EncNovo: De Novo Peptide Sequencing With an Encoder-Only Architectures
peer-reviewed · Rapid Communications in Mass Spectrometry · 2026
| Date | 2026-11-16 |
| Type | peer-reviewed |
| Venue | Rapid Communications in Mass Spectrometry |
| Publisher | Wiley |
| Contribution | algorithm |
| DOI | 10.1002/rcm.70211 |
| Venue 2-year citedness | 1.80 |
Abstract
Rationale De novo peptide sequencing enables interpretation of tandem mass spectrometry (MS/MS) data without relying on sequence databases and is valuable for analyzing proteins from unknown species. However, existing Transformer-based approaches may not fully exploit interactions among precursor information, spectral features, and sequence prediction tokens, potentially limiting sequencing performance. Methods We developed EncNovo, a nonautoregressive de novo peptide sequencing model based on a BERT-inspired encoder-only Transformer architecture. Precursor attributes, spectral peaks, and sequence prediction tokens are jointly encoded within a unified encoder, enabling continuous interaction among the three throughout the encoding process. A residual attention mechanism and a peak mass-difference inductive bias were further incorporated into the self-attention framework to encode amino acid residue mass relationships between spectral peaks. Results Experimental results on the nine-species benchmark dataset demonstrate that EncNovo improves peptide recall at 100% coverage by 5.10 and 4.18 percentage points on average compared with Casanovo and RefineNovo, respectively. The model also achieved higher amino acid recall and demonstrated faster convergence during training. Conclusions Experimental results demonstrate that the proposed encoder-only framework, combined with residual attention and a peak mass-difference inductive bias, effectively improves peptide recall and accelerates model convergence. Collectively, these advantages make EncNovo a reliable framework for de novo peptide sequencing in proteomics.
Methods and tools
- EncNovo: Non-autoregressive sequencer built on a BERT-style encoder-only Transformer: precursor attributes, spectral peaks and the sequence-prediction tokens are encoded together in one encoder, with residual attention and a peak mass-difference bias added to self-attention.
Evaluated on
- Nine-species benchmark (version not stated)