Structure prediction

A variant scan tells you a substitution is tolerable. It does not tell you what the protein looks like afterwards, and some questions only have structural answers: whether a change sits harmlessly on the surface or crowds the active site, whether a loop that matters has come apart, whether a pose for docking is even worth generating. Structure prediction closes that gap. One sequence in, one three-dimensional model out, rendered in the browser a few minutes later.

Two paths, chosen by what the answer has to support. ESMFold handles the fast pass, where the question is simply whether a design folds at all and you have eight candidates to triage. AlphaFold- and Boltz-2-class models handle the cases where confidence has to be high enough to commit to — a construct order, a docking campaign, a season in the greenhouse. The engine records which path produced each model, and the model carries that label everywhere it goes afterwards.

Starting point

SS-0256 — Benchmark C, Benchmark series

Host

Nicotiana benthamiana

Protein

PAMT, vanillylamine branch

Length

~400 residues

Candidates in

8, from one ESM-2 scan

Fast pass

ESMFold — all 8

Dropped on confidence

3

High-confidence pass

Boltz-2 class — 3 variants

Viewer

Mol*, superposed on wild type

Cost

3 credits per structure

Eight candidates, a fast fold each, and three of them discarded because the active-site region came back poorly resolved. That is a real answer rather than a failed run — a model that will not commit to the region you care about has told you something worth knowing for the price of a quick job.

Three survivors earned the expensive prediction. Superposing them on the wild-type structure is where the case for two of them was actually made: the substitution sat clear of the pocket, the backbone held, and the argument stopped being a score and started being a picture. The third went forward to docking against TRPV1, which is the receptor the whole sensation runs through and the next question anyone asks.

Colour by confidence first, before anything else. High-confidence cores with ragged termini are ordinary and usually fine. A low-confidence patch sitting exactly where your substitution went is a warning, and it is the single most common reason a promising variant gets set aside here instead of six weeks later.

Then superpose. Variants are grouped under the enzyme they derive from, so a wild type and its designed relatives can be overlaid without assembling anything by hand, and the structural consequence of a change becomes something you look at rather than infer from a number. Where a structure has already been docked, the poses are attached to it — though a docking score is a ranking hypothesis, not an affinity, and the platform is careful to present it that way.

Enzyme teams fold the shortlist that comes out of every scan. Cultivar designers fold across hosts, checking whether a part that behaves in Capsicum has a structurally similar counterpart in whatever strange organism the series calls for. Anyone preparing a docking run folds first, because a pose built on a doubtful structure is an expensive way to be wrong.

Folding is a GPU engine and runs on prepaid credits: $25 for 250 credits, with one credit covering roughly one Standard job, and a structure priced at 3 credits. Performance and Max classes are available for longer sequences and the higher-accuracy models. Viewing, downloading and re-reading structures you have already folded is free.