Trying to run inference script

[Valentin-Bio]

Hello, I'm learning how to run RFdifussion2 and got a little nit confused:

My goal: design a protein from a theoenzyme that contains 4 residues and one coenzyme (FMN).

For my understanding I need to first specify a ORI atom that specifies the center of mass of my coenzyme. (I did this since i converted a cif file to pdb)

Then I need to know what scripts should I modify. For this I created an own json file (active_site_unindexed_atomic.json) that looks like this:

{
    "active_site_unindexed_atomic":                "inference.input_pdb=theoenzyme_ladA_with_ORI.pdb inference.ligand=\\'FMN\\' contigmap.contigs=[\\'16,B17-17,45,B63-63,15,B79-79,58,B138-138,172,B311-311,129\\'] inference.contig_as_guidepost=True contigmap.contig_atoms=\"'{\\'B17\\':\\'CB,CG,ND1,CD2,CE1,NE2\\',\\'B63\\':\\'CB,CG,CD1,CD2,CE1,CE2,CZ,OH\\',\\'B79\\':\\'N,CA,C,O\\',\\'B138\\':\\'CB,CG,ND1,CD2,CE1,NE2\\',\\'B311\\':\\'CB,CG,ND1,CD2,CE1,NE2\\'}'\""
}

From a previous error message I notice the program also asks me to specify a yaml file and I used this yaml file that I modified of the benchmark config file:

defaults:
  - enzyme_bench_n41
  - _self_

in_proc: True
sweep:
  retries: 2
  command_args: >
    --config-name=aa
    inference.deterministic=True
    inference.ckpt_path=REPO_ROOT/rf_diffusion/model_weights/RFD_173.pt
    inference.seed_offset=43
  
  num_per_condition: 1
  num_per_job: 1
  benchmark_json: config_theoenzyme.json

  slurm:
    p: gpu-bf
    gres: gpu:1

stop_step: sweep
outdir: ${hydra:runtime.cwd}/pipeline_outputs/${now:%Y-%m-%d}_${now:%H-%M-%S}_open_source_demo

as it points out to a enzyme_bench_n41.yaml file, I notice maybe I should delete the - enzyme_bench_n41 part of the yaml file isnt it?

This is the script I'm running (I'm using singularity):

program_dir=/opt/RFdiffusion2

singularity exec --nv --env PYTHONPATH=${program_dir} --env MKL_THREADING_LAYER=GNU --env CUDA_VISIBLE_DEVICES=0 \
	-B /opt:/opt ${program_dir}/rf_diffusion/exec/rfdiffusion_sandbox \
    python3 /opt/RFdiffusion2/rf_diffusion/run_inference.py \
    --config-name=config_theoenzyme \
    sweep.inference=active_site_unindexed_atomic \
    inference.num_designs=10

Aditionally, I'm naively passing the name of the config file to sweep.inference . I get an error message that there is no struct called inference. So in addition, how can I provide my config file ?

I know this may look that I'm lost in trying to run this but is my first time using this type of program. just looking for some clarity and feedback on what I'm trying to do.

[Valentin-Bio]

corrected by removing the config file and by specifying options directly:

CONTIG_ATOMS='{"B17":"CB,CG,ND1,CD2,CE1,NE2","B63":"CB,CG,CD1,CD2,CE1,CE2,CZ,OH","B79":"N,CA,C,O","B138":"CB,CG,ND1,CD2,CE1,NE2","B311":"CB,CG,ND1,CD2,CE1,NE2"}'

singularity exec --nv --env PYTHONPATH=${program_dir} --env MKL_THREADING_LAYER=GNU --env CUDA_VISIBLE_DEVICES=0 \
    -B /opt:/opt \
    -B $HOME:$HOME \
    ${program_dir}/rf_diffusion/exec/rfdiffusion_sandbox \
    python3 /opt/RFdiffusion2/rf_diffusion/run_inference.py \
    inference.input_pdb=theoenzyme_ladA_with_ORI.pdb \
    inference.output_prefix=/home/vberrios/comau/proyecto_tesis/ladA_theoenzyme/diffusion_results \
    inference.ligand='FMN' \
    "contigmap.contigs=['16,B17-17,45,B63-63,15,B79-79,58,B138-138,172,B311-311,129']" \
    inference.contig_as_guidepost=True \
    "contigmap.contig_atoms='${CONTIG_ATOMS}'" \
    inference.num_designs=10 \
    inference.ckpt_path=/opt/RFdiffusion2/rf_diffusion/model_weights/RFD_173.pt \
    diffuser.T=100

[rclune]

Thank you for posting your solution!

You should be able to use a combination of a JSON and YAML file, like you had in your original attempt.

The enzyme_bench_n41.yaml it was pointing to can be found here, and contains other options that you can use in your simulations, that is why you want to include it. You'll notice that is has pipeline as its default, which has it's own set of defaults.

If you are using a JSON file, it is best to put all of your configuration options, like inference.num_designs=10, in the JSON file, instead of a mixture of JSON and command line like in your example above.

I'm not sure sweep.inference is a configuration option that is defined anywhere, were you trying to use sweep.benchmarks?

For usage of RFD2, I recommend using only the aa.yaml and aa_ppi.yaml unless you are trying to do something similar to the benchmark calculations that are discussed in the README.

[Valentin-Bio]

Thanks so much for your response n.n. So if I run RFD2 without specifying any config file, I'm loosing some specifications ? or the inference script looks for some default config that manages options that I'm not directly specifying? For example, I do not specify in my modified script whether RFD2 must use indexed or unindexed design.

[rclune]

I believe that it defaults to using base.yaml, if one exists. The .yaml files define the different configurations that you can use when running RFdiffusion2, like inference.num_designs. So more like what you mentioned in your second question. Follow the path back to where the option for indexed or unindexed design was defined and you should be able to see what the default setting is.