Question: Variable usage in RF3 Validation block (mpnn_outputs vs atom_array)?

[Wassef-M]

Hi Team,

I am currently learning the protein design pipeline using the ipd_design_pipeline_collab Tutorial and had a question regarding the RF3 Validation section.

In the current code, the InferenceInput is created using the original atom_array (the output of RFD3):
input_structure = InferenceInput.from_atom_array(atom_array, ...)

As a beginner, I was wondering: should this block instead use the atom_array from mpnn_outputs[0] to validate the sequences actually designed by ProteinMPNN?

When I tested this on the suggested 80-residue design, I noticed that using the MPNN-designed structure improved the overall pLDDT from 0.872 to 0.900 and lowered the Backbone RMSD from 0.81 Å to 0.69 Å.

Is the use of the initial RFDiffusion atom_array intentional for this validation step, or would using the MPNN output provide a more accurate self-consistency check for the final design?

Thanks for the incredible tools and tutorial!

[Ubiquinone-dot]

Generally for backbone consistencies we use the RFD3 atom array, though the coordinates will be the same between it and the MPNN coordinates