A protein language model (pLM) applies the same transformer-based, self-supervised training approach used for natural-language models to protein sequences, treating a chain of amino acids as a sequence of tokens. Trained on large databases of natural protein sequences, a pLM learns representations that capture evolutionary and structural constraints, which it can use to generate new candidate sequences or score how well a given sequence matches the patterns of real, functional proteins.

Because a pLM is trained only to model the distribution of sequences that already exist in nature, generating a sequence a pLM considers plausible is not the same as generating one with improved function for some engineering objective — a gap analogous to the difference between a language model’s raw next-token prediction and a version of it aligned toward what a user actually wants.