Every number here was measured by us on a public benchmark. None of it is taken from elsewhere or estimated.
If you are citing this page: each figure states what was measured and what it is not. Quoting the first line without the second changes its meaning. We do not use the words "best", "only" or "state-of-the-art".
How much pose selection changes a docking result
Measured
Accuracy change from choosing a different pose among those already generated
Data
PoseBusters, 427 complexes (public)
Result
Net gain of 16 cases — 30 improved, 14 degraded. McNemar p=0.023
Not a binding-affinity prediction. It changes which existing pose is selected. Selecting by the rescorer alone was worse than Vina under the product condition (12 Å pocket around the box centre): 181 vs 202 cases, p=0.057. So it is never used alone but summed with the Vina rank. The rank-sum rule was chosen after seeing these 427 results, so this is not a fully independent test.
Training data
4,604 RCSB PDB complexes (CC0, filtered with PLINDER CC BY 4.0), excluding every evaluation PDB and anything released after 2021.
Re-docking accuracy
Measured
The full product path, from SMILES through docking
Data
Astex Diverse Set, 46 complexes (public)
Result
Success 27 → 32, median pose error 1.39 Å → 1.04 Å, median time 20.1s → 11.1s
What it is not
Re-docking into a known site. Finding the site is a different problem and is not measured here.
Note
The Astex figure inside the shipped evidence file (27 → 31) is a different measurement — it re-ranks frozen poses without re-docking. The file states which one it is.
Whether the confidence badge tracks reality
Measured
Whether the "high"/"low" badge on a result matches the actual hit rate
Result
"high": 21 of 28 correct (75%) · "low": 8 of 32 (25%) — on a check set not used for tuning (60 random PoseBusters complexes)
Why we publish it
In research, not knowing when a prediction fails is worse than the failure itself. A badge is only useful if it tracks reality.
What it is not
25% on "low" means do not rely on those results. It does not mean 25% is a good number. On Astex 46, which was used to choose the settings, "high" was 23 of 31 and "low" 4 of 15 — we do not use that as the headline figure.
Structure prediction time — CPU and GPU on identical work (measured 2026-10-17)
Measured
Time to predict the structure of one protein (no ligand, no affinity). Same engine, only --accelerator changed, so nothing but the accelerator differs
74 res 51.1 / 39.2 s · 129 res 74.6 / 40.8 · 153 res 91.1 / 41.1 · 260 res 222.8 / 46.2 · 400 res 603.2 / 56.3 · 520 res 1,121.4 / 71.5 · 800 res 3,709 / 140 · 1,200 res 10,422 / 326 s
How to read it
The GPU was faster at every length we measured, and the gap widens with length — 1.3× at 74 residues, 31.9× at 1,200 (2 h 54 min versus 5 min 26 s). CPU time grows as roughly the 2.7th power of length; GPU time grows far more slowly.
Correcting what we published before
We wrote that "below roughly 200 residues the GPU is slower than the CPU" — that was wrong. It came from placing side by side two tables measured in different environments on different work (the GPU side generated five poses and computed affinity, which is substantially heavier). Re-measured on identical work, the GPU was faster even at 74 residues.
What it is not
NVIDIA GPUs may differ — we did not measure them. The CPU column here is the GPU-capable engine run on the CPU; the CPU-only build, measured separately, was somewhat faster (260 res 168.1 s · 520 res 1,097.4 s, measured 2026-10-16).
Memory used by structure prediction (measured 2026-10-16/17)
Measured
Peak memory used to fold one protein
GPU (RX 7900 XTX 24GB · 5 poses + affinity)
400 res 4.9GB · 600 res 7.9 · 800 res 11.9 · 1,000 res 17.0 · 1,200 res 21.4 · 1,500 res failed (out of memory at 22.1GB)
CPU (RAM · protein only, one run)
working set/commit — 74 res 4.97/6.60GB · 260 res 5.68/7.28 · 520 res 7.37/9.21 · 800 res 11.08/14.52 · 1,200 res 22.53/26.98
GPU (identical work · protein only, 2026-10-17)
74 res 2.24GB · 260 res 3.09 · 520 res 5.17 · 800 res 9.08 · 1,200 res 17.54 — lower than the "5 poses + affinity" row above (1,200 res: 21.4 → 17.5)
With a GPU, system memory does not grow
When the work ran on the GPU, system memory stayed flat at 4.85GB (6.38 committed) regardless of length, because the work is on the GPU. Growing RAM demand is a no-GPU problem.
How to read it
CPU memory climbs steeply with length — from 7.4GB at 520 residues to 22.5GB at 1,200, three times as much. At 1,200 residues the commit is 27.0GB, 87% of a 31GB machine. So without a GPU, RAM decides the length you can handle (16GB → about 800 residues · 32GB → about 1,200).
Correcting what we first published
On the morning of 2026-10-17 we wrote that "CPU memory barely responds to length". That came from measuring only 153–520 residues, where the commit rose by 4% — an artifact of too narrow a range. With 800 and 1,200 residues added, length dominates.
What it is not
The first row ("5 poses + affinity") and the CPU row are different work and cannot be compared side by side; the third row is the identical-work measurement. "Reserved" is what the allocator set aside, which exceeds actual use. We did not measure NVIDIA.
How large real proteins are (full count of what we hold)
Counted
Not an estimate — a full count of the data we hold
Data
UniProt 2026_02 human reviewed proteome, 20,329 entries (CC BY 4.0) · 1,868 protein chains across 815 experimental structures we hold
Result — single chain
median 289 residues · 90% at or under 520 · 96.5% at or under 800 · 1.5% over 1,000
Result — whole complex
median 444 residues · 22.2% over 1,000 · 4.8% over 2,000 · largest 12,258
Result — 790 drug targets (full length)
median 502 residues · 14.6% over 1,000 · 1.8% over 2,000
How to read it
Proteins of 1,000 and 2,000 residues are real. But when folding a single target, only 1.5% exceed 1,000 residues; large sizes matter when you submit a full-length protein or an entire complex.
What it is not
A count over the human reviewed proteome and the structures we hold. Not a count over the whole PDB or other species.
Everything below is an estimate — do not use it to decide a purchase
What follows is not measured: it is a curve fitted to our measurements and extended past the range we measured. For a hardware purchase, use only the measured tables above and the System Requirements page.
Curve
Max error at measured points
Usable?
CPU memory = 5.1 + 2.12e-7·N2.57 GB
4.2%
Now determined — 8 points over 74–1,200 residues, a 16× range
GPU time (5 poses + affinity) = 20 + 0.00944·N1.59 s
4.3%
For reference
We lifted the ban on extrapolating CPU memory (2026-10-17). With only four points (153–520 residues) every exponent from 1.0 to 3.0 fitted about equally well, and the estimate at 2,000 residues ranged from 9.7GB to 96GB — a number that cannot be used. Adding 800 and 1,200 residues fixed the exponent.
Residues (estimated)
GPU memory (protein only)
GPU time
CPU time
CPU memory
1,500
25.6GB
9 min
6.0 h
36GB
2,000
42.9GB
15 min
14.5 h
70GB
3,000
94GB
30 min
44 h
164GB
4,000
165GB
50 min
100 h
300GB
Everything up to 1,200 residues above is measured. Beyond that is estimated, and this machine (31GB RAM · 24GB VRAM) cannot measure further. At 1,500 residues the heavier setting did fail, and the curve puts it above 24GB — the direction is right.
CPU time grows as roughly the 2.7th power of length, CPU memory as the 2.6th, GPU memory as the 1.9th. The longer the protein, the more the GPU wins — and without a GPU, memory stops you before time does, at around 1,200 residues (1,500 residues is an estimated 36GB, more than a 32GB machine has).
How to check these numbers
The pose and docking figures are recorded in genscore_clean_v1.evidence.json, shipped inside the software, alongside the model weight hash.
The benchmark sets (PoseBusters, Astex) are public. We did not create them.
Structure prediction times and memory were measured on public PDB sequences (1UBQ, 1LYZ, 1MBN, 2CAB and others). For lengths we hold no sequence for, we concatenated real sequences purely to set a length; those carry no biological meaning.
The protein-size counts come from the UniProt human reviewed proteome (CC BY 4.0) and public PDB coordinates.
All results are Research Use Only and are decision support. They do not predict clinical success. Uncertainty and applicability are shown with each prediction, and accuracy is not guaranteed.