SolubleMPNN

SolubleMPNN is not a separate package — it is the ProteinMPNN architecture retrained on a soluble-PDB subset, which shifts the output distribution away from the surface hydrophobics that the full-PDB model happily places (because many of them are buried at crystallographic or membrane interfaces in the training set). Reach for it when the goal is soluble yield in a heterologous host; stick with proteinmpnn when native-like recovery matters more, since the soluble prior trades a few points of recovery for the surface bias. Code and weights are MIT (github.com/dauparas/ProteinMPNN, soluble_model_weights; also exposed via github.com/dauparas/LigandMPNN). The model is small enough to run on CPU — for a handful of sequences on one backbone that is seconds and usually faster than dispatching; a GPU helps for batched campaigns. Either way the repo is cloned in-job (no PyPI dist; checkpoints bundled).

Running it

bash
pip install torch numpy   # if not already present
git clone --depth 1 https://github.com/dauparas/ProteinMPNN.git proteinmpnn
cd proteinmpnn
python protein_mpnn_run.py \
  --pdb_path backbone.pdb --pdb_path_chains "A" \
  --out_folder out --num_seq_per_target 16 \
  --sampling_temp "0.1" --use_soluble_model

The runner uses repo-relative imports, so the cd line is load-bearing — invoking the script by absolute path from elsewhere fails with ModuleNotFoundError. If you want threaded designed-sequence PDBs as well, the LigandMPNN runner accepts --model_type soluble_mpnn (see ligandmpnn for that path; it needs ProDy in addition to torch). The flag surface is otherwise identical to proteinmpnn (or ligandmpnn for the second form), including the string-typed temperature and the fixed-position JSONL keyed by PDB stem — see proteinmpnn for the parsing quirks. The repo ships soluble weights at v_48_010 and v_48_020 only; asking for --model_name v_48_002 --use_soluble_model errors on a missing checkpoint, so leave --model_name at its default.

Output is out/seqs/<stem>.fa with score= and seq_recovery= in each header. Expect recovery against a native structure to drop a few points relative to vanilla — that is the prior working, not a bug.

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.

Hydrophobic surface patches still recur where the fold needs them

Soluble weights shift the distribution; they do not enforce a hydrophobicity ceiling. If a particular surface patch keeps coming back hydrophobic, that patch is likely structurally load-bearing and the network is paying the solubility cost to keep the fold. Layering --omit_AAs "CW" or a per-position bias on top is fine, but check that the resulting designs still fold (via boltz or esmfold2) before assuming the constraint was free.

"Crystallisable" training set ≠ "soluble in your host" — keep an orthogonal filter

The training set is "structures that were soluble enough to crystallise," which correlates with but is not the same as "expresses solubly in E. coli at 37 °C." For campaigns where expression yield is the bottleneck, rank the soluble-MPNN output by an orthogonal sequence-based predictor before committing wet-lab slots; treat the MPNN bias as widening the funnel, not replacing the filter.

Next: fold the designs with boltz or esmfold2 to confirm the backbone is still recovered, then carry survivors into the expression screen.