AMPLIFY

An efficient ESM-2-class protein language model (120M and 350M parameters) pretrained from scratch, from Sarath Chandar's lab as an Amgen–Mila collaboration. AMPLIFY is shipped as a pretrained architecture in NVIDIA BioNeMo and has ~9k HuggingFace downloads, making it one of Mila's most adopted scientific foundation models.

The structure-aligned sibling SaAMPLIFY / SaESM2 (arXiv 2505.16896) aligns the sequence-only pLM to protein-GNN structure via contrastive and structural-token tasks, adding +12.7% ESM-2 contact-prediction accuracy. Released under Apache 2.0.