dIon

algorithm · Transformer (AR)

dIon: algorithm · Transformer (AR). Self-supervised DINO-style pretraining of a spectrum encoder, from a precursor-conditioned spectrum mixture and a partial spectrum with the precursor withheld…

Self-supervised DINO-style pretraining of a spectrum encoder, from a precursor-conditioned spectrum mixture and a partial spectrum with the precursor withheld; initialising a Casanovo-style de novo sequencer from it raises peptide precision over training from scratch.

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

Code

• https://github.com/statisticalbiotechnology/dIon

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

Checkpoints

Version	Trained on	Host	Licence	Size	Checked
v0.1.0	bacterial PXD010000/PXD010613 pretraining corpus (unlabelled)	Hugging Face	Apache-2.0	2.3 GB	live 2026-10-08
v0.1.0	dIon-de-novo-labeled-v1 (DNLv1)	Hugging Face	Apache-2.0	2.7 GB	live 2026-10-08
v0.1.0	dIon-de-novo-labeled-v1 (DNLv1)	Hugging Face	Apache-2.0	2.7 GB	live 2026-10-08

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.

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.

Table1 (Kingdoms de novo)

dIon: Fragmentation-Based Invariance for Self-Supervised Learning of Tandem Mass Spectra, page 7: De novo sequencing on the Kingdoms de novo (all 70 species) and MSKB-final test sets. Within each training regime, scratch and dIon initialization use an identical end-to-end training protocol.

The Proteome Landscape of the Kingdoms of Life	Peptide precision at coverage 1
	Kingdoms
dIon from scratch, end-to-end; trained on MSKB-final	0.449
dIon init., end-to-end; trained on MSKB-final	0.533
dIon from scratch, end-to-end; trained on DNL-v1	0.554
dIon init., end-to-end; trained on DNL-v1	0.602
Casanovo 5.2.1 · released	0.492
InstaNovo 1.2.2 · released	0.570

Peptide precision here is at FULL coverage, unanswered spectra counted wrong, which is Casanovo's precision at coverage 1. Rows are grouped by training corpus, recorded in each variant. InstaNovo 1.2.2 has no MSKB-final value: it 'was trained on MassIVE-KB v1, the source of this test set'.

Table1 (MSKB-final)

dIon: Fragmentation-Based Invariance for Self-Supervised Learning of Tandem Mass Spectra, page 7: De novo sequencing on the Kingdoms de novo (all 70 species) and MSKB-final test sets. Within each training regime, scratch and dIon initialization use an identical end-to-end training protocol.

Casanovo digestion enzyme bias data as deposited	Peptide precision at coverage 1
	MSKB-final
dIon from scratch, end-to-end; trained on MSKB-final	0.747
dIon init., end-to-end; trained on MSKB-final	0.802
dIon from scratch, end-to-end; trained on DNL-v1	0.819
dIon init., end-to-end; trained on DNL-v1	0.842
Casanovo 5.2.1 · released	0.783

Peptide precision here is at FULL coverage, unanswered spectra counted wrong, which is Casanovo's precision at coverage 1. Rows are grouped by training corpus, recorded in each variant. InstaNovo 1.2.2 has no MSKB-final value: it 'was trained on MassIVE-KB v1, the source of this test set'.

Table25

dIon: Fragmentation-Based Invariance for Self-Supervised Learning of Tandem Mass Spectra, page 32: Per-species peptide precision at full coverage on the Kingdoms de novo test set (deterministic run-disjoint split, at most 100k spectra per species).

The Proteome Landscape of the Kingdoms of Life	Peptide precision at coverage 1
	Akkermansia muciniphila	Arabidopsis thaliana (callus)	Arabidopsis thaliana (root)	Arabidopsis thaliana (sprout)	Bacillus subtilis	Bacteroides fragilis	Bacteroides thetaiotaomicron	Bacteroides uniformis	Bacteroides vulgatus	Bifidobacterium adolescentis	Bifidobacterium longum subsp. longum	Blautia obeum
dIon init.; DNL-v1, 200 peaks	0.584	0.617	0.620	0.652	0.559	0.631	0.607	0.632	0.612	0.563	0.525	0.606
dIon from scratch; DNL-v1, 200 peaks	0.535	0.573	0.574	0.605	0.527	0.583	0.561	0.585	0.562	0.519	0.487	0.554
Casanovo 5.2.1 · released	0.487	0.528	0.510	0.537	0.506	0.505	0.493	0.512	0.491	0.471	0.451	0.502
InstaNovo 1.2.2 · released	0.546	0.586	0.581	0.620	0.546	0.574	0.554	0.576	0.554	0.537	0.503	0.572

The Proteome Landscape of the Kingdoms of Life	Peptide precision at coverage 1
	Bos taurus	Caenorhabditis elegans	Canis lupus	Chlamydomonas reinhardtii	Chromera velia	Clostridium bolteae	Clostridium perfringens	Clostridium saccharolyticum	Collinsella aerofaciens	Coprococcus comes	Cricetulus griseus	Danio rerio
dIon init.; DNL-v1, 200 peaks	0.672	0.633	0.631	0.537	0.535	0.587	0.642	0.608	0.570	0.612	0.551	0.603
dIon from scratch; DNL-v1, 200 peaks	0.617	0.584	0.576	0.504	0.492	0.534	0.594	0.556	0.525	0.560	0.500	0.554
Casanovo 5.2.1 · released	0.540	0.519	0.504	0.476	0.450	0.485	0.538	0.502	0.475	0.500	0.451	0.485
InstaNovo 1.2.2 · released	0.649	0.600	0.607	0.520	0.519	0.553	0.600	0.570	0.538	0.569	0.545	0.575

The Proteome Landscape of the Kingdoms of Life	Peptide precision at coverage 1
	Dictyostelium discoideum	Didelphis didelphinae	Dorea formicigenerans	Drosophila melanogaster	Eggerthella lenta	Emiliania huxleyi	Escherichia coli	Eubacterium rectale	Fusarium oxysporum	Fusobacterium nucleatum subsp. nucleatum	Gallus gallus	Glycine max
dIon init.; DNL-v1, 200 peaks	0.622	0.645	0.555	0.640	0.519	0.560	0.619	0.564	0.494	0.641	0.659	0.608
dIon from scratch; DNL-v1, 200 peaks	0.566	0.589	0.514	0.586	0.462	0.515	0.567	0.516	0.457	0.594	0.607	0.561
Casanovo 5.2.1 · released	0.476	0.508	0.485	0.501	0.408	0.449	0.495	0.467	0.438	0.538	0.528	0.494
InstaNovo 1.2.2 · released	0.569	0.611	0.530	0.596	0.482	0.530	0.573	0.521	0.476	0.599	0.622	0.573

The Proteome Landscape of the Kingdoms of Life	Peptide precision at coverage 1
	Gossypium hirsutum	Guillardia theta	Haloarcula marismortui	Halobacterium sp.	Homo sapiens	Lactobacillus paracasei	Mus musculus	Mycoplasma intracellular	Neurospora crassa	Odoribacter splanchnicus	Oryctolagus cuniculus	Oryzias melastigma
dIon init.; DNL-v1, 200 peaks	0.607	0.579	0.652	0.556	0.679	0.586	0.611	0.761	0.553	0.598	0.648	0.593
dIon from scratch; DNL-v1, 200 peaks	0.560	0.533	0.616	0.522	0.624	0.541	0.555	0.698	0.511	0.552	0.582	0.539
Casanovo 5.2.1 · released	0.504	0.472	0.581	0.489	0.543	0.489	0.479	0.636	0.461	0.500	0.497	0.471
InstaNovo 1.2.2 · released	0.575	0.550	0.589	0.518	0.662	0.554	0.596	0.743	0.522	0.551	0.611	0.557

The Proteome Landscape of the Kingdoms of Life	Peptide precision at coverage 1
	Parabacteroides distasonis	Parabacteroides merdae	Phaeodactylum tricornutum	Porphyra umbilicalis	Prevotella copri	Rattus norvegicus	Roseburia intestinalis	Ruminococcus bromii	Ruminococcus gnavus	Ruminococcus torques	Saccharomyces cerevisiae	Stentor coeruleus
dIon init.; DNL-v1, 200 peaks	0.637	0.627	0.555	0.617	0.646	0.580	0.522	0.595	0.623	0.649	0.650	0.615
dIon from scratch; DNL-v1, 200 peaks	0.585	0.577	0.511	0.574	0.598	0.531	0.476	0.549	0.573	0.600	0.602	0.562
Casanovo 5.2.1 · released	0.523	0.520	0.457	0.497	0.537	0.482	0.445	0.505	0.505	0.530	0.517	0.486
InstaNovo 1.2.2 · released	0.588	0.577	0.532	0.583	0.599	0.580	0.497	0.562	0.578	0.604	0.602	0.572

The Proteome Landscape of the Kingdoms of Life	Peptide precision at coverage 1
	Streptococcus parasanguinis	Sulfolobus solfataricus	Sus scrofa	Symbiodinium microadriaticum	Synechocystis nigrescens	Tardigrade	Thalassiosira pseudonana	Triticum aestivum	Veillonella parvula	Vitis vinifera
dIon init.; DNL-v1, 200 peaks	0.627	0.567	0.577	0.551	0.652	0.618	0.562	0.616	0.554	0.581
dIon from scratch; DNL-v1, 200 peaks	0.577	0.514	0.525	0.510	0.615	0.564	0.515	0.569	0.509	0.543
Casanovo 5.2.1 · released	0.518	0.473	0.462	0.477	0.577	0.493	0.454	0.488	0.454	0.510
InstaNovo 1.2.2 · released	0.583	0.543	0.572	0.546	0.633	0.578	0.531	0.577	0.511	0.559

Peptide precision at full coverage per species of dIon's Kingdoms de novo test split. The 'Spectra' column, the number of test spectra per species, is not recorded as a measurement.

Table 26

dIon: Fragmentation-Based Invariance for Self-Supervised Learning of Tandem Mass Spectra, page 33: Complete MSKB-final de novo results, including amino-acid metrics and the dIoninitialized models fine-tuned with a 1,000-peak cap. Within each training regime, scratch and dIon initialization use an identical end-to-end training protocol.

Casanovo digestion enzyme bias data as deposited	Amino acid precision	Amino acid recall	Peptide precision at coverage 1
dIon from scratch, end-to-end; trained on MSKB-final	0.895	0.895	0.747
dIon init., end-to-end; trained on MSKB-final	0.927	0.926	0.802
dIon init., end-to-end, 1,000 peaks; trained on MSKB-final	0.935	0.934	0.822
dIon from scratch, end-to-end; trained on DNL-v1	0.933	0.934	0.819
dIon init., end-to-end; trained on DNL-v1	0.945	0.945	0.842
dIon init., end-to-end, 1,000 peaks; trained on DNL-v1	0.949	0.949	0.856
Casanovo 5.2.1 · released	0.911	0.912	0.783
InstaNovo 1.2.2 · released	0.961	0.961	0.916

MSKB-final, rows grouped by training corpus as in Table 1. InstaNovo 1.2.2 is printed in grey with an asterisk because it 'is documented as training on MassIVE-KB v1, the source corpus of this test set', so its numbers are not a held-out evaluation.
In the paper: column amino acid precision: the original table bolded dIon init., end-to-end, 1,000 peaks; trained on DNL-v1 (0.949).
In the paper: column peptide precision at coverage 1: the original table bolded dIon init., end-to-end, 1,000 peaks; trained on DNL-v1 (0.856).
In the paper: column amino acid recall: the original table bolded dIon init., end-to-end, 1,000 peaks; trained on DNL-v1 (0.949).

Paper describing it

• dIon: Fragmentation-Based Invariance for Self-Supervised Learning of Tandem Mass Spectra (2026, arXiv, preprint)

Authors (5)

Alfred Nilsson, Joel Lapin, Samuel H. Payne, Mathias Wilhelm, Lukas Käll

Seen in the charts

• The architectures
• Code activity
• Models and the authors behind them

Back to the full map