Binding validation, p_bind

[truzmeto]

Hello, thanks for making this repo available.
I tested this code, specifically RF2 part to predict complex structure given target structre+sequence+nanobody sequence, and epitope (hotspot_frac = 1) for a set of experimentally validated binders and some non-binders. Surprisingly all binders and all non-binders perfectly bind to epitope when full hotspot_prop is shown. But, p_bind values are inconsistent. P_bind seems to correlate well with pae, but doesn't correlate at all with experiment. Paper listed two versions for p_bind calculation, is v2 network version for p_bind available? here:

RFantibody/src/rfantibody/rf2/network/predict.py

Line 238 in 8fe3114

has_v2 = 'bind_pred.rbf2attn.weight' in checkpoint['model_state_dict']

Attached image for complex with all non-binders.
Thanks!

[richieYT-wan]

Hi, semi-related and maybe dumb comment but where do you find the output p_bind ? From the command line and using quiver , I only have these scores at the end of the RF2 step:

interaction_pae	pae	pred_lddt	target_aligned_antibody_rmsd	target_aligned_cdr_rmsd	framework_aligned_antibody_rmsd	framework_aligned_cdr_rmsd	framework_aligned_H1_rmsd	framework_aligned_H2_rmsd	framework_aligned_H3_rmsd	framework_aligned_L1_rmsd	framework_aligned_L2_rmsd	framework_aligned_L3_rmsd	tag

[truzmeto]

They have uncommented it, so it doesn't appear in the output file along with other scores. You need to manually, add it to the output in this code rf2/modules/model_runner.py.
In line 144 add this script to extract p_bind from model output dict:
p_bind = output_i["p_bind"].to(device)

then in line 156 add the following to append it to metrics dict:
metrics["p_bind"] = p_bind

It should appear in the ouput .tsv file afterwards.
It was also suggested by another user in this pull request: #44

It didn't work for multiple complexes I tried. Designing negative control is a challenging task in general. So, I wouldn't rely on it.
Let me know if it works in your case.