Struct ESMFold2Runner 

pub struct ESMFold2Runner { /* private fields */ }

Wraps a loaded ESMFold2 model for all-atom structure prediction inference.

Create with:

• from_pretrained — structure head only (~755 MB), backbone stubs with zeros (shape testing only).
• from_pretrained_with_backbone — structure head + ESMC-6B backbone (~12 GB total, real predictions).

Implementations

impl ESMFold2Runner

pub fn from_pretrained( model_variant: ESMFold2Models, device: Device, ) -> Result<Self>

Download and load the ESMFold2 structure head only (~755 MB).

The ESMC-6B backbone is not loaded; hidden states passed to fold_protein will be all-zeros, so predicted coordinates are physically meaningless. Use this for shape/pipeline testing without requiring 12 GB of disk space.

pub fn from_pretrained_with_backbone( model_variant: ESMFold2Models, device: Device, ) -> Result<Self>

Download and load the structure head (~755 MB) and the ESMC-6B backbone (~12 GB). Required for predictions with meaningful pLDDT.

pub fn fold_protein( &self, sequence: &str, num_loops: usize, num_sampling_steps: usize, ) -> Result<ESMFold2Output>

Fold a single protein sequence.

Arguments

• sequence — amino-acid string (standard one-letter codes)
• num_loops — recycling iterations (typically 3)
• num_sampling_steps — diffusion denoising steps (14 fast / 50 quality)

When the runner was created with from_pretrained (no backbone), hidden states are zeros and coordinates will be arbitrary. Use from_pretrained_with_backbone for real predictions.