[BUG] RF3 generates unphysical structures for large complexes

[kaistroh]

Describe the bug
Protein complexes with >1000 residues sometimes have collapsed parts (see cyan blob in screenshot), or misformed helices (radius too large, weird shape). Problems seem to appear only for higher token indices, i.e., chains A and B, and part of chain C are fine, the C-ter of chain C is misfolded. I have seen several cases like this, but have not been able to pinpoint a certain number of tokens after which these problems appear. The test case shown here has three chains, 2250 residues in total. In the screenshot below the misfolded, highlighted helix corresponds to residues 953-973 out of 1259 in chain C. The collapsed bits are further down along the chain. The rest of the predicted structure is in agreement with the structure that is deposited in the PDB. (cf. screenshot 2)

To Reproduce

• MSA from ColabFold pipeline
• weights: rf3_foundry_01_24_latest.ckpt
• everything set to default
• rf3 fold inputs='9o5e_ABC.json' ckpt_path='/path/to/rf3_foundry_01_24_latest.ckpt' out_dir='out'
• input file: 9o5e_ABC.json

RF3 runs without any error message. Am I doing something wrong? Is there a token limit that I overlooked?

[roccomoretti]

PR #236 is attempting to address an issue with large structures in RFD3. If you have an ability to test those fixes, it may be useful to know if that PR addresses your issue.