Seeking guidance on hotspot specification and unexpected results

[schoyeon]

Dear RFdiffusion Developers and Community,

First, I would like to express my sincere appreciation for this powerful and innovative tool.

I am a researcher currently using the RFdiffusion -> ProteinMPNN -> AlphaFold2 pipeline to design binders (~80aa) for a target protein (~200aa).

My current approach for selecting hotspots (typically 1-5 residues) is based on:
Residues that appear exposed and hydrophobic (inspected via ChimeraX).
Metrics like Pocketness, SASA, and Shape Complementarity.
Known critical binding residues from literature.

A Puzzling Observation
I am consistently observing a very drastic difference in results between runs with zero hotspots and runs with at least one (1+) hotspot specified.
Contrary to expectations, when one or more hotspots are specified, my success rate (leading to folded, well-predicted binders) is much lower than in the 0-hotspot case.
The overall binding ratio also seems to decrease.
Additionally, AlphaFold2 metrics such as RMSD and i_pAE are substantially worse (higher/less stable) for the hotspot-specified designs, while the runs without specified hotspots tend to produce better metrics.
This significant discrepancy occurs regardless of whether I use the full-length target (contigs = "C/0 100-100") or a cropped region of the target (contigs = "C311-391/0 100-100").

This massive difference makes me wonder if I am fundamentally misunderstanding a core concept of how hotspots are intended to be used.

My Questions

1. Is this 50x+ performance gap in success rates and metrics (RMSD, i_pAE) between 0 and 1+ hotspots the expected behavior?

2. If so, I would appreciate it if you could help me understand the methodological reason behind why the difference is so significant.

3. Based on my criteria, could you please share any best-practice advice for hotspot selection or any other binding configuration tips to help achieve more consistent and successful results?

4. I am currently testing designs using a cropped region of the target (e.g., contigs = "C311-391/0 100-100"). I am concerned about whether a binder designed against this cropped region might fail to bind properly to the full-length protein. Is this target cropping a common and valid approach? What are the potential risks or considerations I should be aware of?

RFdiffusion Settings
contigs: "C/0 100-100" (full target) or "C311-391/0 100-100" (cropped target)
hotspot: "" (for 0-hotspot runs) or [Please add your hotspot spec here] (for 1+ hotspot runs)
iterations: 50
num_designs: 200
visual: "image"

ProteinMPNN Settings
num_seqs: 8
initial_guess: True
num_recycles: 3
use_multimer: True
rm_aa: "C"
mpnn_sampling_temp: 0.0001

Thank you very much for your time and any insights you can share. Your help is greatly appreciated.

Best regards.

[alphafp]

Hi Shoyeon,

For hotspots I found the selection of those make a critical difference in the final ipAE results. Authors recommended that 3 hotspots at least to be supplied, but I found 5 hotspots worked the best (I tried 4-6 as well). Also have to make sure that for the surface that the binder sits on, the hotspot is really assessible to the binder (not at a too concave spot). This is my experience but exact details and tips of hotspot designs may be needed to consult the Baker lab.

And in #389 , When you said you run RFdiffusion with 200 or 500 iterations, do you mean the noising step or the denoising step? It seems to me that changing diffusion.T=200 alone would not change the number of timesteps in the inference (denoising) step, and it will also result in this error:
Traceback (most recent call last): File "~/RFdiffusion/./scripts/run_inference.py", line 103, in main denoised_xyz_stack = torch.stack(denoised_xyz_stack) RuntimeError: stack expects a non-empty TensorList Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.

Would you mind enlightening me as to how we can run RFdiffusion with increased iterations as you have said? What parameters are we to modify in the HYDRA configs before running the inference:
../scripts/run_inference.py inference.output_prefix=example_outputs/design_ppi inference.input_pdb=input_pdbs/insulin_target.pdb 'contigmap.contigs=[A1-150/0 70-100]' 'ppi.hotspot_res=[A59,A83,A91]' inference.num_designs=10 denoiser.noise_scale_ca=0 denoiser.noise_scale_frame=0

Please feel free to find me at alphafp20@gmail.com , also, so that we can share more on the RFdiffusion experiments.

Many thanks!

[schoyeon]

Hi Shoyeon,

For hotspots I found the selection of those make a critical difference in the final ipAE results. Authors recommended that 3 hotspots at least to be supplied, but I found 5 hotspots worked the best (I tried 4-6 as well). Also have to make sure that for the surface that the binder sits on, the hotspot is really assessible to the binder (not at a too concave spot). This is my experience but exact details and tips of hotspot designs may be needed to consult the Baker lab.

And in #389 , When you said you run RFdiffusion with 200 or 500 iterations, do you mean the noising step or the denoising step? It seems to me that changing diffusion.T=200 alone would not change the number of timesteps in the inference (denoising) step, and it will also result in this error: Traceback (most recent call last): File "~/RFdiffusion/./scripts/run_inference.py", line 103, in main denoised_xyz_stack = torch.stack(denoised_xyz_stack) RuntimeError: stack expects a non-empty TensorList Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.

Would you mind enlightening me as to how we can run RFdiffusion with increased iterations as you have said? What parameters are we to modify in the HYDRA configs before running the inference: ../scripts/run_inference.py inference.output_prefix=example_outputs/design_ppi inference.input_pdb=input_pdbs/insulin_target.pdb 'contigmap.contigs=[A1-150/0 70-100]' 'ppi.hotspot_res=[A59,A83,A91]' inference.num_designs=10 denoiser.noise_scale_ca=0 denoiser.noise_scale_frame=0

Please feel free to find me at alphafp20@gmail.com , also, so that we can share more on the RFdiffusion experiments.

Many thanks!

Thank you so much for your valuable insights regarding the hotspots. Based on your advice, I will try increasing the number of hotspots to 5 and carefully check the surface accessibility (avoiding concave spots) in ChimeraX. This is extremely helpful for my next designs.

Regarding your question about the iterations (200 or 500) and the RuntimeError: I realized that I am not running the raw CLI commands directly. Instead, I use a Python wrapper script that automatically handles the argument parsing. This is likely why I haven't encountered the stack error you mentioned.

However, based on how my wrapper behaves, I suspect the issue might be related to the difference between standard and partial diffusion. Since your contig setting (A1-150...) implies a partial diffusion task (starting from an existing structure), simply setting diffuser.T might cause a conflict in the inference schedule.

My wrapper seems to automatically detect this mode and adjusts diffuser.partial_T (or the starting timestep) instead of just setting diffuser.T. You might want to check if using diffuser.partial_T solves your empty tensor error.
I hope this helps! I will send you an email at alphafp20@gmail.com later to share more results.

Best regards