ESM-2 variant scoring

Protein engineering has a shortlist problem. A 450-residue enzyme offers something like nine thousand single substitutions, and almost all of them are bad in ways that only become apparent after someone has spent a month finding out. The scanner moves that filtering step forward. Meta’s ESM-2 650M protein language model scores substitutions against the patterns it learned across evolutionary sequence data, and what it returns is an ordering: which changes look consistent with a protein that still works, and which look like vandalism.

Scores are a proxy for tolerance, not a measurement of activity. That distinction matters enough that the interface keeps restating it. What the engine is genuinely good at is telling you where not to spend effort, which in practice is most of the value — a ranking that removes eight thousand candidates is worth more than a confident prediction about one.

Starting point

SS-0042 — White Hot, Scorch series

Host

Capsicum chinense

Protein

Capsaicin synthase (pun1 locus)

Length

~450 residues

Mode

Exhaustive — every position

Variants scored

~9,000

Model

Meta ESM-2 650M

GPU class

Standard

Cost

1 credit

Shortlist taken forward

S39L, L345G, and 6 others

Nine thousand substitutions, one job, one credit. The heatmap came back with the shape you hope for and do not always get: a clearly constrained core clustered around the active site, a relaxed periphery where the model has no strong opinion, and a modest band of interesting cells in between. Eight candidates came out of that band.

Two of them — S39L and L345G — are the pair that ended up in the White Hot construct and in a good deal of the catalog besides. The scan did not prove they would work. It put them in a group of eight instead of a group of nine thousand, and the folding, docking and grown material that followed did the rest.

Start with the columns. A position where nearly every substitution scores badly is one the model considers load-bearing, and those positions are usually structural or catalytic. Leave them alone. The useful cells are bright squares sitting inside otherwise relaxed stretches — somewhere a change might alter behaviour without destabilizing the fold.

Reading across a row tells you something different: how the model treats one substitution everywhere in the protein, which often exposes a chemical preference rather than a positional one. Domain boundaries, active-site residues and binding regions can be overlaid on the map, so the model’s opinion and the known biology sit in one window instead of two.

Finished scans are kept whole. The shortlist is what you export, but the full matrix stays in the project, which means a question about a position you dismissed last quarter is answerable without paying for the scan a second time.

Enzyme teams use it as the first filter in any optimization campaign. Cultivar designers use it to check whether a substitution that worked in one host has any support in a homolog from another. Groups running comparative work queue a scan per organism as a batch and read the maps side by side.

Scanning is a GPU engine and runs on prepaid credits: $25 for 250 credits, with one credit covering roughly one Standard job. Performance and Max classes are there for long sequences and large sweeps. Re-reading a scan you have already run costs nothing, now or later.