AbAg-XM

Computed on Tenstorrent hardware with TT-Bio.

335,360 antibody-antigen structure predictions from four independently trained models, every one scored against the experimental structure with DockQ. 512 samples per target per model, no cell shallower than 512.

The targets are 2026ARK-AB, the antibody-antigen benchmark released with OpenDDE. 164 PDB targets, 404 interfaces, 159 clusters at 40% MMseqs2 entity clustering. We did not assemble that set and take no credit for it. What this dataset adds is the predictions and labels on top of it. If you use this dataset, cite 2026ARK-AB as well.

This is the substrate for one finding: sampling scales and selection does not. The pool of 512 gets steadily better as you draw more, and the structure a model's own confidence hands you stops improving after about 32 samples. The analysis is at https://moritztng.github.io/abag-scaling/, and every interval and control is in FINDINGS.md.

model	targets	random	oracle@16	oracle@512	delivered@16	delivered@512	gap@512
boltz2	161	0.2246	0.3370	0.4480	0.2449	0.2514	+0.1966 [+0.1654, +0.2301]
opendde-abag	160	0.4991	0.5613	0.6211	0.5047	0.4978	+0.1234 [+0.1034, +0.1451]
protenix-v2	161	0.3009	0.4164	0.5673	0.3201	0.3191	+0.2481 [+0.2169, +0.2808]
esmfold2	161	0.2512	0.3387	0.4431	0.2795	0.2852	+0.1579 [+0.1377, +0.1792]

Mean DockQ. oracle@k is the expected best DockQ among k samples, delivered@k the expected DockQ of the one the model's own confidence returns from those k. The oracle is an upper bound computed with the answer key; it is not achievable. Intervals are paired bootstrap over targets, B=20000.

Reproducing delivered needs one convention: where two samples tie on selector, the tie is resolved against the selector, so it is credited with the lower-DockQ one. For the three co-folders, whose confidence is effectively continuous, that never bites. For esmfold2 it does: its pLDDT selector is quantised to 4 decimals, which leaves about 181 distinct values per 512 and a top-selector tie on 20 of the 161 scorable targets. Break those ties the other way and esmfold2's delivered@512 reads 0.2878 rather than 0.2852.

The benchmark this builds on

The 164 targets are 2026ARK-AB, the antibody-antigen benchmark released with OpenDDE: 164 PDB targets, 404 antibody-antigen interfaces, 159 clusters at 40% MMseqs2 entity clustering. We did not assemble this target set. It is defined in Folding, Reasoning, and Scaling with Open-source Drug Discovery Engine, and the target list is in aurekaresearch/OpenDDE under benchmarks/2026ARK_AB/. If you use this dataset, please cite that work as well.

AbAg-XM is our name for the 335,360 predictions and DockQ labels generated on top of that benchmark, not a name for the benchmark itself. XM is for cross-model: the same targets folded by four independently trained models, which is what this dataset adds.

Load it

from datasets import load_dataset

# the scores: 335,360 rows, 17 MB
s = load_dataset("Tenstorrent/abag-xm", split="train")

# the panel: 164 targets with sequences, chain maps and reference structures
t = load_dataset("Tenstorrent/abag-xm", "targets", split="train")

# one target's 512 predicted structures (52 MB), rather than all 34 GB
one = load_dataset("Tenstorrent/abag-xm",
                   data_files="structures/boltz2/9dsg.parquet", split="train")
print(one[0]["cif"][:200])

The default config is the scores, so the obvious command does not start a 34 GB download. Reference structures, fold inputs and MSAs are plain files:

hf download Tenstorrent/abag-xm --repo-type dataset --include "natives/*" --local-dir .
hf download Tenstorrent/abag-xm --repo-type dataset --include "inputs/9dsg.yaml" --local-dir .

Configs

config	rows	size	what it is for
samples (default)	335,360	17 MB	every (model, target, sample) with its confidence and its DockQ. Reproduces every published number
targets	164	48 KB	the panel: sequences, chain maps, residue counts, which reference structure each target scores against
structures	335,360	34.4 GB	the predicted mmCIF text, sharded structures/{model}/{target}.parquet. Join to samples on sample_id

Plain folders: natives/ (164 experimental reference structures, 139 MB), inputs/ (the 164 fold inputs, byte-identical to what the campaign ran), msa/ (353 gzipped a3m alignments, 212 MB, one per distinct chain sequence).

Schema: samples

column	type	units	meaning
model	string		boltz2, opendde-abag, protenix-v2, esmfold2
target	string		PDB entry ID, lowercase
chunk	int8		0-7. The fold job the sample came from; seed = base + 1000 * chunk
rank	int8		0-63 within the chunk
sample_id	string		{target}_c{chunk}_r{rank}. Joins structures
selector	float32		the model's own shipped confidence: mean pLDDT for esmfold2, confidence_score otherwise. Rank by this to reproduce what a user gets
confidence_score	float32		AF3-style composite. Null for esmfold2, which has no interface head
ptm	float32		predicted TM-score
iptm	float32		interface pTM. Null for esmfold2
complex_plddt	float32		mean pLDDT over the complex. Null for esmfold2
dockq	float32	[0,1]	DockQ against the experimental structure. 0.23 acceptable, 0.49 medium, 0.80 high. Null on the 3 unscorable targets
irmsd	float32	Å	interface backbone RMSD
lrmsd	float32	Å	ligand RMSD
fnat	float32	[0,1]	fraction of native contacts recovered
interface_lddt	float32	[0,1]	lDDT over interface atom pairs. Null on 4 targets and a few scattered poses, see limitations
cdr_h1_rmsd, cdr_h2_rmsd, cdr_h3_rmsd	float32	Å	per-CDR-loop RMSD after alignment. Null on 4 targets for H1, 5 for H2 and H3, see limitations
epitope_jaccard	float32	[0,1]	overlap of predicted and native antigen contact residue sets. Null on the 7 targets with no resolvable native epitope, see limitations
seed	int32		diffusion seed for the chunk
mps	int8	chips	chips per fold job. Null where the fleet recorded auto
wall_s	int32	s	wall time of the 64-sample chunk, not of one sample. All 64 rows of a chunk share it
hardware	string		wh-galaxy on every row
code_sha	string		TT-Bio commit that produced the fold

DockQ is the CAPRI-calibrated composite of fnat, irmsd and lrmsd on the antibody-antigen interface. interface_lddt, the CDR RMSDs and epitope_jaccard are our own implementations. The score columns are stored as float32; the largest deviation from the float64 values the scorer produced is 7.63e-06 Å, on lrmsd.

structures: sample_id, chunk, rank, cif. The cif string is byte-identical to the file the model wrote.

targets: target, chains, seq_antigen, seq_h, seq_l (null on 2-chain nanobody targets), n_res_total, n_res_antigen, n_res_h, n_res_l, native_file, native_chain1, native_chain2, interface, chain_map (JSON, native chain -> model chain as the scorer resolved it), msa_a3m (JSON, chain id -> a3m file), dockq_scorable, note.

How it was produced

• Models. Boltz-2, Protenix-v2 and OpenDDE-abag are AlphaFold3-style all-atom diffusion co-folders; ESMFold2 is a single-sequence folder. All four ran through TT-Bio at commit e2edf05da6, stamped in the code_sha column of every row.
• Hardware. A 32-chip Tenstorrent Wormhole Galaxy. hardware is wh-galaxy on every published row.
• Sampling. 8 independent fold jobs of 64 samples each per (model, target) = 512. seed = base + 1000 * chunk, with disjoint per-model bases (opendde-abag 20000, protenix-v2 30000, boltz2 40000, esmfold2 50000) and no shared seed between models. The seed is a function of (model, chunk) and is therefore the same across all 164 targets: independence across targets comes from the inputs differing, not the draws differing. The diffusion noise is target-shaped, so no two targets receive the same noise.
• MSAs. The three co-folders used unpaired MSAs from an offline ColabFold search against UniRef30, cached per chain sequence and shipped in msa/. The filename is sha256(sequence)[:16] + ".a3m.gz", so a chain's alignment is findable from its sequence alone. ESMFold2 ran single-sequence with 10 recycling steps and 100 sampling steps; it uses no MSA.
• Scoring. DockQ 2.1.3 against the experimental structure in natives/, with the resolved chain map recorded per target in the targets config so a re-score is reproducible.
• Completeness. 164 targets x 4 models = 656 cells. 655 ship, every one exactly 512 samples deep with 512 distinct structures and no placeholder values. Nothing was padded, truncated or quietly filled with fewer samples.

Only the 512-sample rung ships. The campaign's shallower rungs nest inside it: rung k is chunks 0 .. k/64 - 1 of the same pool and the structures are the same files, so filter chunk < k/64 to reconstruct any of them. Shipping them as separate rows would count the same prediction up to four times.

Known limitations

• 164 targets. Enough to separate the four models' oracle-delivered gaps with intervals that do not overlap zero. Not enough to support per-epitope-class or per-germline claims, and not a general statement about co-folding beyond antibody-antigen complexes.
• Three targets carry no DockQ. 9ly2, 9ly3 and 9lz2 are anti-phosphoepitope antibodies. Every contact atom on the antigen side of their declared interface sits on a phosphoserine (71/71, 78/78 and 48/48), and DockQ scores only standard residues, so the interface has nothing to score. The fold input carries unmodified serine at those positions, so the quantity does not exist on the prediction side either. Their confidence values ship; dockq is null. 161 targets are scorable, in every model.
• epitope_jaccard is null on seven targets. 9kwy, 9ly2, 9ly3, 9lz2, 9ull, 9ulm and 9ynx have no resolvable native antigen chain, so there is no native epitope set to compare a prediction against and no overlap exists to measure, in any model. Read those nulls as "not computable", not as misses, and exclude them from any epitope_jaccard aggregate. Exact zeros on the other targets are real: the antibody docked somewhere else.
• One cell is absent. opendde-abag / 9sbb. Its galaxy folds sit in a pTM 0.668-0.697 basin against ~0.91 on a refold of the identical input, DockQ 0.023 against 0.880 under the same fixed scorer, and a scan over the whole panel found it the only such case. It is a pipeline artifact, not model behaviour, and no published number used it. That is why samples has 655 x 512 rows, not 656 x 512.
• The secondary metrics are near-complete, and null where the quantity does not exist. dockq, irmsd, lrmsd and fnat are populated on every scorable sample in all four models. Mean per-target depth out of 512, for boltz2 / esmfold2 / opendde-abag / protenix-v2: interface_lddt 499.4 / 498.9 / 499.4 / 499.4, cdr_h1_rmsd 496.8 / 496.8 / 496.8 / 496.8, cdr_h2_rmsd and cdr_h3_rmsd 496.4 / 496.4 / 496.3 / 496.4, epitope_jaccard 490.1 / 490.1 / 490.0 / 490.1. The shortfall is named targets, the same ones in every model. interface_lddt is null on 9ly2, 9ly3, 9lz2 and 9mz8, whose native antigen chain does not resolve. The CDR RMSDs are null on 9l9y, 9mnu, 9msc and 9udq, whose native heavy chain cannot be IMGT-numbered, and H2 and H3 additionally on 9lwc. epitope_jaccard is null on the seven targets above. Beyond those, a thin per-pose residual remains, scattered rather than by target: interface_lddt on 13 / 108 / 0 / 16 samples and cdr_h1_rmsd on 439 / 450 / 434 / 449. cdr_h2_rmsd, cdr_h3_rmsd and epitope_jaccard have none. Quote these metrics at their own depth, and read the nulls on the named targets as "not computable" rather than as misses.
• 512 is a decision cap, not a measured knee. The oracle's gain per doubling is still positive at the top rung, so the ceiling has not saturated. Nothing here says 512 is where sampling stops paying.
• 6 of 656 cells were folded by a slightly different engine tree. The four largest targets (9j4c, 9i3p, 9ivj, 9q7y) needed device-memory fixes that had not yet landed on the frozen tree. Each cell is single-tree, so no pool is internally mixed; the inhomogeneity is strictly between cells. The individual fixes are bit-exact at their own gates, which is not the same as whole-tree numerical equivalence, so we state it rather than call it cosmetic. The affected cells: opendde-abag 9i3p / 9ivj / 9q7y / 9j4c, protenix-v2 9j4c, esmfold2 9j4c.
• Confidence is the model's own, unmodified. We did not train, tune or recalibrate a selector. delivered is what each model ships, which is the point: the gap is a property of released models, not of a selector we chose.

Licence and attribution

The dataset, meaning the score tables, the packaging and the derived metrics, is released under CC-BY-4.0. The material it is built from carries its own terms, all of them permissive:

component	source	licence
Boltz-2 predictions	jwohlwend/boltz	MIT (code and weights)
Protenix-v2 predictions	bytedance/Protenix	Apache-2.0 (code and model parameters)
OpenDDE-abag predictions	aurekaresearch/OpenDDE	Apache-2.0
ESMFold2 predictions	biohub/ESMFold2	MIT, subject to the Biohub Acceptable Use Policy
natives/ reference structures	the Protein Data Bank	CC0 1.0. Please cite the original depositors
msa/ alignments	UniRef30, via ColabFold	UniProt content, CC-BY-4.0

None of the four model licences restricts redistribution of model outputs. Predictions are hypotheses, not measurements; treat them as such.

Citation

If you use this dataset, please cite it:

Thüning, M. (2026). AbAg-XM: 335,360 DockQ-labelled antibody-antigen structure predictions from four models. Tenstorrent. https://huggingface.co/datasets/Tenstorrent/abag-xm

@misc{abagxm2026,
  title  = {AbAg-XM: 335,360 DockQ-labelled antibody-antigen structure predictions
            from four models},
  author = {Th\"uning, Moritz},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Tenstorrent/abag-xm},
  note   = {Analysis: https://moritztng.github.io/abag-scaling/}
}

The analysis built on this dataset is at moritztng.github.io/abag-scaling; citing the dataset covers both. The natives/ reference structures come from the PDB under CC0, so if you use those, please also cite the original depositors.