henhaobio
Sign in

bioRxivBioinformaticsPreprint

De novo mass spectrometry peptide sequencing with a transformer model

Melih YilmaziD, William E. FondrieiD, Wout BittremieuxiD, Sewoong OhiD, William Stafford NobleiD

PDFDOICode
In-depth full-text digest available · Jump to digest ↓

Abstract

Tandem mass spectrometry is the only high-throughput method for analyzing the protein content of complex biological samples and is thus the primary technology driving the growth of the field of proteomics. A key outstanding challenge in this field involves identifying the sequence of amino acids--the peptide--responsible for generating each observed spectrum, without making use of prior knowledge in the form of a peptide sequence database. Although various machine learning methods have been developed to address this de novo sequencing problem, challenges that arise when modeling tandem mass spectra have led to complex models that combine multiple neural networks and post-processing steps. We propose a simple yet powerful method for de novo peptide sequencing, Casanovo, that uses a transformer framework to map directly from a sequence of observed peaks (a mass spectrum) to a sequence of amino acids (a peptide). Our experiments show that Casanovo achieves state-of-the-art performance on a benchmark dataset using a standard cross-species evaluation framework which involves testing with spectra with never-before-seen peptide labels. Casanovo not only achieves superior performance but does so at a fraction of the model complexity and inference time required by other methods.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (18 Jun 2022) · CC BY 4.0

Casanovo, a transformer encoder-decoder, translates unbinned tandem mass spectrum peaks into peptide sequences, beating DeepNovo and PointNovo on peptide-level precision across nine held-out species while running faster than both.

De novo peptide sequencing assigns peptides to tandem mass spectra without a sequence database, which matters for variant peptides, immunopeptidomics, antibody sequencing and metaproteomics. Earlier deep learning tools such as DeepNovo and PointNovo combine convolutional and recurrent networks, mostly bin the mass-to-charge (m/z) axis and rely on dynamic-programming post-processing; Casanovo replaces this with one transformer over unbinned peaks plus a simple precursor-mass filter.

Key results

  1. ShownAcross all nine held-out species, Casanovo has higher peptide-level precision than DeepNovo and PointNovo. The margin is large when its precursor filter is on and small when it must predict every spectrum.

    Mean peptide-level precision gain of 0.373 over DeepNovo and 0.310 over PointNovo at a mean Casanovo coverage of 0.60 (Table 2). With the filter off (coverage 1.0), the gains fall to 0.076 and 0.013. DeepNovo was run from its pre-trained weights; the PointNovo numbers are its published results. · 5.1 Casanovo outperforms state-of-the-art…; Tables

  2. ShownAt the peptide level, the advantage over DeepNovo holds over a range of confidence thresholds, not only at the filter's operating point. At the amino-acid level it is clearest in the high-precision part of the curves.

    Area under the peptide precision-coverage curve (AUC) is higher than DeepNovo's by 0.131 on average, and Casanovo's AUC exceeds DeepNovo's in 8 of 9 species. The filter's operating point lies above DeepNovo's curve in all 9 species. Amino-acid curves favour Casanovo mostly in the high-precision region. Error bars for human and mouse come from five random initialisations. · 5.1 Casanovo outperforms state-of-the-art…; Figure legends

  3. ShownCasanovo has fewer parameters than DeepNovo and runs inference faster than both baselines. The authors argue it is fast enough to sequence spectra in real time as the instrument acquires them.

    47M parameters vs 86M for DeepNovo (not reported for PointNovo). Inference runs at 119 spectra/s on an RTX 2080, vs 36 spectra/s for DeepNovo and PointNovo's reported 20 spectra/s on an RTX 2080 Ti. Instruments produce MS2 spectra at about 20 to 40 Hz. · 5.1 Casanovo outperforms state-of-the-art…; 2 Background: De novo peptide sequencing; 6 Discussion

  4. SuggestedOn yeast, the simple precursor m/z filter beats both no post-processing and DeepNovo-style knapsack dynamic programming. The knapsack step adds little peptide precision and lowers amino-acid precision; a qualitative review suggests early decoding errors make it prune correct later residues.

    Peptide precision at full coverage rises only from 0.561 to 0.565 with knapsack, which also lowers AUC (the DeepNovo paper reported a 12.4% drop when this step was removed). Amino-acid precision: 0.769 with no post-processing, 0.726 with knapsack, and 0.965 with the m/z filter (0.576 when forced to cover all spectra). · 5.2 Precursor m/z post-processing; Figure legends

  5. ClaimedFilling the spectra that Casanovo's filter rejects with DeepNovo predictions is reported to improve precision further and to surpass PointNovo.

    The Discussion reports up to 10% higher peptide precision than Casanovo alone, and better results than PointNovo on all evaluation metrics across species. No table, figure or per-species numbers are given for this combination. · 6 Discussion

Methods, data and reproducibility

A nine-layer transformer encoder-decoder (~47M parameters) embeds each peak as sinusoidal m/z plus linear intensity, then greedily decodes amino acids (24-token vocabulary) conditioned on precursor mass and charge; outputs whose mass misses the precursor beyond a ppm tolerance are discarded. It is trained with cross-entropy on database-search labels and tested by leave-one-species-out cross-validation.

  • Nine-species benchmark: about 1.5 million MS2 spectra from nine experiments on one instrument, labelled by database search at 1% false discovery rate (FDR); about 300,000 unique peptides · nine species (human, mouse and yeast named in the text) · n = about 1.5 million spectra; ~300,000 unique peptides; 9 species
  • Human held-out fold, used to show train/test separation: unique peptide labels compared with those of the other eight species · human · n = ~26,000 unique peptides (7% overlap with ~250,000 from other species)
Methods & tools
Casanovo, DeepNovo, PointNovo, spectrum utils
Assay conditions & compute
Training on 2 RTX 2080 GPUs for 30 epochs, roughly two days; inference at 119 spectra/s on an RTX 2080.
Reproducibility gaps
The text gives no link or accession for the benchmark spectra and states no software versions or environment files. PointNovo could not be rerun (its weights and benchmark predictions were unavailable), so its numbers come from its publication.
Limitations (4)
  • Even with the precursor filter, many test spectra get no plausible prediction (mean coverage 0.60). The authors leave recovering these to future work and suggest using precursor mass as a stronger prior (mass-constrained beam search or a mass-penalising loss); neither is tested here. · stated by the authors
  • The precursor m/z filter trades away precision at full coverage: on yeast, amino-acid precision drops to 0.576 when all spectra must be predicted, vs 0.769 with no post-processing. · stated by the authors
  • PointNovo is compared only through its published numbers, its speed was measured on a different GPU (RTX 2080 Ti), and its parameter count is unknown. · noted in this summary
  • All labels come from 1% FDR database search on one benchmark acquired on a single instrument, so performance on spectra that database search cannot identify is untested. The post-processing and embedding/loss ablations use only the yeast fold. · noted in this summary

Read this if you build or use deep learning models for database-free peptide identification from tandem mass spectra, or want a worked example of treating spectrum-to-peptide as transformer sequence translation.

Citations
55

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISYilmaz '24Lapin '25Eloff '25Chang '25Ananth '24Klaproth-An '23
92% semantic matchBioinformatics
Melih Yilmaz, William E. Fondrie, Wout Bittremieux+5 more·Nature Communications·Jul 30, 2024· 119 citations