ESM-2

ESM-2 (Evolutionary Scale Modeling) is a family of protein language models developed by Meta's Fundamental AI Research (FAIR) team, ranging from 8M to 3B parameters, that learn structural and functional representations directly from protein sequence. ESM-2 is the base model for ESMFold's single-sequence structure prediction and underpins a wide range of genomics-adjacent variant-effect-on-protein and protein-design workflows. Its source code is confirmed MIT licensed (the ESM Metagenomic Atlas data, a separate artefact, carries a separate CC-BY-4.0 licence not applicable to the model weights/code catalogued here).

Registry note on model_card_uri: this points to the model's standard Hugging Face model card, not a GA4GH Genomic AI Model Card v0.6.0-formatted document. No model has yet published a card in the GA4GH v0.6.0 format, since it is a newly drafted specification. ESM-2 is protein-sequence-native rather than DNA-native; it is catalogued here because protein-level variant-effect modelling is a common downstream step in genomic AI pipelines, per PRD.md Section 4.1's task-relevance test rather than a DNA-specific one.