Implementing AlphaFold 3 on non-Slurm cluster

[AnastasiiaMamaeva]

Hi again! I'm using LSF cluster computer and I have been trying to run APD as you suggested in a previous topic (#564 (comment)) for a week with help of many different AI, but never succeeded. I have moved to version of APD without snakemake, but as far as I understand it cannot use AlphaFold 3 unlike snakemake-adapted version. Thus, a question comes: can I use the instruction from previous APD version to run last version of APDSnakemake to utilize AP3. Or may be there is another way to incorporate up to date models on non-Slurm computers?

Thank you,
Anastasiia

[DimaMolod]

Hi Anastasiia,

yes, you can run AP on a GPU node via the CLI, either using our AF3 container from Docker Hub:

singularity pull docker://kosinskilab/alphafold3:latest
singularity shell /path/to/your/container
# then run the scripts from the old documentation...

or by creating a fresh conda env using the instructions from the AF3 Dockerfile:
https://github.com/KosinskiLab/AlphaPulldown/blob/main/docker/alphafold3.dockerfile

In the best case, you can keep the AF3-capable AlphaPulldownSnakemake pipeline and just swap the executor by installing the official LSF plugin:

pip install "snakemake>=8" snakemake-executor-plugin-lsf

Then create a profile like config/profiles/lsf/config.yaml so that LSF-specific accounting stays out of the Snakefile. A minimal starting point could be:

executor: lsf

default-resources:
  - mem_mb=8000
  - runtime=60
  - lsf_project=my_project
  - lsf_queue=short

set-resources:
  - structure_inference:lsf_queue=gpu
  - structure_inference:walltime=1440
  - structure_inference:lsf_extra="-gpu num=1 -R 'select[gpu && mem>64000]'"

use-singularity: true

envvars:
  - APPTAINER_BINDPATH
  - SINGULARITY_BINDPATH
  - APPTAINER_NV
  - SINGULARITY_NV

printshellcmds: true
latency-wait: 600

You’ll need to adapt the project/queue names and the lsf_extra flags to match the exact bsub syntax on your cluster.

Hope this helps,
Dima

[AnastasiiaMamaeva]

Thank you!
I have already tried the last option that you have suggested for Snakemake, I will try again with your config
I'll update if I find a working way

[TobyL98]

Hi all,
Just adding to this as I am trying to do similar but with PBS rather than LSF. I cannot get AlphaPulldown snakemake to run properly on my cluster. It seems that the snakemake pipeline is defaulting to running create_features on my login node rather than running as a job.

I am using the cluster-generic executor as there is no specific executor for PBS. Here is my config file (I've tried various different modifications of this but no luck so far): executor: cluster-generic

executor: cluster-generic

# submission command
cluster-generic-submit-cmd: >
  qsub
  -P {resources.project}
  -q {resources.queue}
  -l walltime={resources.runtime_hrs}
  -l select=1:ncpus={threads}:mem={resources.mem_gb}gb{resources.pbs_extra}

jobs: 100

default-resources:
  queue: "v1_small24"
  runtime_hrs: "01:00:00"
  mem_gb: 80
  project: "AP_project"
  pbs_extra: ""

set-resources:
  structure_inference:
    queue: "v1_gpu72"
    runtime_hrs: "24:00:00"
    mem_gb: 64
    pbs_extra: ":ngpus=1"

use-singularity: true
latency-wait: 600
printshellcmds: true

[AnastasiiaMamaeva]

@TobyL98 Hi! I think, I had same problem with not being able to run create features normally, I had to twig many resources parameters and some other things specific to LSF and to my particular cluster, so I doubt my solution will be helpful for you. But regarding snakemake running on login node: don't run snakemake -flag1 -flag2 etc, submit this snakemake command itself to cluster
like for me it would be
bsub snakemake -flag1 -flag2...

then snakemake will run on cluster and will be submitting jobs to cluster (at least it works with LSF)