Workflow
Optimize an enzyme
Protein language models turn a wall of possible substitutions into a ranked shortlist. This workflow takes one pathway enzyme from wild-type sequence to a handful of candidates worth building.
650M
ESM-2 parameters
20
substitutions per site
GPU
engine class
4
stages
01
The goal
A four-hundred-residue enzyme admits roughly eight thousand single substitutions. Almost all of them are neutral or quietly destructive, and the handful that matter are not obvious from the sequence alone. The point of this workflow is triage: move from eight thousand possibilities to five or ten that justify the cost of actually making them.
The enzymes that repay this attention in the capsaicinoid pathway are the ones sitting on the vanillylamine branch, which carries about ninety percent of flux — PAMT, and capsaicin synthase at the pun1 locus. Tuning an enzyme downstream of a bottleneck rarely moves the heat rating; tuning the bottleneck does.
02
Stage by stage
Load the sequence
Pull the wild-type enzyme from the project, from the catalog, or paste a FASTA record. Biopython handles parsing and the annotation engine marks domains and active-site residues so you know which regions you are about to perturb.
Run the scan
Meta ESM-2 650M scores every substitution at every position in a single pass and returns a full matrix. Positions the model considers highly constrained usually correspond to structural or catalytic necessity — useful as a warning as much as a ranking.
Fold the shortlist
Promising variants go to ESMFold for a fast structure, and to AlphaFold- or Boltz-2-class prediction when confidence matters. Mol* renders the result in the browser with per-residue confidence shown directly on the model.
Carry it forward
Chosen variants become part sequences in the project. Codon adaptation for the host, verification primers and construct assembly all pick up from there, so an optimization never dead-ends as a loose protein sequence.
03
Worked example
- Project
- SS-0188 — Tier 4 habanero line
- Enzyme
- Capsaicin synthase (pun1 locus)
- Sequence length
- ~440 residues
- Scan
- ESM-2 650M, full single-substitution matrix
- Shortlist
- 9 variants across 6 positions
- Structures
- ESMFold for all 9, Boltz-2 for 3
- GPU class
- Standard for the scan, Performance for folding
The scan output reads as a heat map: positions across one axis, the twenty amino acids across the other. Two patterns are worth looking for. Columns that are dark almost everywhere mark positions the model will not tolerate changing — leave them alone. Scattered bright cells in otherwise tolerant regions are where a substitution might shift substrate handling without compromising the fold.
Nine candidates survived the first pass here. Folding dropped three on confidence grounds, docking against TRPV1 reordered the rest, and four went into constructs. That ratio is typical, and it is the honest version of the story: the models narrow the search, they do not settle it.
04
What you get
A saved scan keeps the full matrix, not just the shortlist, so a later question about a position you ignored has an answer without paying for the run again. Each variant carries its score, its structure, its confidence metrics, and a link back to the project that requested it.
Variant scanning and structure prediction are GPU engines and run on prepaid credits — $25 for 250 credits, one credit for roughly one Standard job. A full-matrix scan of a single enzyme is a Standard job. The downstream work that turns a variant into DNA is CPU and free.
Ready to run it
Try this in the Splice Platform.
CPU engines are free to run. GPU engines run on prepaid credits.
Related