RFdifussion and RFdiffusion2 hardware and GPU requirements

[rubenalv]

Hello, could you direct me to a resource with the minimal hardware requirements for inference and generative design? I am thinking on 200-500 aminoacid models.
In a laptop:

• how many GPU Gb and RAM would make it practical (e.g. setting up and configure, or run some experiments, and then burst to cloud or HPC)?
• laptops overheat quickly, so if you think that laptop GPU is not worth it please let me know
• for CPU-only tasks how much RAM would be reasonable? I have read 1-2Gb RAM/core, but what would be the timeframe to run the above experiments?

[HangboZhu]

Hello, I used the PDB file of the tau protein from the example on GitHub. This file contains 700 amino acids. When I ran the script named design_ppi_flexible_peptide_with_secondarystructure_specification.sh, on my local RTX 3090 graphics card, it only occupied less than 3GB of video memory. The memory usage was 16GB. After 10 iterations, it took about 10 minutes. The version of PyTorch is 1.13+cu117. If you want to conduct high-concurrency mining, it is recommended that the memory be above 128GB and the video memory of a single GPU card be above 24GB, which can meet your requirements. Of course, the larger the memory and the video memory of the GPU, the better. Hahaha, I hope this is helpful to you.

[neru1234]

It there open source link for RFdiffsuion2?

[NishikaGolani]

Sized recommendation from production runs: for protein backbone generation up to ~200 residues, a single A10G (24GB) is plenty; ~400+ residues you want H100 (80GB) for the full diffusion trajectory without offloading. CPU memory: 32-64GB per rank. We size Clusterra clusters this way out of the box (managed Slurm, RFdiffusion + AlphaFold + Vina pre-installed, scale-to-zero between runs, your own AWS account). [keystone demo]"
Skip these repos (no signal in 12-mo window): ProteinMPNN, AutoDock-Vina, ColabFold, LocalColabFold, GROMACS GitLab issues. ColabFold/Vina/MPNN scaling pain shows up on Reddit, not the trackers.

[BJWiley233]

Is there anyway to run rf2 say on an HLT of 225 residues for the antibody and the transmembrane domain of a dimer or tetramer with 1200 residues and 24 GB of GPU memory? Basically antibodies are only going to target transmembrane or cell surface antigens which are usually oligomers and 400 to 1200 residues over the Fab. Can someone ask David or Yuan or Nvidia contact if there are any models to download that are trained to be used with smaller arch. on the *ConvSE3 layers?

Brian
"Get it done. Any way possible"

[rubenalv]

@BJWiley233 I'll leave requirements to the developers, but 1200 residues sounds like a lot for that GPU. If it does not fit, you could try two things:

• predict the structure of the 1200-residue protein. Identify if there is a minimal domain which may maintain its original structure in isolation from the rest of the protein, and retains the surface you want to bind against. Predict the structure of the minimal domain and get its RMSD against the original, and if you are satisfied try fitting that minimal structure into your RFD2 approach.
• renting a large machine with 80Gb comes at about $5-$10/h in e.g. Google Cloud. If you are just understanding what you need in terms of hardware, and whether your RF2 approach can work, it's worth getting at least a proof of concept via cloud. There are small grants for academics for Google Cloud and AWS that you could apply for, and larger grants for industry.

[BJWiley233]

Thanks. I starting reading more an it seems like I could just use smaller subset of residues. I am trying again and this seems to be okay but I got a lot of clashes so trying the RFDiffusion from the start with less residues to see if I made some error along the way.

[BJWiley233]

I guess another question I would have is for the proteinmpnn part of RFAntibody, should we be using only the backbone atoms of the target/hotspot residues of the target? It seems like it would miss out on where the side chains of the antigen are when designing sequences.

[rubenalv]

@BJWiley233 I don't have enough experience on that to advise, sorry.