2021.09.30 NOTE: we are in the process of recomputing the following results and will update the page shortly
The following are the baseline methods we evaluated
Levenshtein: Levenshtein distance to wild-type.
BLOSUM62: BLOSUM62-score relative to wild-type.
Ridge regression: Ridge regression model on one-hot encoding.
Convolutional network: Simple convolutional network on one-hot encoding.
ESM-untrained: 750M parameter transformer with randomly-initialized weights
ESM-1b: 750M parameter transformer pretrained on UniRef50.
ESM-1v: 750M parameter transformer pretrained on UniRef90. Only one element of ensemble used due to compute constraints.