Predict protein mutation effects.

Mutation

—

Model

—

Protein sequence

—

Ranking head output (near-duplicate of the score)

Not returned

Direction (probability above assay median)

Not returned

Uncertainty (predicted SD of the score)

Not returned

Structure input

—

Input source

—

Computational prediction only; this result is not experimental evidence and should be experimentally validated where appropriate.

How to read this: the score is on the model's training scale, a per-assay robust z-score where 0 is the median variant of an experiment. It is not an absolute fitness value and not relative to the wild-type sequence. Direction is the model's estimate that the variant scores above that median. Uncertainty is the model's own predicted spread; on held-out validation proteins it tracked the real error only modestly (correlation about 0.23), so treat it as a weak guide. Pasted sequences have no 3D structure, so this runs in sequence-only mode (ESM-C embedding plus sequence and mutation features); accuracy in this mode has not been measured separately. Across held-out proteins the full model reached a rank correlation of roughly 0.43 to 0.50, and individual proteins vary widely.