XuanjiNovo

algorithm · Transformer (NAR)

XuanjiNovo: algorithm · Transformer (NAR). Billion-scale pretraining

Billion-scale pretraining

Kind algorithm
Deep learning yes
Acquisition DDA
Family Transformer (NAR)

Code

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

Data

What the papers describing this method trained it on and reported its results on, as each paper states it. A version is named only where the paper names one.

Trained on

Evaluated on

Deposited

Also used, purpose not recorded

Reported comparisons (4)

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.

Supplementary Fig. 3a

MassNet: billion-scale AI-ready mass spectral corpus empowering scalable deep learning in proteomics, page 5: Evaluation of de novo peptide sequencing performance on other benchmark datasets. (A) Peptide recall across nine species using seven models: InstaNovo+, π-HelixNovo, ContraNovo, CasaNovo V2, π-PrimeNovo, XuanjiNovo-100M, and XuanjiNovo-130M. (B) Peptide recall on the HCC dataset. (C) Peptide recall on the PT dataset. The π-HelixNovo model is publicly available and was trained on datasets derived from eight species, excluding human. XuanjiNovo-100M: base model; XuanjiNovo-130M: fine-tuned on 30M PSMs from MassIVE.

Nine-species benchmark
original (DeepNovo, 2017) 		Peptide recall
Bacillus subtilis Candidatus Thiodiazotropha endoloripes Apis mellifera Homo sapiens Methanosarcina mazei Mus musculus Vigna mungo Solanum lycopersicum Saccharomyces cerevisiae
π-HelixNovo · released 0.596 0.388 0.473 0.392 0.56 0.483 0.623 0.56 0.568
InstaNovo+ 0.576 0.417 0.467 0.456 0.528 0.471 0.565 0.585 0.56
Casanovo V2 0.622 0.446 0.499 0.455 0.565 0.486 0.594 0.624 0.598
ContraNovo 0.688 0.487 0.576 0.622 0.63 0.566 0.677 0.672 0.674
π-PrimeNovo 0.721 0.531 0.603 0.574 0.65 0.567 0.702 0.697 0.697
XuanjiNovo 100M 0.43 0.393 0.435 0.589 0.468 0.502 0.499 0.536 0.407
XuanjiNovo 130M 0.689 0.503 0.574 0.6 0.628 0.682 0.595 0.681 0.67

Read from the paper's source data: every value is a cell of 'Supplementary Data 1.xlsx', sheet SFigure3A, behind a figure that prints no values. Nothing was estimated from bar heights. Values from the paper's Supplementary Data 1 (sheet SFigure3A); the figure prints none. The sheet names InstaNovo where the legend draws InstaNovo+. THE TEXT AND THE SHEET DISAGREE once: p. 8 puts XuanjiNovo-130M on M. musculus at 'approximately 0.67', the sheet at 0.682.

Supplementary Fig. 3b

MassNet: billion-scale AI-ready mass spectral corpus empowering scalable deep learning in proteomics, page 5: Evaluation of de novo peptide sequencing performance on other benchmark datasets. (A) Peptide recall across nine species using seven models: InstaNovo+, π-HelixNovo, ContraNovo, CasaNovo V2, π-PrimeNovo, XuanjiNovo-100M, and XuanjiNovo-130M. (B) Peptide recall on the HCC dataset. (C) Peptide recall on the PT dataset. The π-HelixNovo model is publicly available and was trained on datasets derived from eight species, excluding human. XuanjiNovo-100M: base model; XuanjiNovo-130M: fine-tuned on 30M PSMs from MassIVE.

Proteomics identifies new therapeutic targets of early-stage hepatocellular carcinoma
as deposited 		Peptide recall
HCC
Casanovo V2 0.168
InstaNovo+ 0.087
ContraNovo 0.373
π-PrimeNovo 0.382
XuanjiNovo 0.676

Read from the paper's source data: every value is a cell of 'Supplementary Data 1.xlsx', sheet SFigure3B, behind a figure that prints no values. Nothing was estimated from bar heights. Values from Supplementary Data 1 (sheet SFigure3B); the figure prints none. The figure draws six bars and the sheet has five: π-HelixNovo, whose bar is barely above zero, has no row, so it is not recorded rather than read off the bar. XuanjiNovo is XuanjiNovo-100M, per the caption of Fig. 4.

Supplementary Fig. 3c

MassNet: billion-scale AI-ready mass spectral corpus empowering scalable deep learning in proteomics, page 5: Evaluation of de novo peptide sequencing performance on other benchmark datasets. (A) Peptide recall across nine species using seven models: InstaNovo+, π-HelixNovo, ContraNovo, CasaNovo V2, π-PrimeNovo, XuanjiNovo-100M, and XuanjiNovo-130M. (B) Peptide recall on the HCC dataset. (C) Peptide recall on the PT dataset. The π-HelixNovo model is publicly available and was trained on datasets derived from eight species, excluding human. XuanjiNovo-100M: base model; XuanjiNovo-130M: fine-tuned on 30M PSMs from MassIVE.

ProteomeTools Peptide recall
PT
Casanovo V2 0.459
InstaNovo+ 0.479
ContraNovo 0.574
π-PrimeNovo 0.586
XuanjiNovo 0.605

Read from the paper's source data: every value is a cell of 'Supplementary Data 1.xlsx', sheet SFigure3C, behind a figure that prints no values. Nothing was estimated from bar heights. Values from Supplementary Data 1 (sheet SFigure3C); the figure prints none. The figure draws six bars and the sheet has five: π-HelixNovo (about 0.26 by its bar) has no row, so it is not recorded. THE TEXT AND THE SHEET DISAGREE: p. 8 says XuanjiNovo-100M 'achieved a recall of 0.62' on PT, the sheet gives 0.605.

Fig. 4b

MassNet: billion-scale AI-ready mass spectral corpus empowering scalable deep learning in proteomics, page 7: Comparison of XuanjiNovo and existing de novo peptide sequencing models on the 15-species benchmark dataset. b, Cross-species peptide recall aggregated by species, illustrating overall performance differences across taxa.

15-species benchmark
Nature Methods (307 files) 		Peptide recall
Acinetobacter calcoaceticus Aedes aegypti Arabidopsis thaliana Aspergillus fumigatus Bacteroides thetaiotaomicron Drosophila melanogaster Escherichia coli Homo sapiens Mus musculus Pyrococcus furiosus Saccharomyces cerevisiae Toxoplasma gondii
π-HelixNovo · released 0.41 0.403 0.183 0.286 0.413 0.188 0.114 0.184 0.307 0.398 0.386 0.186
Casanovo V2 · released 0.6 0.57 0.266 0.395 0.652 0.319 0.329 0.224 0.458 0.516 0.598 0.262
InstaNovo+ · released 0.632 0.606 0.262 0.412 0.661 0.346 0.345 0.259 0.524 0.545 0.6 0.241
ContraNovo 0.654 0.615 0.283 0.436 0.678 0.373 0.368 0.299 0.54 0.544 0.626 0.277
XuanjiNovo 0.779 0.757 0.333 0.582 0.793 0.444 0.476 0.449 0.676 0.668 0.691 0.355
π-PrimeNovo 0.733 0.686 0.323 0.51 0.78 0.429 0.466 0.324 0.604 0.616 0.692 0.308
15-species benchmark
Nature Methods (307 files) 		Peptide recall
Trypanosoma brucei Zea mays Solanum lycopersicum
π-HelixNovo · released 0.348 0.443 0.001
Casanovo V2 · released 0.588 0.557 0.011
InstaNovo+ · released 0.623 0.649 0.004
ContraNovo 0.641 0.653 0.062
XuanjiNovo 0.762 0.799 0.252
π-PrimeNovo 0.738 0.731 0.06

Read from the paper's source data: every value is a cell of 'Source Data.xlsx', sheet Figure4B, behind a figure that prints no values. Nothing was estimated from bar heights. Values from the paper's Source Data file (sheet Figure4B); the figure prints none. XuanjiNovo is XuanjiNovo-100M, per the caption. THE TEXT AND THE SHEET DISAGREE once: p. 5 says the other models on S. lycopersicum ranged 'from as low as 0.01 (π-HelixNovo)', the sheet gives 0.001.

Compared against in other papers (1)

1 table from 1 paper that report this method as a baseline, newest first. The caveats of the comparisons above hold twice over: each paper ran, retrained or quoted this method its own way, so its number here is that paper’s claim, not a measurement of the method. Where a paper says which it did, the word follows the method’s name.

Learning from tandem mass spectra at scale with a self-supervised foundation model for proteomics (2026) · 1 table

Table S12

Learning from tandem mass spectra at scale with a self-supervised foundation model for proteomics, page 84: Peptide recall across the biological validation datasets for the de novo sequencing benchmark. The best model per dataset is in bold.

a dataset per evaluation set Peptide recall
HeLa degradome Candidatus Scalindua brodae Snake venoms HeLa single-shot Nanobodies Wound exudates
InstaNovo-FM fine-tuned 0.821 0.740 0.230 0.662 0.528 0.366
InstaNovo v1.2 0.868 0.764 0.227 0.708 0.517 0.431
Casanovo 0.767 0.743 0.217 0.531 0.465 0.262
XuanjiNovo 0.808 0.741 0.233 0.673 0.529 0.466

No accession is printed for the six validation sets. They are mapped from the Methods, which call them InstaNovo's application datasets minus 'the Immuno and Herceptin datasets': GluC is InstaNovo's 'HeLa GluC degradome', TPL Antibodies its nanobodies, and Hela QC, by elimination, its HeLa single-shot set.

Papers describing it (2)

Authors (14)

Jun A, Xiang Zhang (Shanghai AI Lab), Xiaofan Zhang, Jiaqi Wei, Te Zhang, Yamin Deng, Pu Liu, Zongxiang Nie, Yi Chen, Nanqing Dong, Zhiqiang Gao, Siqi Sun, Tiannan Guo, Yiwen Yu

Seen in the charts

Back to the full map

Back to top