Data
Structure gallery
Every pathway enzyme and designed variant the catalog has folded, in one browsable collection — with confidence shown on the model rather than buried in a summary statistic.
Mol*
viewer
ESMFold
fast path
Boltz-2
high-confidence path
GPU
engine class
01
What the dataset is
Structures accumulate as a side effect of design work. Every time a variant is worth a closer look, it gets folded; every fold is kept. The gallery is the resulting collection — wild-type pathway enzymes across the catalog’s hosts, and the designed variants that have been considered seriously enough to model.
Three prediction paths feed it. ESMFold handles the fast pass, where the question is whether a variant folds at all. AlphaFold- and Boltz-2-class models handle the cases where confidence has to be high enough to make a decision on. Which path produced a given model is recorded on the model.
02
What the gallery shows
Interactive models
Mol* renders every structure in the browser — rotate, slice, colour by chain or by confidence, and jump to a residue by number without downloading anything.
Confidence on the structure
Per-residue confidence is painted onto the model. A high average hiding a low-confidence active-site loop is a different situation from a uniformly solid prediction, and only the painted view shows that.
Wild-type and variant together
Variants are grouped under the enzyme they derive from and can be superposed on it, so the structural consequence of a substitution is something you look at rather than infer.
Pathway grouping
Structures are organized by their position in the capsaicinoid pathway, which keeps the vanillylamine branch enzymes — the ones carrying most of the flux control — together.
Docking context
Where a structure has been docked, the poses are attached. TRPV1 docking with AutoDock Vina and DiffDock is the recurring case, since that is the receptor the sensation runs through.
Provenance
Each model names the engine, the version, the GPU class and the run that produced it, and links back to the project. No orphan structures.
03
A worked look
- Enzyme
- Capsaicin synthase (pun1 locus)
- Wild-type model
- Boltz-2 class
- Variants folded
- 9, from one ESM-2 scan
- Fast pass
- ESMFold, all 9
- High-confidence pass
- 3 variants
- Dropped on confidence
- 3
- Docking
- AutoDock Vina and DiffDock vs TRPV1
A typical cluster. Nine variants came out of a single scan, all nine got a fast fold, and three were dropped because the active-site region came back poorly resolved — which is a real answer, not a failure. Three survivors earned the expensive prediction, and the superposition against wild-type was where the argument for two of them was actually made.
Docking reordered the remaining candidates. That reordering is worth treating as a hypothesis rather than a result, and the gallery presents it that way: poses and scores, attached to the structure, with the method named.
04
What you can learn from it
Whether a substitution disturbs a fold or sits harmlessly on the surface. Where confidence is weak and a prediction should not be leaned on. How the pathway’s enzymes relate structurally across hosts. Which variants have been tried before, by whom, and what happened to them.
Browsing the gallery is free. Producing new structures is GPU work on prepaid credits — $25 for 250 credits, one credit for roughly one Standard job — with Performance and Max classes when a model needs to be good enough to decide on.
Ready to run it
Try this in the Splice Platform.
CPU engines are free to run. GPU engines run on prepaid credits.
Related