Using constraints seems to have little to no effect

[jmrussell]

We have experimental data showing contacts between residues <35 angstroms and we wanted to evaluate chai fold's ability to improve on previously predicted structures using the constraints feature.

After running our inferences with and without the constraint files, I find almost no difference in the distance between our residues of interest (each group of dots is a residue pair we hoped to see 'constrained', I calculated distance for all 5 ranked inferences):

Any idea why the non-constrained and constrained results would be close to each other? I would expect to see at least some bias toward closer contact. Here is an example line from our constraints file:

K70,K490,A,A,contact,1.0,0.0,35.0,xlink,restraint_3

[wukevin]

Thanks for trying the restraints! The restraints in Chai-1 are trained as inter-chain contacts; based on the line you've provided, it looks like you're trying to specify an intra-chain contact. Intra-chain contacts are more akin to structure templates, which you may be able to hack into the TemplateContext.

[arogozhnikov]

Hi @jmrussell if you have some (preferably small) example to share, that would help us in future debugging.