ESMFold is a folding oracle that wraps Meta’s ESMFold model for single-sequence protein structure prediction. It returns 3D atomic coordinates along with confidence metrics (pLDDT, pTM, PAE) that are consumed by structure-based energy terms.
For multimer prediction, ESMFold uses either a glycine linker between chains or a gap in positional embeddings (default: skip of 512) to signal chain boundaries. These can be configured via the config parameter.
ESMFold 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 ESMFold reference for backend configuration details.
Parameters
Whether to run ESMFold on Modal’s serverless GPU infrastructure. Set to
True for serverless execution (no local GPU required), or False for local GPU execution.Model-specific configuration. Common keys:
glycine_linker(str): linker sequence between chains (e.g.,"GGGG"). Default: no linker.position_ids_skip(int): gap in positional embeddings between chains. Default:512.
Optional Modal app context for reusing an existing Modal session.
Attributes
The result class for this oracle.
ESMFoldResult extends FoldingResult with ESMFold-specific attributes (pLDDT, pTM, PAE arrays).Methods
fold
Fold a list of chains using ESMFold. ParametersThe chains to fold. For multimer prediction, pass multiple chains.
