Alphafold2

AlphaFold2 (ColabFold runner)

This skill wraps AlphaFold2 and AlphaFold2-Multimer through colabfold_batch, which replaces DeepMind's local-database MSA pipeline with a call to the public MMseqs2 server — so a prediction is one command and one FASTA, not a 2 TB database mount. AF2 remains the reference monomer predictor and the multimer model is still a strong protein–protein validator, but it does not handle ligands or nucleic acids; for those, route to boltz, chai1, or openfold3. The ColabFold code is MIT (github.com/sokrypton/ColabFold) and the AlphaFold2 code is Apache-2.0 (github.com/google-deepmind/alphafold); the AF2 model parameters are CC-BY-4.0 with DeepMind's terms of use.

Running it

colabfold_batch input.fasta out \
  --num-recycle 3 \
  --model-type alphafold2_multimer_v3

The input is a plain FASTA. For a complex, put every chain on one sequence line separated by : — colabfold_batch builds a paired MSA per segment and runs the multimer model when it sees the colon (so the explicit --model-type alphafold2_multimer_v3 above is belt-and-braces). For monomers omit --model-type and the colon. --templates and --amber add PDB templates and OpenMM relaxation respectively; both are off by default and both add minutes per model.

ColabFold runs all five AF2 model weights by default and ranks them by pLDDT (pTM/ipTM for multimer), so output per query lands in out/ as five ranked PDBs <name>_unrelaxed_rank_00{1..5}_*.pdb (b-factor column carries pLDDT) and a matching <name>_scores_rank_00{N}_*.json with plddt, ptm, and — for multimer — iptm and the pae matrix. Rank-1 is the model to read first; ipTM > 0.5 is the usual soft pass for an interface.

Unified-memory defaults loop forever under gVisor — the env patches them out

colabfold/batch.py hard-sets TF_FORCE_UNIFIED_MEMORY=1 and XLA_PYTHON_CLIENT_MEM_FRACTION=4.0 on import. Under a gVisor sandbox unified memory is unsupported, so JAX's device_put loops indefinitely allocating host RAM during AF2 parameter load — the job appears hung, never errors. Override both before the import (TF_FORCE_UNIFIED_MEMORY=0, fraction 0.95), or sed-patch the two assignments out of batch.py in the image build, or the first fold never starts.

The MSA server is the wall-clock bottleneck, and it is shared

colabfold_batch defaults to --msa-mode mmseqs2_uniref_env, which posts your sequence to api.colabfold.com. That server is a public, rate-limited resource: the wait dominates short folds and occasionally times out under load. For campaigns, run the MSA stage once with --msa-only, keep the resulting .a3m files, and feed the directory back as the input on subsequent runs — the GPU stage then starts immediately and the server is not hit again.

Wisp execution

Use python only for bounded interactive checks. For a long or GPU-backed workload, require a selected and probed ssh:<alias> context and load remote-compute-ssh. Put the documented invocation in a self-contained project script, activate the remote environment explicitly, stage only small files with input_paths, and make the command write to a known absolute remote result path. Submit it with run_in_context and register that exact ssh:// path in output_specs. Call monitor_run once when waiting is needed, get_run once for a snapshot, or cancel_run to stop. Do not send a scheduler submission through the SSH-direct runner.

Errors worth recognizing

You see	It means / do this
Job hangs silently during "Running model_1" with host RAM climbing	Unified-memory loop under gVisor — see the gotcha above; override or patch batch.py.
RESOURCE_EXHAUSTED / OOM during XLA compile	XLA_PYTHON_CLIENT_MEM_FRACTION too high for the GPU — drop below the 0.95 default to 0.9 or so.
MSA stage hangs at Submitting job	Public MMseqs2 server is rate-limiting — wait, or pre-compute with --msa-only and re-run from the cached .a3m.

---

Next: for designed-sequence validation, superpose the rank-1 model onto the design backbone with US-align and gate on pLDDT/ipTM thresholds; for ligand-bearing complexes, hand the same chains to boltz or chai1.