π-PrimeNovo

algorithm · Transformer (NAR)

π-PrimeNovo: algorithm · Transformer (NAR). NAR Transformer (CTC)

NAR Transformer (CTC)

Kind algorithm
Deep learning yes
Acquisition DDA
Also known as PrimeNovo
Family Transformer (NAR)

Code

Live stars, open issues and last-push figures are on the Code activity chart.

Checkpoints

Version Trained on Host Licence Size Checked Backup
MassIVE MassIVE-KB Google Drive MIT 387 MB verified 2026-10-02 copy
phosphorylation fine-tuned from model_massive.ckpt on the 2020-Cell-LUAD dataset; predicts one extra token B for Phosphorylation (+79.97) Google Drive MIT 387 MB verified 2026-10-02 copy

A host marked archival has a DOI and keeps what it is given. The others can move or disappear, which is why they are checked rather than merely listed. verified means the bytes were fetched and hashed on the date shown; live means only that the host answered when last asked. Where a copy is linked, it is a backup of someone else’s weights kept in case the original link goes stale; the original is the link to cite and to prefer.

Benchmarks

  • denovo_benchmarks: median peptide-level average precision 0.802 over 86 datasets, median rank 5 of 14 (version bm-1.0.0).
  • ProteoBench, on the nine-species benchmark, ProteoBench selection: peptide-level AUC 0.773; precision 0.682 at 88% coverage; amino-acid AUC 0.839 (version 1.0, beam search + PMC [1 Da], submitted 2026-08-07).

Both are mass-based matches on the tool’s most recent run. What these numbers mean.

Reported comparisons (5)

The comparison tables this method’s own papers print, standardised: every value on a 0-1 scale, methods down the side, the measure and then the species across. These are numbers papers report about themselves and their baselines. They are not a leaderboard, and they do not compare across tables: each was produced by a different group, on the dataset named in its corner, with each baseline either retrained, run from released weights or quoted from another paper. Where the paper says which, it follows the method’s name (hover it for the sentence); most papers do not say. Bold is the best value in a column and underline the runner-up, our ranking rather than the paper’s own marks.

SupplementaryTable5

π-PrimeNovo: an accurate and efficient non-autoregressive deep learning model for de novo peptide sequencing, page 33: Peptide recall on the revised nine-species benchmark dataset. Note: the bold text indicates the highest performance in each row.

Nine-species benchmark
revised (main) 		Peptide recall
Apis mellifera Bacillus subtilis Candidatus Thiodiazotropha endoloripes Homo sapiens Methanosarcina mazei Mus musculus Saccharomyces cerevisiae Solanum lycopersicum Vigna mungo Average
Casanovo 0.3578 0.4637 0.3055 0.4468 0.4927 0.4063 0.4657 0.4354 0.3977 0.4191
Casanovo V2 0.57 0.71 0.51 0.65 0.68 0.52 0.73 0.69 0.7 0.64
π-PrimeNovo 0.70 0.80 0.59 0.75 0.78 0.60 0.83 0.78 0.84 0.74

Supplementary Table 3

π-PrimeNovo: an accurate and efficient non-autoregressive deep learning model for de novo peptide sequencing, page 26: Peptide recall on the nine-species benchmark dataset. Note: the bold text indicates the highest performance in each row.

The same table is also printed in π-PrimeNovo: an accurate and efficient non-autoregressive deep learning model for de novo peptide sequencing (Supplementary Table 3, page 31).

Nine-species benchmark Peptide recall
Apis mellifera Bacillus subtilis Candidatus Thiodiazotropha endoloripes Homo sapiens Methanosarcina mazei Mus musculus Saccharomyces cerevisiae Solanum lycopersicum Vigna mungo Average
PEAKS 0.29 0.39 0.20 0.28 0.36 0.20 0.43 0.40 0.36 0.32
PepNet 0.28 0.42 0.24 0.23 0.38 0.28 0.37 0.43 0.32 0.33
DeepNovo 0.33 0.45 0.25 0.29 0.42 0.29 0.46 0.45 0.44 0.38
PointNovo 0.40 0.52 0.30 0.35 0.48 0.36 0.53 0.51 0.51 0.44
Casanovo 0.41 0.54 0.33 0.34 0.48 0.43 0.49 0.52 0.51 0.45
Casanovo V2 0.49 0.62 0.45 0.45 0.56 0.48 0.60 0.62 0.59 0.54
π-PrimeNovo 0.60 0.72 0.53 0.57 0.65 0.57 0.70 0.70 0.70 0.64

SupplementaryTable4

π-PrimeNovo: an accurate and efficient non-autoregressive deep learning model for de novo peptide sequencing, page 27: Amino acid precision on the nine-species benchmark dataset. Note: the bold text indicates the highest performance in each row.

The same table is also printed in π-PrimeNovo: an accurate and efficient non-autoregressive deep learning model for de novo peptide sequencing (SupplementaryTable4, page 33).

Nine-species benchmark Amino acid precision
Apis mellifera Bacillus subtilis Candidatus Thiodiazotropha endoloripes Homo sapiens Methanosarcina mazei Mus musculus Saccharomyces cerevisiae Solanum lycopersicum Vigna mungo Average
PEAKS 0.63 0.72 0.59 0.64 0.67 0.60 0.75 0.73 0.64 0.66
DeepNovo 0.63 0.74 0.60 0.61 0.69 0.62 0.75 0.73 0.68 0.67
PointNovo 0.64 0.77 0.59 0.61 0.71 0.63 0.78 0.73 0.73 0.69
Casanovo 0.63 0.75 0.60 0.59 0.68 0.69 0.68 0.72 0.67 0.67
Casanovo V2 0.71 0.79 0.68 0.68 0.76 0.76 0.75 0.79 0.75 0.74
π-PrimeNovo 0.76 0.84 0.72 0.72 0.80 0.78 0.80 0.82 0.82 0.79

SupplementaryTable6

π-PrimeNovo: an accurate and efficient non-autoregressive deep learning model for de novo peptide sequencing, page 29: The average performance of PrimeNovo compared to Casanovo and Casanovo V2 across four distinct large-scale MS/MS datasets.

The same table is also printed in π-PrimeNovo: an accurate and efficient non-autoregressive deep learning model for de novo peptide sequencing (SupplementaryTable6, page 35).

one per row (Data availability) Amino acid precision Peptide recall
HCC IgG1-Human-HC PT Three-species HCC IgG1-Human-HC PT Three-species
Casanovo 0.0908 0.34465 0.4443 0.5633 0.0001 0.1154 0.2873 0.5037
Casanovo V2 0.4968 0.605917000 0.6284 0.8167 0.161 0.4065 0.4543 0.6873
π-PrimeNovo 0.5793 0.693133000 0.7307 0.8767 0.3817 0.5416 0.5857 0.7867

SupplementaryTable7

π-PrimeNovo: an accurate and efficient non-autoregressive deep learning model for de novo peptide sequencing, page 30: The performance of PrimeNovo compared to Casanovo and Casanovo V2 on six different proteolytic enzymes in the IgG1-Human-HC dataset.

The same table is also printed in π-PrimeNovo: an accurate and efficient non-autoregressive deep learning model for de novo peptide sequencing (SupplementaryTable7, page 36).

Monoclonal antibody de novo assembly Amino acid precision Peptide recall
HC AspN HC Chymotrypsin HC GluC HC LysC HC Proteinase K HC Trypsin HC AspN HC Chymotrypsin HC GluC HC LysC HC Proteinase K HC Trypsin
Casanovo 0.3089 0.2114 0.2878 0.4359 0.3844 0.4395 0.0726 0.0298 0.0686 0.2108 0.089 0.1703
Casanovo V2 0.5236 0.4794 0.5214 0.7002 0.6735 0.7374 0.2611 0.2493 0.3133 0.5224 0.4568 0.5483
π-PrimeNovo 0.6059 0.6227 0.6427 0.749 0.7567 0.7818 0.3593 0.446 0.4625 0.6204 0.5972 0.6649

Papers describing it (2)

Paper using it

An application or evaluation that ran this method. It is not counted among its authors below.

Authors (18)

Xiang Zhang (Shanghai AI Lab), Tianze Ling, Zhi Jin, Sheng Xu, Zhiqiang Gao, Boyan Sun, Zijie Qiu, Nanqing Dong, Guangshuai Wang, Guibin Wang, Leyuan Li, Muhammad Abdul-Mageed, Laks V.S. Lakshmanan, Wanli Ouyang, Cheng Chang, Siqi Sun, Jiaqi Wei, Fuchu He

Seen in the charts

Back to the full map

Back to top