Skip to main content
ESM2 is an embedding oracle that wraps Meta’s ESM-2 protein language model. It produces high-dimensional per-residue embeddings that capture biochemical and evolutionary context, which are used by embedding-based energy terms like EmbeddingsSimilarityEnergy. By default, BAGEL uses the 650M parameter version of ESM-2, but any ESM-2 model size can be specified via the config parameter. For multimers, ESM-2 uses the same linker and positional encoding approach as ESMFold.
ESM-2 inference is powered by boileroom. boileroom handles model loading, GPU execution, and dependency isolation — either serverlessly via Modal or locally via Apptainer. See the boileroom ESM-2 reference for backend configuration details.

​
Parameters

​
use_modal
bool
default:"False"
Whether to run ESM-2 on Modal’s serverless GPU infrastructure. Set to True for serverless execution (no local GPU required), or False for local GPU execution.
​
config
dict[str, Any]
default:"{}"
Model-specific configuration. Can be used to specify model size, linker parameters, and other options.
​
modal_app_context
App | None
default:"None"
Optional Modal app context for reusing an existing Modal session.

​
Methods

​
embed

Calculate the embeddings of the residues in the chains. Parameters
​
chains
list[Chain]
required
The chains to embed. Sequences are concatenated with appropriate linker handling.

​
Example

import bagel as bg

# Create an ESM-2 oracle using Modal
esm2 = bg.oracles.ESM2(use_modal=True)

# Extract reference embeddings from a conserved region
reference_embeddings = esm2.predict(chains).embeddings

# Use with EmbeddingsSimilarityEnergy to maintain functional similarity
state = bg.State(
    chains=[chain],
    energy_terms=[
        bg.energies.EmbeddingsSimilarityEnergy(
            oracle=esm2,
            weight=1.0,
            residues=[conserved_residues],
            reference_embeddings=reference_embeddings,
        ),
    ],
    name="enzyme_variant",
)