ThermoMPNN-D

1-5 minutesPaperCodeProduction Ready

Double protein point mutation recommendation for increased thermostability

Readiness

Readiness: Production Ready
Smoke: passed
Artifacts: passed
Production: production_validated_runner_api_slurm_thermompnn_d
Storage: <5GB
Strategy: shared PyTorch SIF + tool ext3 overlay; avoid per-tool torch duplication
Next: Production Ready. Optional hardening: suppress pandas/pyarrow ABI warning noise, expose chain/threshold controls in UI, and add GPU path for larger batches.

SOP agent staged upstream ThermoMPNN-D revision df9a75aaddb674a7c4c193005031fc0536d325fb at /media/nik/seagate_nik/bio_server/tools/thermompnn-d/source. Included ThermoMPNN and ThermoMPNN-D checkpoint files; source+weights about 153M before interrupted env install. Standard venv failed on Seagate because lib64 symlink is rejected; attempted pip --target at envs/thermompnn-d_py312_cuda_site was incomplete. No Slurm benchmark submitted. 2026-07-09T11:58:39+08:00 SOP production validation: promoted from Production Blocked to Production Ready. Added real thermompnn-d backend adapter using shared python312_slim.sif and staged Kuhlman-Lab/ThermoMPNN-D source at /media/nik/seagate_nik/bio_server/tools/thermompnn-d/source. Adapter creates a run-local v2_ssm.py CPU compatibility copy because upstream script hardcodes CUDA device calls, reuses ThermoMPNN site_cpu plus sequence_design_torch_cpu_py312, and validates the bundled examples/pdbs/1VII.pdb benchmark. Ubuntu runner API job 2680534edeb7 / Slurm job 585 completed in 11.161 s and produced 720 finite single-mutant ddG predictions; min ddG -0.8899725 for KA30L, max ddG 3.3779905 for MA13P, mean ddG 0.7829697047752778. Artifacts verified: summary.json, results.csv, thermompnn_d.csv, thermompnn_d_result.json, thermompnn_d_stdout.txt, thermompnn_d_stderr.txt, v2_ssm_cpu.py, 1VII.pdb, slurm.out, slurm.err.