Transformer (AR)

37 methods · 2022–2026

Transformer (AR): Autoregressive transformer decoders, which emit the peptide one residue at a time conditioned on the spectrum and on the residues already chosen. The dominant architecture since Casanovo, and the one most benchmarked tools share.

Autoregressive transformer decoders, which emit the peptide one residue at a time conditioned on the spectrum and on the residues already chosen. The dominant architecture since Casanovo, and the one most benchmarked tools share.

The earliest of its 37 methods is Casanovo (2022); 36 more have followed.

Methods 37
Papers describing them 58
Authors 234
Active 2022-02-07 to 2026-09-24
Deep learning 37 of 37
Kinds algorithm (33), adjacent (2), downstream-application, post-processor
Acquisition DDA (32), DIA (3), both (2)

Methods (37)

Oldest first, by the paper that describes each one.

  • Casanovo (2022): First Transformer
  • DPST (2022): Amino-acid-aware transformer
  • PaSER Novor (2023): Real-time 4D-proteomics
  • BiATNovo (2023): Bidirectional self-attention
  • DpNovo (2023): Transformer + dynamic programming
  • π-HelixNovo (2023): Complementary spectra
  • InstaNovo (2023): Knapsack beam search
  • SeqNovo (2023): Seq2Seq for IoMT
  • DiaTrans (2023): Transformer for DIA
  • Casanovo-DB (2024): Database search scoring
  • NovoB (2024): Bidirectional decoding
  • AdaNovo (2024): Mutual info for PTMs
  • ContraNovo (2024): Contrastive learning
  • Cascadia (2024): Transformer for DIA
  • PowerNovo (2024): Transformer + BERT ensemble
  • TransNovo (2024): Transformer-based sequencing
  • π-xNovo (2024): Explainable AI sequencing
  • SearchNovo (2024): DB-search + de novo fusion
  • RankNovo (2024): Universal reranking
  • DIANovo (2024): Transformer-based de novo sequencer for DIA: disentangles coeluted/multiplexed precursor spectra. Presented in Zheng Ma’s arXiv preprint (2024) and his PhD thesis (2025).
  • ReNovo (2024): Retrieval-based sequencing
  • XA-Novo (2025): NAR knowledge distillation
  • TSARseqNovo (2025): Semi-autoregressive
  • PepGo (2025): Tree search-based decoding
  • DiNovo (2025): Mirror proteases + DL
  • Pairwise (2025): Mass difference attention
  • InstaNovo-P (2025): Phosphoproteomics
  • Casanovo Foundation (2025): Foundation model for tandem-MS proteomics: pre-trained the Casanovo spectrum encoder on 30M labelled spectra from MassIVE-KB, then reused the encoder off-the-shelf for downstream tasks (de novo sequencing, spectrum quality, chimericity, phosphorylation, glycosylation prediction). Successor in spirit to Casanovo v1/v2/v5.
  • LIPNovo (2025): Latent imputation
  • InstaNovo glycopeptide fine-tuning (2025): Adaptation of InstaNovo, a transformer de novo peptide sequencer, to glycoproteomics by fine-tuning on glycopeptide spectra. Learning from glyco spectra is measurable but limited, and every fine-tuning setting suffers catastrophic forgetting, losing accuracy on peptides the base model handled well; fine-tuning on unfiltered spectra that partly overlap the original training set gives the strongest learning signal but does not remove the effect. PCA of spectrum embeddings shows a large domain shift that explains the difficulty, and the study argues for treating overlapping glyco spectra as informative examples rather than outliers to discard, alongside multi-task and contrastive objectives and embedding transformations that pull glyco spectra toward the original distribution.
  • RNovA (2025): Zero-shot open PTM discovery
  • pUniFind (2025): Multimodal pre-trained transformer for mass spectra that unifies peptide-spectrum scoring and zero-shot de novo sequencing in a single model. Trained on >100M open-search-derived spectra; reports +60% PSMs over prior de novo methods with 1,300+ modifications supported, and a DL-based QC step that recovers 38.5% additional peptides.
  • Modanovo (2025): Unified PTM-aware model
  • CausalNovo (2026): Causality-informed framework
  • MemNovo (2026): Autoregressive transformer de novo sequencer with a memory mechanism that lets the decoder revisit the input spectrum at each step to balance error accumulation.
  • π-HelixNovo2 (2026): Successor to π-HelixNovo with an emphasis on availability: an online inference service alongside the model architecture refinement. Same Tsinghua / Pengcheng Lab / NCPSB collaboration as the original.
  • GyroNovo (2026): Attacks missing b- and y-ion fragments on two fronts. Rather than treating imputation as a fixed reconstruction task, it uses the decoder errors seen during training to steer the imputation objective toward the fragments that actually cause mistakes, and to build easy and hard augmented views of each spectrum so the decoder learns under varying spectral corruption. It also gives self-attention a mass-aware inductive bias, using rotary embeddings to encode pairwise mass differences between peaks. Inference needs no extra inputs or search. Reports about 9 points of peptide-level and 7 points of amino-acid-level precision over the previous best on NovoBench.

How they score

7 of the 37 have been run on denovo_benchmarks, which ranks 17 tools over 84 datasets. The family’s best median rank is 1.

  • InstaNovo: median peptide-level average precision 0.853, median rank 1 of 17
  • π-HelixNovo: median peptide-level average precision 0.833, median rank 5 of 17
  • ContraNovo: median peptide-level average precision 0.808, median rank 5 of 17
  • Pairwise: median peptide-level average precision 0.808, median rank 5 of 17
  • Casanovo: median peptide-level average precision 0.773, median rank 7 of 17
  • AdaNovo: median peptide-level average precision 0.514, median rank 12 of 17
  • BiATNovo: median peptide-level average precision 0.469, median rank 13 of 17

Read these next to the rest of the field, not on their own: what the numbers mean.

Applied in

Glycoproteomics

Papers describing them (58)

Authors (234)

Abdellah El Mekki, Alfred Nilsson, Ali Ghodsi, Amandla Mabona, Andreas Hougaard Laustsen, Andreas Pichlmair, Anna L. Kaysheva, Anne Ljungars, Annekatrine Kirketerp-Møller, Arthur T. Kopylov, Baozhen Shan, Bin Ma, Binyang Li, Bo Meng, Bo Wen, Boyan Sun, Bozhen Hu, Carlo F. Melendez, Chao Peng, Chen Yang, Cheng Chang, Cheng Lai, Cheng Tan, Chiara Francavilla, Chien-Ming Chen, Chris Hsu, Christian Nix, Christine C. Wu, Daniela Klaproth-Andrade, Denis V. Petrovskiy, Dennis Trede, Di Zhang, Dong An, Dongxin Lyu, Eric Stone, Erwin M. Schoof, Esperanza Rivera-de-Torre, Fan Liu, Fan Xu, Fangzheng Li, Fuchu He, George Rosenberger, Gwenneth Straub, Haipeng Wang, Haiteng Deng, Han Wen, Hao Chi, Hongxin Xiang, Hyunwoo Kim, Ida Sofie Goldschmidt, Isaac H.J. Houngue, Jacob H. Russell, Jakob Berg Jespersen, Jennifer Ferguson, Jeroen Van Goey, Jesper Lauridsen, Jiale Zhao, Jiangli Hu, Jiaqi Wei, Jiaxiang Ding, Jiaxing Dai, Jiaxing Qi, Jiazhen Chen, Jin Xiao, Jingbo Zhou, Jinghan Yang, Jingyi Wang, Jinze Huang, Joel Lapin, Jonathan Krieger, Julien Gagneur, Jun Xia, Justin Sanders, Kai Zou, Kaifei Wang, Karim Beguir, Karsten Kristiansen, Ke Wang, Kevin Eloff, Khandaker Asif, Kirill S. Nikolsky, Konstantinos Kalogeropoulos, Kristina A. Malsagova, Laks V.S. Lakshmanan, Lecheng Zhang, Lei Wang, Lei Xin, Lennart Martens, Lequan Yu, Leyuan Li, Li Kang, Lijin Yao, Ling Luo, Lingwen Xu, Linhai Xie, Liudmila I. Kulikova, Liyuan Pan, Lukas Käll, Maha Driss, Marcin J. Skwark, Marina Pominova, Mathias Wilhelm, Matin Ahmadi, Melih Yilmaz, Michael J. MacCoss, Michael Riffle, Ming Li, Mingjia Zhu, Mingjie Xie, Muhammad Abdul-Mageed, Nanqing Dong, Nanxi Yu, Nicholas M. Riley, Nicolas Lopez Carranza, Ning-Shao Xia, Oliver Morell, Pathmanaban Ramasamy, Paul Fullwood, Paul Rudnick, Pengzhi Mao, Ping Wu, Ping Xu, Piyu Zhou, Qian Zhao, Qianqiu Zhang, Qingfang Bu, Qixin Liu, Quan Yuan, Qunying Wang, Rachel Catzel, Ranfei Chen, Rich Johnson, Rongshan Yu, Rowan Nelson, Rui Zhang, Ruitao Wu, Ruixue Zhang, Sam P. B. van Beljouw, Sam van Puyenbroeck, Sangjeong Lee, Saru Kumari, Sewoong Oh, Shafin Rahman, Shaorong Chen, Sheng Xu, Shiva Ebrahimi, Shu Yang, Shujun Wang, Shuo Wang, Shuqi Lu, Siqi Liu, Siqi Sun, Siu-Ming Yiu, Siyu Wu, Sizhe Liu, Stan J. J. Brouns, Stan Z. Li, Tatiana V. Butkova, Tharan Srikumar, Thippa Reddy Gadekallu, Tiannan Guo, Tianze Ling, Timothy P. Jenkins, Tine Claeys, Tingpeng Yang, Ulrich auf dem Keller, Vahap Canbay, Valter Bergant, Varun Ananth, Vladimir R. Rudnev, Wadii Boulila, Wanli Ouyang, Wanyu Lin, Wassim Gabriel, Weijie Zhang, Weinan E, Wenbin Jiang, Wenjie Du, Wesley Williams, William E. Fondrie, William Stafford Noble, Wout Bittremieux, Xiang Fang, Xiang Zhang (Shanghai AI Lab), Xiansong Huang, Xiaohong Ji, Xiaohui Liang, Xiaoqing Chen, Xin Zhang, Xinhua Dai, Xinming Li, Xuan Guo, Xueli Peng, Yan Fu, Yan Yang, Yanchang Li, Yang Zhao, Yangtao Wu, Yanik Bruns, Yao Zhang, Yaoguang Wei, Yaoyu He, Yaping Peng, Ye Du, Yijie Qiu, Yiming Li, Yizhou Li, Yonghan Yu, Yonghong He, Yu Wang, Yuanliang Zhang, Yucheng Liao, Yue Liu, Yue Yu, Yueting Xiong, Yuling Chen, Yuqi Chang, Yuqiang Li, Yuxiaomei Liu, Zakir Hossain, Zeping Mao, Zexuan Yi, Zhangyang Gao, Zhendong Liang, Zheng Ma, Zhenjian Jiang, Zhenxin Fu, Zhi Jin, Zhiqiang Gao, Zhiyuan Cheng, Zhongzhi Luan, Zicheng Liu, Zijie Qiu, Zixuan Cao

Seen in the charts

Back to the full map

Back to top