Transformer Architectures for De Novo Peptide Sequencing and Peptide Property Prediction in LC–MS/MS Proteomics
peer-reviewed · Artificial Intelligence in the Life Sciences · 2026
| Date | 2026-09-01 |
| Type | peer-reviewed |
| Venue | Artificial Intelligence in the Life Sciences |
| Publisher | Elsevier BV |
| Contribution | review |
| DOI | 10.1016/j.ailsci.2026.100184 |
| Citations (OpenAlex) | 0 |
Abstract
Transformer architectures are increasingly used to interpret peptide-centred liquid chromatography-tandem mass spectrometry (LC-MS/MS) data, yet reported gains remain difficult to compare because studies differ in tasks, training corpora, peptide-overlap controls, acquisition settings, metrics and deployment criteria. This scoping review mapped Transformer and Transformer-hybrid applications in de novo peptide sequencing and peptide property prediction, with attention to architecture, evaluation validity, determinants of performance, reproducibility and research-workflow readiness. PubMed/MEDLINE, Scopus and Web of Science Core Collection were searched for English-language studies published from January 2017 through June 2026. Of 439 records identified, 207 remained after deduplication, 49 full texts were assessed and 27 studies were included. Thirteen studies primarily addressed de novo sequencing, 12 addressed property prediction and two covered both. Encoder-decoder models predominated in spectrum-to-sequence generation, whereas encoder-style and hybrid architectures were common for fragment intensity, retention time, collision cross section and ion-mobility prediction. Performance depended on architecture, but also on training-data quality, acquisition context, peptide length, fragment completeness, post-translational-modification representation, domain adaptation, physicochemical constraints and decoding strategy. Cross-study comparison was limited by heterogeneous benchmarks, incomplete sequence-overlap auditing, scarce calibrated uncertainty and inconsistent reporting of compute and reproducibility. Property prediction offered the clearest route to reusable workflow integration, while de novo sequencing remained a confidence-sensitive complement to established identification methods. Future evaluations require versioned benchmarks, auditable data independence, broader distribution-shift testing and standardised reporting of uncertainty and efficiency.
Methods and tools
- Transformer architectures in LC-MS/MS proteomics (scoping review): Scoping review of Transformer and Transformer-hybrid models across de novo peptide sequencing and peptide property prediction. Searched PubMed/MEDLINE, Scopus and Web of Science for January 2017 to June 2026: 439 records, 207 after deduplication, 49 full texts assessed, 27 studies included, of which 13 were primarily de novo sequencing, 12 property prediction and 2 both. Finds encoder-decoder models dominant for spectrum-to-sequence generation and encoder-style or hybrid architectures common for fragment intensity, retention time, collision cross section and ion mobility. Its main argument is that reported gains are hard to compare at all, because benchmarks are heterogeneous, sequence-overlap auditing is incomplete, calibrated uncertainty is scarce and reporting is inconsistent.