Running RFD3 - network_input (X_noisy_L) for example_id ?

[Katrina818]

My design specifications are as follows:
{"trimmed_hotspot_1": {
"dialect": 2,
"infer_ori_strategy": "hotspots",
"input": "input_pdbs/trimmed.pdb",
"contig": "65-80,/0,A42-191,209-371",
"select_hotspots": {
"A346": "CD1,CE2",
"A347": "CB",
"A349": "CB"
},
"is_non_loopy": true
}

What is causing this assertion error?
AssertionError: network_input (X_noisy_L) for example_id: trimmed_hotspot_1_44: Tensor contains NaNs!

[rclune]

Have you double checked your input structure? It's possible that since it's a trimmed protein there is some additional cleaning that needs to happen before RFD3 can fully interpret the structure.

This error also seems to be coming from models/rfd3/src/rfd3/trainer/rfd3.py and I'm not sure why it would be getting called during an inference run. Could you share the command you are using to run RFD3 or any other details about the calculation?

[Katrina818]

Yes I've double checked the input. It's from a high confidence region in an AlphaFold prediction. Would choosing discontinuous sections of the PDB matter (ie. A42-191,209-371)?

Yes the rfd3 command is from the tutorial:
rfd3 design out_dir=out ckpt_path=../../checkpoints/rfd3_latest.ckpt inputs=trimmed.json inference_sampler.step_scale=3 n_batches=150 +inference_sampler.gamma=0.2

I think the models/rfd3/src/rfd3/trainer/rfd3.py is running OK - it works on other targets I'm designing binders for.

[rclune]

Thank you for clarifying, I see where the error is likely coming from now. You have specified hotspot residues that are not part of the input structure RFD3 can see:

If you are trying to include residues 42-191 and 209-371 from chain A, you need to change your contig string to "65-80,/0,A42-191,A209-371" - but this may cause some issues since residues A191 and A209 are not connected in the input and are automatically held in place (in coordinate space) by RFD3. Right now, RFD3 is interpreting your contig string as:

• design 65-80 residues
• chain break
• residues A42-191 from the input
• design a chain of residues between 209 and 371 amino acids in length.

I would recommend either putting a chain break between the discontiguous sections of chain A you want to include or having RFD3 design a few residues between them.

TL,DR: Putting the chain label in front of each comma-separated portion of your contig is how RFD3 determines if the residues are taken from the input structure or designed. Since your contig was not taking residues 209-371 from your input, RFD3 wasn't able to determine what the hotspot residues are.

[Katrina818]

Oh awesome thank you! Another common error I'm getting is this one below. I'm running the same or similar json files for different targets and the infer_ori_strategy is always the same. Do you know what could be the cause?

AssertionError: Atom array must contain 'is_atom_level_hotspot' annotation to
infer ori from hotspots.

[rclune]

If you are supplying a few hotspots as you did in the above example then the infer ORI strategy should work.

Can you provide an example where you got this error?