HyperMPNN

5-30 minutesPaperCodeProduction Ready

Design thermostable proteins

Readiness

Readiness: Production Ready
Smoke: passed
Artifacts: passed
Production: production_ready_api_slurm_hypermpnn_weighted_proteinmpnn
Storage: <5
Strategy: shared PyTorch SIF + tool ext3 overlay; avoid per-tool torch duplication
Next: Production Ready. Optional hardening: expose chain selection/fixed positions and alternate HyperMPNN retrained model weights in the UI.

HyperMPNN source cloned from meilerlab/HyperMPNN rev 1877ea181d23fe84fb33147748e1367e5a02b6a9; footprint ~343 MB and retrained ProteinMPNN weights are included. README says to use original ProteinMPNN with --path_to_model_weights/--model_name; no Slurm benchmark/API adapter run yet. 2026-07-05 production-path evidence: direct backend Slurm job 418 reused the validated ProteinMPNN CPU stack with staged HyperMPNN retrained weight v48_020_epoch300_hyper on 6p6f.pentamer.renumbered.pdb. Job generated out/seqs/input.fa with 2 FASTA headers plus summary.json and results.csv. An earlier job 417 reached ProteinMPNN and failed only because this repo does not accept the --device flag. 2026-07-08T20:59:25+08:00: Production Ready promotion: backend runner API job a53c55f77da0 / Slurm job 566 completed through the web runner API using ProteinMPNN with staged HyperMPNN retrained weights v48_020_epoch300_hyper.pt. Benchmark used bundled 6p6f.pentamer.renumbered.pdb from the HyperMPNN source, shared python312_slim.sif, and sequence_design_torch_cpu_py312. Produced zero-byte slurm.out/slurm.err, empty hypermpnn_stderr.txt, parsed_pdbs.jsonl, designed_6p6f.pentamer.renumbered.fa, and results.csv with one designed sample, finite score/global_score 0.8256, seq_recovery 0.3427.