How it works

The experiment behind the results.

1. PDBPick a protein whose real structure is already known, so the answer can be checked.
2. MSASearch databases for related sequences, then restrict alignment depth.
3. GPURun the selected AlphaFold2 settings and record runtime.
4. ScoreCompare every structure to the real one, and find the cheapest that was still good.

The 12 settings

Two dials, every combination: MSA depth 32 / 128 / 256 crossed with 0 / 1 / 3 / 8 recycles. Everything else is held fixed — AlphaFold2 v2.3.2, model_3, seed 2026, no templates.

What counts as good

The selected study defines which checks must pass.

TM-score	≥ 0.50	overall shape match
CA-lDDT	≥ 0.60	local accuracy
CA-RMSD	≤ 5.0 Å	average atom distance
Coverage	≥ 0.95	how much of the real structure was compared
Identity	≥ 0.99	guard: is this even the right protein

The last two are not accuracy measures. They exist so a good-looking score cannot come from quietly comparing the wrong thing.

Why it matters

Better decisions before compute is committed.

Optifold explores whether sequence features can help prioritize folding experiments for labs with limited compute. Its research asks which sequences pass a defined structural gate, and which tested settings cost less.

Savings and regret

Saving compute is easy if you are allowed to be wrong sometimes. So the saving is always reported with its regret rate — how often the cheap recommendation fails when an expensive one would have worked. Both numbers or neither.

—