deepfold.losses.auxillary.experimentally_resolved_loss

deepfold.losses.auxillary.experimentally_resolved_loss(logits: Tensor, atom37_atom_exists: Tensor, all_atom_mask: Tensor, resolution: Tensor, min_resolution: float, max_resolution: float, eps: float = 1e-08) → Tensor[source]

Predicts if an atom is experimentally resolved in a high-res structure.

Jumper et al. (2021) Suppl. Sec. 1.9.10 ‘“Experimentally resolved” prediction’

Parameters:

  • logits – logits of shape [*, N_res, 37]. Log probability that an atom is resolved

  • representation (in atom37)

  • sigmoid. (can be converted to probability by applying)

  • atom37_atom_exists – labels of shape [*, N_res, 37]

  • all_atom_mask – mask of shape [*, N_res, 37]

  • resolution – resolution of each example of shape [*]

Note

This loss is used during fine-tuning on high-resolution X-ray crystals and cryo-EM structures resolution better than 0.3 nm. NMR and distillation examples have zero resolution.