Eight Peptide-Sequencing Models Hit the Same Wall: Missing Evidence.
A benchmark of eight deep-learning systems found that incomplete mass spectra create ambiguity no model can reliably infer away, and that common post-processing brought little improvement.
Direct answer
Why did the peptide-sequencing models remain ambiguous?
The benchmark found that many errors were driven by the spectrum itself: missing complementary ions, indistinguishable masses and noise can leave several peptide sequences consistent with the observation. The best individual model reached 50.4% accuracy against the study's database-search gold standard, while an idealised ensemble that counted any correct answer from eight models reached 64.2%. Rescoring, reranking and iterative refinement produced no or only modest gains.
A benchmark designed around disagreement
Sam van Puyenbroeck, Denis Beslic, Tomi Suomi and colleagues compared AdaNovo, Casanovo, ContraNovo, InstaNovo, NovoB, PepNet, π-HelixNovo and π-PrimeNovo on three public proteomics datasets. The Nature Methods article, published on 3 November 2026, also tested three post-processing approaches and analysed where model predictions diverged from database-derived peptide-spectrum matches.
The comparison retained multiple possible database identifications for chimeric spectra and counted a de novo result as correct if it matched any accepted candidate. That is an important evaluation choice: a single reported reference sequence can overstate certainty when the measurement itself permits alternatives.
The data ceiling
About a fifth of errors were localised to one site. In 68.5% of two-amino-acid transposition cases, the spectra lacked diagnostic fragment ions capable of distinguishing the alternatives; around half of isobaric tags were similarly indistinguishable. On average, 20% of wrong predictions were almost entirely different from the gold-standard sequence.
The models performed far better on complete amino-acid ladders and dropped sharply on timsTOF data outside the HCD-Orbitrap distribution that dominated training. This is not one model's bug. It is a combination of evidence loss, domain shift and confidence calibration.
Why uncertainty should be an output
A model that always emits one full sequence can look decisive even when several sequences fit the same peaks. Native scores may also be overconfident, and the target–decoy logic used in database searching does not transfer cleanly to a de novo system with a vast candidate space.
The authors argue for reliable sequence tags or partial matches with uncertain regions marked explicitly. That is the practical validation lesson: when the observation cannot resolve the claim, the system should expose the ambiguity rather than convert it into a precise-looking answer.
Primary sources and version boundary
Published figures and claims are reported from the linked source versions. The featured researchers and organisations did not commission or endorse this coverage. This article is independent Discovery News reporting and is not a Haivian audit result.
Independent validation
Need a computational claim examined before publication?
Haivian offers scoped research audits with explicit evidence, limitations and accountable human sign-off. Payment does not purchase a positive result or publication.
Explore research audits