peptidebrowser

So representation matters: amino-acid tokens lose chemistry that SMILES/chemical language models can retain. Thus: smaller domain models will work, but only if the input format preserves the molecular information the assay actually sees. More: biorxiv.org/content/10.648…

And while PepBERT is a good sequence-model result, PeptideMTR gets closer to the real chemistry problem. Therapeutic peptides are often not just linear amino-acid strings. They include N/C-terminal mods, cyclization, non-canonical residues, conjugates, lipidation, etc.

PepBERT is a useful counterpoint to the 'bigger model wins' narrative. Peptides are not the open internet. They are short, structured biological sequences where the right training distribution matters more than brute scale.