PaperlayerView exact revision on arXiv

Source-linked AI summary

GyroNovo: Error-Guided Fragment Imputation with Mass-Aware Attention for \textit{De Novo} Peptide Sequencing

Abdellah El Mekki, Laks V. S. Lakshmanan, Muhammad Abdul-Mageed

arXiv:2609.30542v1cs.LG

TL;DR

De novo peptide sequencing is difficult because spectra can be sparse, noisy, and incomplete, while existing imputation does not prioritize fragments linked to decoder errors and peak representations leave mass differences implicit. GyroNovo guides imputation and augmented spectrum views with decoder errors and adds mass-aware rotary attention. On NovoBench, it reports gains of about 9 percentage points in peptide-level precision and 7 percentage points in amino-acid-level precision over the state-of-the-art baseline.

  • Problem

    Sparse, noisy, incomplete spectra leave informative fragments unobserved, while existing imputation treats reconstruction as fixed rather than prioritizing fragments relevant to decoder errors, and peak representations do not explicitly model mass differences.

    Source: Introduction · p. 1

  • Method

    GyroNovo uses decoder errors to prioritize cleavage-local fragment imputation and construct easy and hard spectrum views, while continuous mass-based rotary attention encodes relative peak mass differences.

    Source: Introduction · p. 2

  • Results

    About 9 percentage points in peptide-level precision and 7 percentage points in amino-acid-level precision over the state-of-the-art baseline are reported on NovoBench.

    Source: Introduction · p. 2

  • Takeaways & Limitations

    The results indicate that aligning fragment representation learning with decoder uncertainty and adding a relative-mass inductive bias is a more effective approach for learning from sparse and noisy tandem mass spectra.

    Source: Conclusion · p. 7

    Multi-view training entails greater computation per epoch.

    Source: Time and Memory Overhead · p. 7

Abstract

from arXiv · show

De novo peptide sequencing from tandem mass spectra is essential for identifying peptides without relying on reference databases. Despite advances in deep learning, accurate sequencing remains challenging because experimental spectra are often sparse, noisy, and incomplete, leaving informative b- and y-ion fragments unobserved. Existing methods attempt to recover this missing evidence via latent-space imputation before autoregressive decoding. However, they typically treat imputation as a fixed reconstruction task, without considering which missing fragments are most relevant to decoder errors. Moreover, existing peak representations do not explicitly model mass differences between peaks, despite their fundamental importance. We introduce GyroNovo, a framework with two main contributions. First, we use decoder errors observed during training to adapt the imputation objective, prioritizing fragments associated with frequent decoding errors. We further use the decoder error distribution to construct easy and hard augmented views of each spectrum, enabling the decoder to learn under varying degrees of spectral corruption and missing-fragment severity. Second, we introduce a mass-aware inductive bias into self-attention by using rotary embeddings to encode pairwise mass differences between spectral peaks. Together, these components align missing-fragment recovery with decoder behavior while explicitly incorporating the mass relationships that underlie peptide fragmentation. At inference time, GyroNovo retains a standard encoder-imputer-decoder architecture and requires neither additional inputs nor auxiliary search procedures. Experiments on NovoBench show gains of about 9 percentage points in peptide-level precision and 7 percentage points in amino-acid-level precision over the state-of-the-art baseline. Code: https://github.com/UBC-NLP/gyronovo.

Introduction

De novo peptide sequencing infers peptide sequences directly from tandem mass spectra, but sparse, noisy spectra and missing fragmentation evidence continue to challenge accurate decoding. GyroNovo uses decoder errors to guide fragment imputation and spectrum augmentation, and adds rotary attention based on mass differences.

Introduction

  • Missing fragmentation evidence can leave autoregressive decoders predicting residues with weak or absent cleavage evidence.LIPNovo imputes theoretical fragment representations, but its supervision does not prioritize fragments most useful for correcting current decoding errors.Source: Introduction · p. 1
  • Spectra encode chemically meaningful mass separations, but standard Transformer encoders use absolute m/z and intensity and leave relative relationships implicit.The passage identifies residue transitions, fragmentation ladders, complementary ions, and neutral losses as sources of characteristic separations.Source: Introduction · p. 1
  • GyroNovo maps token-level decoder errors to nearby theoretical fragment cleavages, assigning greater imputation weight to evidence associated with difficult residue decisions.The same feedback constructs easier and harder peptide-conditioned spectrum views by selectively modifying theoretical fragments while preserving observed peaks.Source: Introduction · p. 2
  • About 9- and 7-percentage-point gains in peptide- and amino-acid-level precision, respectively, are reported for GyroNovo on NovoBench.The evaluation is described as state-of-the-art in the introduction.Source: Introduction · p. 2
  • GyroNovo combines decoder-conditioned imputation, observed-anchor peptide-conditioned spectral views, and continuous mass-based rotary self-attention.Its stated contributions include imputation and decoding supervision for the augmented views and relative mass differences in the spectrum encoder.Source: Introduction · p. 2

Related Work

De novo sequencing approaches have developed from spectrum-graph search and handcrafted scoring to neural models that predict residues from spectra. Transformer methods decode sequences from peak representations, while latent imputation and mass-difference attention address missing fragments and relative peak structure.

De Novo Peptide Sequencing

  • Classical methods use spectrum-graph path finding or dynamic programming, while neural methods progress from next-residue prediction to peak-set and graph representations.Sequencing accuracy remains sensitive to noise and missing fragment ions.Source: De Novo Peptide Sequencing · p. 2
  • LIPNovo differs from methods that immediately decode observed spectra by imputing latent representations of theoretical fragment peaks before prediction.This design directly targets incomplete fragmentation.Source: De Novo Peptide Sequencing · p. 2

Transformer-based De Novo Peptide Sequencing

  • Transformer models translate spectra into peptide sequences, and conventional peak embeddings encode absolute m/z while leaving chemically meaningful mass differences implicit.Recent methods incorporate pairwise mass differences through attention biases.Source: Transformer-based De Novo Peptide… · p. 2

Proposed Method

GyroNovo builds on latent fragment imputation to address incomplete spectra with decoder-guided training and continuous-mass attention. Training uses error-conditioned spectrum views; inference predicts missing-fragment representations from the observed spectrum.

Problem Formulation

  • The task is to recover a peptide sequence from a tandem mass spectrum without a protein database, despite absent, weak, or noisy informative b- and y-ions.The model represents the peptide sequence autoregressively, conditioning each amino acid on the preceding sequence, spectrum, and precursor information.Source: Problem Formulation · p. 2

Latent Imputation Backbone

  • The LIPNovo backbone encodes observed peaks, uses learned queries to predict candidate missing-fragment representations and confidence scores, then decodes from confident imputed and observed representations.Training constructs theoretical b- and y-ion targets and aligns predicted queries with them; at inference, missing-fragment representations are predicted directly from the observed spectrum.Source: Latent Imputation Backbone · p. 3

Error-Guided Fragment Imputation Supervision

  • Decoder errors reweight theoretical-fragment imputation toward cleavages near poorly predicted residues and guide easier and harder augmented spectrum views.The easy view strengthens difficult fragment evidence, while the hard view weakens or removes it; both views receive imputation and sequence supervision during training.Source: Imputation Loss Re-Weighting · p. 4, Observed Mass Spectrum Augmentations · p. 4

Continuous Mass Rotary Attention

  • Continuous-m/z rotary attention makes spectrum self-attention depend on relative peak-mass displacement, reflecting informative residue-dependent gaps and fragmentation ladders.The method retains baseline peak embeddings and applies rotations to query and key vectors based on each peak’s continuous m/z coordinate.Source: Continuous Mass Rotary Attention · pp. 4–5

Training Objective and Inference

  • Training combines peptide decoding and fragment imputation on the observed spectrum, the theoretical spectrum, and easy and hard augmented views; inference uses observed-spectrum encoding, imputation, and peptide decoding.At inference, only the observed spectrum and precursor metadata are required; the theoretical spectrum, decoder-conditioned weights, and augmented views are discarded.Source: Training Objective and Inference · p. 5

Experiments

Experiments use three peptide-spectrum datasets, compare GyroNovo with established sequencing methods, and assess performance at amino-acid, PTM, and complete-peptide levels.

Datasets

  • Testing uses Nine-species, Seven-species, and human-derived HC-PT spectra, with yeast held out in the first two datasets.HC-PT includes canonical proteins and isoforms, alternative-protease peptides, and HLA-associated peptides.Source: Datasets · p. 5

Baselines

  • GyroNovo is compared with six methods; reproduced LIPNovo results serve as the baseline for a fair comparison.The baselines include DeepNovo, PointNovo, CasaNovo, AdaNovo, LIPNovo, and concurrent LIPNovo+.Source: Baselines · p. 5

Evaluation Metrics

  • Evaluation reports amino-acid and PTM precision and recall, exact-match peptide precision, and precision–recall AUC.Amino-acid correctness uses residue-mass and prefix-or-suffix-mass tolerances.Source: Evaluation Metrics · p. 5

Implementation Details

  • The model follows Du et al.’s configuration, with nine-layer encoder and decoder, a three-layer imputer, and training on four A100 GPUs for up to 30 epochs.Inference uses beam search with width five.Source: Implementation Details · pp. 5–6

Results and Analysis

GyroNovo outperforms the reproduced LIPNovo baseline across datasets and evaluation metrics, with gains also over concurrent LIPNovo+. Ablations and neighbor-retrieval analysis report gains from augmented views and mass-aware rotary attention, while efficiency comparisons quantify its overhead and memory use.

Main Results

  • GyroNovo leads across datasets: amino-acid precision exceeds LIPNovo by 6.4 points, peptide precision by 8.7 points, and PTM precision by 5.6 points.The paper reports highest amino-acid precision and recall on all datasets and improvements across peptide-level metrics.Source: Main Results · p. 6, Table 1 · p. 7, Table 2 · p. 7
  • Against LIPNovo+, precision gains average 1.20 amino-acid, 2.93 peptide, and 2.10 PTM percentage points across all three datasets.GyroNovo is reported to outperform LIPNovo+ across all metrics on all three datasets.Source: Main Results · p. 6, Table 1 · p. 7, Table 2 · p. 7

Ablation Study and Analysis

  • On Nine-species, augmented views give the largest increments (+3.1 amino-acid precision and recall; +5.3 peptide precision), while rotary adds +1.9, +1.9, and +3.5 points, respectively.The complete model improves over LIPNovo by 5.4 points in amino-acid precision, 5.2 in recall, and 9.1 in peptide precision.Source: Component Ablation · p. 6, Table 3 · p. 7
  • Hit@1 rose approximately 13% to 19% and Hit@20 approximately 58% to 65% with rotary; shuffled masses brought retrieval near random and below vanilla attention.The task measures retrieval of adjacent b- or y-ion ladder neighbors among attention-ranked peaks.Source: Figure 2 · p. 6, Mechanistic Analysis of Mass Rotary… · p. 6
  • At inference, GyroNovo and GyroNovo-Pairwise both maintain baseline throughput, while GyroNovo uses 1.01× baseline memory compared with 3.86× for GyroNovo-Pairwise.During training, GyroNovo uses 2.08× baseline memory; GyroNovo-Pairwise has 0.27× baseline throughput and uses 4.02× baseline memory.Source: Time and Memory Overhead · p. 7, Table 4 · p. 7
  • Both LIPNovo and GyroNovo stopped at epoch 13, but GyroNovo’s multi-view training entails greater computation per epoch.The maximum training budget was 30 epochs.Source: Time and Memory Overhead · p. 7

Conclusion

GyroNovo targets missing-fragment supervision and relative-mass modeling with decoder-guided imputation and continuous mass-based rotary attention. It retains latent imputation at inference and reports that these components provide a more effective approach to sparse, noisy spectra.

Conclusion

  • Decoder-guided learning prioritizes fragments relevant to unresolved residues and guides easy and hard spectrum augmentations.The augmentations provide clearer evidence for difficult residues while improving robustness to incomplete fragmentation ladders.Source: Conclusion · p. 7
  • Continuous mass-based rotary attention explicitly incorporates relative mass differences into spectrum self-attention without additional learned positional parameters.Source: Conclusion · p. 7
  • At inference, GyroNovo retains latent imputation and requires neither theoretical spectra nor an augmentation procedure.Source: Conclusion · p. 7
  • The authors conclude that decoder-uncertainty alignment and relative-mass inductive bias offer a more effective approach to learning from sparse, noisy spectra.They identify task-aware spectrum representation learning and physically informed attention as directions for further investigation.Source: Conclusion · pp. 7–8

A Decoder-Conditioned Spectrum Augmentation

The augmentation procedure uses decoder-derived difficulty to focus fragment targets and construct easy and hard spectral views. It varies corruption over training while retaining shared peptide supervision across the views.

A Decoder-Conditioned Spectrum Augmentation

  • Decoder-derived token losses condition fragment weights, with difficult residues emphasizing nearby cleavage targets.The focus is normalized within each spectrum, and a maximum keeps one difficult residue from being diluted by easier neighbors.Source: A Decoder-Conditioned Spectrum… · p. 10
  • Corruption strength interpolates from zero toward a maximum across peak missingness, ion dropout, ladder gaps, intensity and m/z jitter, structured noise, and PTM visibility reduction.Source: A Decoder-Conditioned Spectrum… · p. 10
  • Easy and hard views use different operator strengths, while fixed retention, insertion, and decoder-focus adjustments remain active throughout training.PTM-containing peptides receive a 1.2 multiplier on the PTM component; operator strengths may act as probabilities, Gaussian standard deviations, or Poisson count parameters.Source: A Decoder-Conditioned Spectrum… · p. 10
  • The procedure matches theoretical fragments to observed peaks within 0.5 m/z, then retains matched and unannotated peaks with probabilities 0.90 and 0.70, respectively.Retained peaks keep measured m/z; maximally focused peaks are strengthened by 20% in the easy view or attenuated by 20% in the hard view.Source: A Decoder-Conditioned Spectrum… · p. 10
  • A theoretical fragment can be inserted when it lacks an observed match or its matched peak was removed, restoring missing or dropped evidence.Inserted fragment intensities and locations are perturbed.Source: A Decoder-Conditioned Spectrum… · pp. 10–11
  • Ladder-gap events select consecutive fragment indices, with a configured maximum run of three applied to both ion ladders.Hard views remove observed peaks and disable insertion at selected indices; easy views apply focus-based exceptions.Source: A Decoder-Conditioned Spectrum… · p. 11
  • Structured-noise peaks model water or ammonia losses with m/z jitter; the assembled spectrum is filtered, capped at 150 peaks, and sorted by m/z.If no peak survives, the strongest baseline theoretical fragment is retained.Source: A Decoder-Conditioned Spectrum… · p. 11
  • Both views retain precursor information and peptide labels, use the same encoder-imputer-decoder, and receive equally weighted imputation and sequence losses.Augmentation hyperparameters were selected by random search, with perturbation types motivated by validation-set error analysis.Source: A Decoder-Conditioned Spectrum… · p. 11
Loading 2609.30542v1…