LocalColabFold runs AlphaFold2 (and DeepFold) with cluster-local weights and MMseqs2 MSA databases.

Usage: ColabFold / LocalColabFold.

module load colabfold

This puts colabfold_search / colabfold_batch on PATH and sets COLABFOLD_DATA to /mnt/scratch/colabfold. Always pass --data $COLABFOLD_DATA to colabfold_batch.

Two jobs: MSA on CPU, then folding on GPU.

1. Create a working directory and put sequences in query.fasta. One chain per record; for complexes, separate chains with : in a single sequence (ColabFold convention).

2. Create a Slurm job from the ColabFold MSA template.

MSA needs about 128 GB RAM and no GPU. Guest partition (16 GB) is not enough. Prefer high-RAM nodes.

3. After the MSA job finishes, create a job from the ColabFold template. Use NVIDIA RTX A5000 or A6000.

Do not pass a FASTA file to colabfold_batch. That queries the public MSA server. Pass the msa_out directory (a3m files) from the local search.

4. Adjust the job script (for more details see Slurm Guide) and start the job.

colabfold_batch: --model-type auto (default; monomers use AF2-ptm, complexes use multimer v3), --num-models 5, --templates, --amber / --use-gpu-relax.

colabfold_search: default prefilter is k-mer (high RAM). --prefilter-mode 1 uses ungapped search (more CPU, less RAM). The databases include GPU MMseqs2 indexes; --gpu 1 uses them if the job also requests a GPU. The MSA template stays CPU-only (128 GB), as in the ColabFold batch-search docs.