Discrepancies in Absolute Coordinates & Missing Side Chains in Output

[patricia-rocha]

hi,

I've noticed the following while experimenting with RFdiffusion:

1. The absolute coordinates of atoms in the output PDB files differ from those in the reference PDB files, even for regions of the structure that have been fixed.

2. Additionally, the output PDB files only contain backbone atoms, with all side chains removed. Given that side chains have degrees of freedom and are essential for accurate structural modeling, shouldn't they also be fixed in the correct orientation? Even though inverse folding models might be able to infer side-chain positions, it seems counterproductive to lose this information during motif scaffolding.

Question: Is there any known workaround or setting that can preserve the absolute coordinates and retain the side chains?

Any guidance or suggestions would be greatly appreciated, thank you!

[simonlevine]

Hi Patricia, not an author but have worked with this model.

1. It sounds like you are working on motif scaffolding. In this setting, the entire complex is initially translated to the alpha carbon centroid of the motif region. Generations are then all with respect to this global frame of reference. As such, the relative coordinates of the motif region should not be changed, but the coordinates would be different.
2. This model is backbone-only (N, Ca, C with O/Cb interpolation), with the potential to use motif side-chain coordinates as well as some side-chain angle information as conditioning. You may find these arguments deep within the feature and model code as well as the lines that indicate the final written coordinates.