Masked Models

Input: one or more protein or codon-aligned nucleotide sequences, depending on the toolkit.
Output: sequence embeddings, per-position token probabilities, sampled mutations, or naturalness scores.

Protein- and coding-sequence language models trained to fill in masked positions using surrounding context from both directions. They produce sequence embeddings, per-position token probabilities, sampled mutations, and naturalness scores. These outputs are sequence-level priors, useful for representation, local editing, and ranking rather than structural or functional validation.