguity issue returns in full force. Recent deep learning-based models have nevertheless pushed de novo performance to new levels27. These networks typically predict peptide sequences autoregressively, use beam search to explore multiple candidates in parallel and apply precursor mass control. Since the development of the Casanovo model28, attention mechanisms have become central to learning the relationships between peak intensities and masses in an MS2 spectrum and the peptide subsequence under construction. Nevertheless, questions remain regarding (1) what the strengths and weaknesses of different de novo models are and (2) to what extent they can deal with the identification ambiguity problem.In this study, we therefore systematically benchmark eight deep learning-based de novo sequencing models (AdaNovo29, Casanovo28, ContraNovo30, InstaNovo31, NovoB32, PepNet33, π-HelixNovo34 and π-PrimeNovo35) alongside three post-processing methods (InstaNovo+31, MS2Rescore18 and Spectralis36). First, we evaluate their ability to reproduce database search results. We then assess the extent to which post-processing methods, including rescoring and reranking algorithms, can improve the original de novo results. In cases where de novo and database-derived PSMs disagree, we analyze how prediction errors relate to specific spectral characteristics. Finally, we demonstrate that although de novo sequencing excels under specific conditions, and often times parts of the sequence can be unambiguously identified, its limitations, particularly with regard to low-quality spectra, mirror those of database-based approaches.ResultsWe first benchmark eight state-of-the-art models and assess their agreement, then evaluate post-processing methods, and finally perform an error analysis to characterize the limitations of the current models.Assessment of accuracy and consensus across de novo modelsModel performance is reported against a gold standard of database search identifications, obtained with Sage11 in chimeric mode and MSGF+37 at a q-value Article PubMed CAS Google Scholar Schulte, D. & Snijder, J. A handle on mass coincidence errors in de novo sequencing of antibodies by bottom-up proteomics. J. Proteome Res. 23, 3552–3559 (2024).Article PubMed PubMed Central CAS Google Scholar Verheggen, K. et al. Anatomy and evolution of database search engines-a central component of mass spectrometry based proteomic workflows. Mass Spectrom. Rev. 39, 292–306 (2020).Article PubMed CAS Google Scholar Tyanova, S., Temu, T. & Cox, J. The MaxQuant computational platform for mass spectrometry-based shotgun proteomics. Nat. Protoc. 11, 2301–2319 (2016).Article PubMed CAS Google Scholar Kong, A. T., Leprevost, F. V., Avtonomov, D. M., Mellacheruvu, D. & Nesvizhskii, A. I. MSFragger: ultrafast and comprehensive peptide identification in mass spectrometry-based proteomics. Nat. Methods 14, 513–520 (2017).Article PubMed PubMed Central CAS Google Scholar Lazear, M. R. Sage: an open-source tool for fast proteomics searching and quantification at scale. J. Proteome Res. 22, 3652–3659 (2023).Article PubMed CAS Google Scholar Alves, G. & Yu, Y.-K. Robust accurate identification of peptides (RAId): deciphering MS2 data using a structured library search with de novo based statistics. Bioinformatics 21, 3726–3732 (2005).Zhang, J. et al. PEAKS DB: de novo sequencing assisted database search for sensitive and accurate peptide identification. Mol. Cell. Proteomics 11, M111.010587 (2012).Article PubMed Google Scholar Nesvizhskii, A. I. Proteogenomics: concepts, applications and computational strategies. Nat. Methods 11, 1114–1125 (2014).Article PubMed PubMed Central CAS Google Scholar Van Den Bossche, T. et al. The microbiologist’s guide to metaproteomics. iMeta 4, e70031 (2025).Noble, W. S. Mass spectrometrists should search only for peptides they care about. Nat. Methods 12, 605–608 (2015).Article PubMed PubMed Central CAS Google Scholar Sticker, A., Martens, L. & Clement, L. Mass spectrometrists should search for all peptides, but assess only the ones they care about. Nat. Methods 14, 643–644 (2017).Article PubMed CAS Google Scholar Declercq, A. et al. MS2Rescore: data-driven rescoring dramatically boosts immunopeptide identification rates. Mol. Cell. Proteomics 21, 100266 (2022).Article PubMed PubMed Central CAS Google Scholar Degroeve, S. & Martens, L. MS2PIP: a tool for MS/MS peak intensity prediction. Bioinformatics 29, 3199–3203 (2013).Article PubMed PubMed Central CAS Google Scholar Gessulat, S. et al. Prosit: proteome-wide prediction of peptide tandem mass spectra by deep learning. Nat. Methods 16, 509–518 (2019).Article PubMed CAS Google Scholar Bouwmeester, R. et al. Transfer learning in DeepLC improves LC retention time prediction across substantially different modifications and setups. Nat. Commun. 17, 2601 (2026)Guo, R. et al. Highly accurate and large-scale collision cross sections prediction with graph neural networks. Commun. Chem. 6, 139 (2023).Article PubMed PubMed Central Google Scholar Devreese, R. et al. Collisional cross-section prediction for multiconformational peptide ions with IM2Deep. Anal. Chem. 97, 15113–15121 (2025).Article PubMed PubMed Central CAS Google Scholar Meier, F. et al. Deep learning the collisional cross sections of the peptide universe from a million experimental values. Nat. Commun. 12, 1185 (2021).Article PubMed PubMed Central CAS Google Scholar Fondrie, W. E. & Noble, W. S. mokapot: Fast and flexible semisupervised learning for peptide detection. J. Proteome Res. 20, 1966–1971 (2021).Article PubMed PubMed Central CAS Google Scholar Käll, L., Canterbury, J. D., Weston, J., Noble, W. S. & MacCoss, M. J. Semi-supervised learning for peptide identification from shotgun proteomics datasets. Nat. Methods 4, 923–925 (2007).Article PubMed Google Scholar Bittremieux, W. et al. Deep learning methods for de novo peptide sequencing. Mass Spectrom. Rev. 45, 507–526 (2026).Article PubMed CAS Google Scholar Yilmaz, M. et al. Sequence-to-sequence translation from mass spectra to peptides with a transformer model. Nat. Commun. 15, 6427 (2024).Article PubMed PubMed Central CAS Google Scholar Xia, J. et al. AdaNovo: Towards robust de novo peptide sequencing in proteomics against data biases. Adv. Neural Inf. Process. Syst. 37, 1811–1828 (2024).Jin, Z. et al. ContraNovo: a contrastive learning approach to enhance de novo peptide sequencing. In Proc. AAAI Conference on Artificial Intelligence (eds Wooldridge, M., Dy, J. & Natarajan, S.) Vol. 38, 144–152 (AAAI Press, 2024).Eloff, K. et al. InstaNovo enables diffusion-powered de novo peptide sequencing in large-scale proteomics experiments. Nat. Mach. Intell. 7, 565–579 (2025).Lee, S. & Kim, H. Bidirectional de novo peptide sequencing using a transformer model. PLoS Comput. Biol. 20, e1011892 (2024).Article PubMed PubMed Central CAS Google Scholar Liu, K., Ye, Y., Li, S. & Tang, H. Accurate de novo peptide sequencing using fully convolutional neural networks. Nat. Commun. 14, 7974 (2023).Article PubMed PubMed Central CAS Google Scholar Yang, T. et al. Introducing π-HelixNovo for practical large-scale de novo peptide sequencing. Brief. Bioinform. 25, bbae021 (2024).Zhang, X. et al. π-PrimeNovo: an accurate and efficient non-autoregressive deep learning model for de novo peptide sequencing. Nat. Commun. 16, 267 (2025).Klaproth-Andrade, D. et al. Deep learning-driven fragment ion series classification enables highly precise and sensitive de novo peptide sequencing. Nat. Commun. 15, 151 (2024).Article PubMed PubMed Central CAS Google Scholar Kim, S., Gupta, N. & Pevzner, P. A. Spectral probabilities and generating functions of tandem mass spectra: a strike against decoy databases. J. Proteome Res. 7, 3354–3363 (2008).Article PubMed PubMed Central CAS Google Scholar Van Puyvelde, B. et al. A comprehensive LFQ benchmark dataset on modern day acquisition strategies in proteomics. Sci. Data 9, 126 (2022).Van Den Bossche, T. et al. Critical assessment of metaproteome investigation (CAMPI): a multi-laboratory comparison of established workflows. Nat. Commun. 12, 7305 (2021).Wen, B. & Noble, W. S. A multi-species benchmark for training and validating mass spectrometry proteomics machine learning models. Sci. Data 11, 1207 (2024).Article PubMed PubMed Central Google Scholar Yang, H., Chi, H., Zeng, W.-F., Zhou, W.-J. & He, S.-M. pNovo 3: precise de novo peptide sequencing using a learning-to-rank framework. Bioinformatics 35, i183–i190 (2019).Yang, K. L. et al. MSBooster: improving peptide identification rates using deep learning-based features. Nat. Commun. 14, 4539 (2023).Buur, L. M. et al. MS2Rescore 3.0 is a modular, flexible, and user-friendly platform to boost peptide identifications, as showcased with MS Amanda 3.0. J. Proteome Res. 23, 3200–3207 (2024).Miller, S. E., Rizzo, A. I. & Waldbauer, J. R. PostNovo: postprocessing enables accurate and FDR-controlled de novo peptide sequencing. J. Proteome Res. 17, 3671–3680 (2018).Beslic, D., Tscheuschner, G., Renard, B. Y., Weller, M. G. & Muth, T. Comprehensive evaluation of peptide de novo sequencing tools for monoclonal antibody assembly. Brief. Bioinform. 24, bbac542 (2023).Muth, T. & Renard, B. Y. Evaluating de novo sequencing in proteomics: already an accurate alternative to database-driven peptide identification?. Brief. Bioinform. 19, 954–970 (2018).Article PubMed CAS Google Scholar Tran, N. H. et al. NovoBoard: a comprehensive framework for evaluating the false discovery rate and accuracy of de novo peptide sequencing. Mol. Cell. Proteomics 23, 100849 (2024).Zhou, J. et al. NovoBench: benchmarking deep learning-based de novo peptide sequencing methods in proteomics. Adv. Neural Inf. Process. Syst. 37, 104776–104791 (2024).Devreese, R. et al. ProteoBench: the community-curated platform for comparing proteomics data analysis workflows. Preprint at bioRxiv https://doi.org/10.64898/2025.12.09.692895 (2025).Kalhor, M. et al. Prosit-XL: enhanced cross-linked peptide identification by accurate fragment intensity prediction to study protein-protein interactions and protein structures. Nat. Commun. 16, 5429 (2025).McDonnell, K., Howley, E. & Abram, F. The impact of noise and missing fragmentation cleavages on de novo peptide identification algorithms. Comput. Struct. Biotechnol. J. 20, 1402–1412 (2022).Article PubMed PubMed Central Google Scholar Devabhaktuni, A. & Elias, J. E. Application of de novo sequencing to large-scale complex proteomics data sets. J. Proteome Res. 15, 732–742 (2016).Article PubMed CAS Google Scholar Guthals, A., Clauser, K. R., Frank, A. M. & Bandeira, N. Sequencing-grade de novo analysis of MS/MS triplets (CID/HCD/ETD) from overlapping peptides. J. Proteome Res. 12, 2846–2857 (2013).Article PubMed PubMed Central CAS Google Scholar O’Bryon, I., Jenson, S. C. & Merkley, E. D. Flying blind, or just flying under the radar? The underappreciated power of de novo methods of mass spectrometric peptide identification. Protein Sci. 29, 1864–1878 (2020).Article PubMed PubMed Central Google Scholar Lin, A., Plubell, D. L., Keich, U. & Noble, W. S. Accurately assigning peptides to spectra when only a subset of peptides are relevant. J. Proteome Res. 20, 4153–4164 (2021).Article PubMed PubMed Central CAS Google Scholar Yu, F. et al. Identification of modified peptides using localization-aware open search. Nat. Commun. 11, 4065 (2020).Article PubMed PubMed Central CAS Google Scholar Lauridsen, J. et al. InstaNovo-P: a de novo peptide sequencing model for phosphoproteomics. Nat. Commun. 17, 9277 (2026).Wang, Y. et al. Transforming de novo peptide sequencing by explainable AI. Preprint at https://doi.org/10.21203/rs.3.rs-4716013/v1 (2024).Sundararajan, M., Taly, A. & Yan, Q. Axiomatic attribution for deep networks. In Proc. 34th International Conference on Machine Learning (eds Precup, D. & Teh, Y. W.) Vol. 70, 3319–3328 (PMLR, 2017).Alves, G. & Yu, Y.-K. Confidence assignment for mass spectrometry based peptide identifications via the extreme value distribution. Bioinformatics 32, 2642–2649 (2016).Johnson, R. S. et al. Assessing protein sequence database suitability using de novo sequencing. Mol. Cell. Proteomics 19, 198–208 (2020).Article PubMed CAS Google Scholar Frank, A. M., Savitski, M. M., Nielsen, M. L., Zubarev, R. A. & Pevzner, P. A. De novo peptide sequencing and identification with precision mass spectrometry. J. Proteome Res. 6, 114–123 (2007).Article PubMed PubMed Central CAS Google Scholar Karunratanakul, K., Tang, H.-Y., Speicher, D. W., Chuangsuwanich, E. & Sriswasdi, S. Uncovering thousands of new peptides with sequence-mask-search hybrid de novo peptide sequencing framework. Mol. Cell. Proteomics 18, 2478–2491 (2019).Article PubMed PubMed Central CAS Google Scholar Tabb, D. L., Ma, Z.-Q., Martin, D. B., Ham, A.-J. L. & Chambers, M. C. DirecTag: accurate sequence tags from peptide MS/MS through statistical scoring. J. Proteome Res. 7, 3838–3846 (2008).Ma, B. & Johnson, R. De novo sequencing and homology searching. Mol. Cell. Proteomics 11, O111.014902 (2012).Article PubMed Google Scholar Sunyaev, S., Liska, A. J., Golod, A., Shevchenko, A. & Shevchenko, A. MultiTag: multiple error-tolerant sequence tag search for the sequence-similarity identification of proteins by mass spectrometry. Anal. Chem. 75, 1307–1315 (2003).Vande Moortele, T. et al. Unipept in 2024: expanding metaproteomics analysis with support for missed cleavages and semitryptic and nontryptic peptides. J. Proteome Res. 24, 949–954 (2025).Article PubMed Google Scholar Potgieter, M. G. et al. MetaNovo: an open-source pipeline for probabilistic peptide discovery in complex metaproteomic datasets. PLOS Comput. Biol. 19, e1011163 (2023).Gabriels, R., Declercq, A., Bouwmeester, R., Degroeve, S. & Martens, L. psm_utils: a high-level python API for parsing and handling peptide-spectrum matches and proteomics search results. J. Proteome Res. 22, 557–560 https://doi.org/10.1021/acs.jproteome.2c00609 (2023).Gonnelli, G. et al. A decoy-free approach to the identification of peptides. J. Proteome Res. 14, 1792–1798 (2015).Article PubMed CAS Google Scholar Tran, N. H., Zhang, X., Xin, L., Shan, B. & Li, M. De novo peptide sequencing by deep learning. Proc. Natl Acad. Sci. USA 114, 8247–8252 (2017).Article PubMed PubMed Central CAS Google Scholar Noble, W. A multi-species benchmark for training and validating large scale mass spectrometry proteomics machine learning models. Zenodo https://doi.org/10.5281/zenodo.13685813 (2024).van Puyenbroeck, S. et al. Data for ‘Limitations of de novo peptide sequencing in resolving sequence ambiguity’. Figshare https://doi.org/10.6084/m9.figshare.30084355 (2026).Van Den Bossche, T. et al. The metaproteomics initiative: a coordinated approach for propelling the functional characterization of microbiomes. Microbiome 9, 243 (2021).Article PubMed PubMed Central Google Scholar Download referencesAcknowledgementsThis work has benefited from collaborations facilitated by the Metaproteomics Initiative (https://metaproteomics.org/), whose goals are to promote, improve and standardize metaproteomics74. The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.FundingT.C., R.B., S.v.P., L.M. and T.V.D.B. acknowledge funding from the Research Foundation Flanders (FWO) (grant nos. 12A8W25N, 12A6L24N, 1SA7M26N, 1286824N, G010023N and G0GDV23N). L.M. acknowledges the European Union’s Horizon Europe project BAXERNA 2.0 (grant agreement no. 101080544) and COMBINE (grant agreement no. 101191739). L.L.E. acknowledges support from the Research Council of Finland (341342 and 364700), the Sigrid Jusélius Foundation, Cancer Foundation Finland, Biocenter Finland and ELIXIR Finland. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or HORIZON-RIA. Neither the European Union nor the granting authority can be held responsible for them.Author informationAuthor notesThese authors jointly supervised this work: Tim Van Den Bossche, Tine Claeys.Authors and AffiliationsVIB-UGent Center for Medical Biotechnology, Ghent, BelgiumSam van Puyenbroeck, Tanja Holstein, Lennart Martens, Robbin Bouwmeester, Tim Van Den Bossche & Tine ClaeysDepartment of Biomolecular Medicine, Ghent University, Ghent, BelgiumSam van Puyenbroeck, Tanja Holstein, Lennart Martens, Robbin Bouwmeester, Tim Van Den Bossche & Tine ClaeysImage Analysis (ZKI-PH3), Centre for Artificial Intelligence in Public Health Research, Robert Koch Institute, Berlin, GermanyDenis BeslicTurku Bioscience Centre, University of Turku and Åbo Akademi University, Turku, FinlandTomi Suomi & Laura L. EloData Competence Centre (MF2), Robert Koch Institute, Berlin, GermanyTanja Holstein & Thilo MuthBioOrganic Mass Spectrometry Laboratory (LSMBO), IPHC UMR 7178, University of Strasbourg and CNRS, Strasbourg, FranceTanja Holstein & Lennart MartensInfrastructure Nationale de Protéomique ProFI–FR2048, Strasbourg, FranceTanja Holstein & Lennart MartensInstitute of Biomedicine, University of Turku, Turku, FinlandLaura L. EloAuthorsSam van PuyenbroeckView author publicationsSearch author on:PubMed Google ScholarDenis BeslicView author publicationsSearch author on:PubMed Google ScholarTomi SuomiView author publicationsSearch author on:PubMed Google ScholarTanja HolsteinView author publicationsSearch author on:PubMed Google ScholarThilo MuthView author publicationsSearch author on:PubMed Google ScholarLaura L. EloView author publicationsSearch author on:PubMed Google ScholarLennart MartensView author publicationsSearch author on:PubMed Google ScholarRobbin BouwmeesterView author publicationsSearch author on:PubMed Google ScholarTim Van Den BosscheView author publicationsSearch author on:PubMed Google ScholarTine ClaeysView author publicationsSearch author on:PubMed Google ScholarContributionsS.v.P. conceptualized the study, implemented the benchmarking pipeline, performed all analyses, generated the figures and wrote and revised the paper. D.B., T.S., T.H., T.M., L.L.E. and T.V.D.B. contributed to the conceptualization of the study, participated in the discussion and interpretation of the results, and revised the paper. R.B., T.V.D.B. and T.C. supervised the work. L.M. provided overall supervision and revised the paper. All authors read and approved the final paper.Corresponding authorCorrespondence to Lennart Martens.Ethics declarationsCompeting interestsThe authors declare no competing interests.Peer reviewPeer review informationNature Methods thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editor: Arunima Singh, in collaboration with the Nature Methods team. Peer reviewer reports are available.Additional informationPublisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.Extended dataExtended Data Fig. 1 Characteristics of the gold-standard PSMs for PXD028735-Orbitrap.(A-C) Counts of PSMs identified by Sage, MS-GF+, or both: (A) by rank, as defined by the MS2Rescore PSM score; (B) by the ppm mass error between the candidate and experimental m/z after correction for the closest isotopic deviation; and (C) by isotopic error. (D–F) Distribution of modifications in the gold standard, (D) by search engine, (E) by rank, and (F) by isotopic error.Source dataExtended Data Fig. 2 Composition of de novo predictions matching the gold standard on PXD028735-Orbitrap.For predictions from each de novo model that match a gold-standard PSM, the percentage that (A) correspond to gold-standard PSMs of rank 1–7, and (B) originate from each search engine. (C) Performance of the de novo models on PXD028735-Orbitrap, split by evaluation type.Source dataExtended Data Fig. 3 Pairwise overlap of spectra yielding correct and incorrect de novo PSMs.Pairwise overlap of spectra yielding (A) correct and (B) incorrect de novo PSMs. Each value is the proportion of predictions from the model on the vertical axis that are also predicted by the model on the horizontal axis (the number of predictions shared between the two models, divided by the total number of predictions from the model on the vertical axis).Source dataExtended Data Fig. 4 Example spectra for which all de novo models converge on a sequence not matching the gold standard but present in the database.Mirror plots of the experimental spectrum matched against the assigned gold-standard sequence (top) versus the same spectrum matched against the sequence predicted in common by all eight de novo models (bottom), for four representative spectra per sample from the PXD028735-Orbitrap dataset for (top to bottom) E. coli, Human, QC, and Yeast samples. In each case, none of the gold-standard PSMs match the shared de novo prediction, yet that predicted sequence is present in the sequence database, indicating either (i) an incorrect gold-standard identification, or (ii) a chimeric spectrum containing a co-fragmenting precursor not captured by the gold standard. Red lines indicate y-ions, and blue lines indicate b-ions. The peptidoform notation indicates the identified peptide with tags in brackets () having no b- and y-ion annotations.Extended Data Fig. 5 Effect of sequence refinement on de novo PSM correctness, per model.Sankey plots showing the effect of refinement by (left) InstaNovo+ and (right) Spectralis on PSM correctness, for de novo PSMs generated by (A) AdaNovo, (B) ContraNovo, (C) InstaNovo, (D) NovoB, (E) PepNet, (F) π-HelixNovo, (G) π-PrimeNovo, and (H) Casanovo. Red indicates a correct prediction changed to an incorrect one; green, a correct refinement; grey, a status-quo outcome (a correct prediction left unrefined, or an incorrect prediction refined to a different incorrect sequence); and blue, an incorrect prediction that the post-processor did not attempt to refine.Source dataExtended Data Fig. 6 Rank of the gold-standard PSM among candidates from different scoring methods.Rank of the gold-standard PSM among the candidates proposed by each scoring method, for spectra where the gold-standard PSM was (A) among the top 10 Casanovo candidates or (B) absent from them. Scoring methods: Casanovo, Casanovo without precursor-mass post-processing (Casanovo no PP), MS2Rescore-CS, MS2Rescore-d, Spectralis, and InstaNovo+.Source dataExtended Data Fig. 7 Score distributions of the gold-standard and MS2Rescore-selected candidate PSMs among spectra where the gold-standard PSM was among the top 10 Casanovo candidates.(A) Casanovo and (B) MS2Rescore-d score distributions for the gold-standard PSM (blue) and the top-ranked MS2Rescore-d PSM (orange). Difference between the two PSMs’ scores (gold-standard minus MS2Rescore-selected) for (C) Casanovo and (D) MS2Rescore-d scores.Source dataExtended Data Fig. 8 Most common tag pairs by de novo error category.Most common tag pairs within each error category, colored by their percentage of errors in that category: (A) tags not classified under B, C, or D, (B) amino acid permutations, (C) isobaric tags, and (D) single amino acid variants.Source dataExtended Data Fig. 9 Relationship between amino-acid reordering errors and their natural occurrence in the gold standard.Each point represents one pair of amino-acid substrings that are permutations of each other (for example, AG ↔ GA) and that was identified as a single-site transposition error (Damerau-Levenshtein distance < 5, edit distance ≠ 1, one differing alignment site) between a de novo prediction and the gold-standard peptide sequence, summed across all eight de novo models (AdaNovo, Casanovo, ContraNovo, InstaNovo, NovoB, PepNet, π-HelixNovo, and π-PrimeNovo) on PXD028735-Orbitrap. The x-axis shows how often this reordering occurred as a prediction error; the y-axis shows how often either substring of the pair occurs in the gold-standard peptide sequences of the dataset, independent of any prediction error. The 100 variants with the highest gold-standard occurrence are shown. A positive relationship would indicate that reordering errors occur more often simply because the underlying amino-acid combination is common in real peptides, rather than reflecting residue pairs that are inherently harder to order correctly from fragment ion evidence.Source dataExtended Data Fig. 10 Ambiguity and localization analysis of amino acid permutation and isobaric substitution errors.(A-C) Ambiguity map showing, for localized (A) permutation tags, (B) isobaric tags, and (C) non-isobaric tags, the proportion of ambiguous tags, defined as tags that lack all complementary ions in both PSMs or have at least one fragment ion at every cleavage site. (D–F) Location analysis of de novo PSMs with amino acid permutation and isobaric differences: (D) proportion of tags located at the C- or N-terminus, and the starting position of the sequence mismatch counted from the closest terminus, at the (E) C-terminus or (F) N-terminus.Source dataSupplementary informationSupplementary Information (download PDF )Supplementary Notes 1–3 and Figs. 1 and 2 with legends.Reporting Summary (download PDF )Peer Review File (download PDF )Source dataSource Data Fig. 1 (download XLSX )Source Data Fig. 2 (download XLSX )Source Data Fig. 3 (download XLSX )Source Data Fig. 4 (download XLSX )Source Data Fig. 5 (download XLSX )Source Data Extended Data Fig. 1 (download XLSX )Source Data Extended Data Fig. 2 (download XLSX )Source Data Extended Data Fig. 3 (download XLSX )Source Data Extended Data Fig. 5 (download XLSX )Source Data Extended Data Fig. 6 (download XLSX )Source Data Extended Data Fig. 7 (download XLSX )Source Data Extended Data Fig. 8 (download XLSX )Source Data Extended Data Fig. 9 (download XLSX )Source Data Extended Data Fig. 10 (download XLSX )Rights and permissionsOpen Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.Reprints and permissionsAbout this article