google-deepmind / google-deepmind/alphafold
the setting of interface_fape loss is different with paper description
Open
- Dominant language
- Python
- Stars
- 14.9k
- Forks
- 2.9k
- PR merge metrics
- No merged PRs in 30d
Description
from multimer paper:
> We achieve this by splitting the FAPE loss into two. The first is clamped at 10 Å and applied to within-chain residue-pairs; the second is clamped at 30 Å and applied to between-chain residue-pairs. These two losses are then summed (which also has the effect of upweighting between-chain residue-pairs).
While in code:
https://github.com/deepmind/alphafold/blob/918434f597c0f8a7bd999f878de474aeb0a0b4ae/alphafold/model/config.py#L624-L627
BTW, it not clear about the sidechain fape loss. Does it also split into two losses?
Contributor guide
Assessment
This issue has not been assessed yet.