Data
Variant effect maps
Every position of an enzyme against every amino acid it could become, scored in one pass. The map tells you where a protein will tolerate change long before it tells you which change to make.
ESM-2
650M parameters
20
substitutions per site
Full
matrix retained
GPU
engine class
01
What the dataset is
A variant effect map is a matrix: sequence positions along one axis, the twenty amino acids along the other, and in each cell a score for replacing the residue at that position with that amino acid. Meta ESM-2 650M produces the whole matrix in a single forward pass, which is what makes exhaustive scanning practical instead of aspirational.
The scores are not predictions of activity. They reflect how consistent a substitution is with the patterns the model learned across evolutionary sequence data — a proxy for tolerance, which correlates usefully with function but is not the same thing. Treating it as the former is the most common way to misread one of these maps.
02
How to read one
Dark columns are warnings
A position where almost every substitution scores poorly is one the model considers constrained. Those positions usually turn out to be structural or catalytic, and they are the ones to leave alone.
Bright cells in tolerant regions
Scattered high scores in a region that is otherwise relaxed are where a substitution might change behaviour without destabilizing the fold. This is where shortlists come from.
Rows show chemistry
Reading across a row shows how the model treats one substitution everywhere in the protein, which often reveals a chemical preference rather than a positional one.
Overlay the annotation
Domains, active-site residues and binding regions can be overlaid on the map, so the model's opinion and the biology's are visible together rather than in two windows.
Carry it to structure
Any cell opens a path to folding, so a candidate can go from a bright square to an ESMFold model to a Boltz-2 model to a docking pose without leaving the project.
The whole matrix is kept
Saved scans retain every cell, not just the shortlist. A question about a position you dismissed three months ago is answerable without paying for the scan again.
03
A worked look
- Enzyme
- PAMT, amine branch
- Length
- ~400 residues
- Cells scored
- ~8,000
- Model
- Meta ESM-2 650M
- GPU class
- Standard
- Constrained positions
- Clustered near the active site
- Shortlist
- 11 cells across 7 positions
PAMT is the enzyme most worth scanning, because the amine branch carries roughly ninety percent of flux control and PAMT is where that branch tends to constrict. Its map has the shape you hope for: a clearly constrained core, a relaxed periphery, and a modest number of interesting cells in between.
Eleven candidates came out of this one. The map narrowed eight thousand possibilities to eleven in a single Standard job; folding, docking and eventually growing material did the rest. The scan is the cheapest step in that sequence by a wide margin, which is precisely why it goes first.
04
What you can learn from it
Where an enzyme is structurally untouchable. Which regions have slack. Whether a substitution that worked in one homolog has any support in another. Where to spend structure-prediction credits, and — more often than teams expect — where not to bother.
Scanning is a GPU engine on prepaid credits: $25 for 250 credits, one credit for roughly one Standard job, with Performance and Max classes for longer sequences and larger sweeps. Reading a scan you have already run is free and stays free.
Ready to run it
Try this in the Splice Platform.
CPU engines are free to run. GPU engines run on prepaid credits.
Related