Protein Language Models Fail to Rank Mutations or Prioritize Novel Functions

Aggregating many experiments into one zero-shot protein language model score obscures that current models cannot meaningfully rank a set of fit mutations or prioritize new-to-nature functions