DeepRank-Ab

1-5 minutesPaperCodeNeed >5GB

Score and rank antibody-antigen complex models by predicted DockQ

Readiness

Readiness: Need >5GB
Smoke: partial
Artifacts: partial
Production: need_gt5gb_esm2_cache_and_full_inference_benchmark
Storage: 5-20GB
Strategy: dedicated/shared CUDA PyTorch SIF plus staged ESM2_t33_650M_UR50D cache
Next: Pre-stage ESM2_t33_650M_UR50D and contact-regression assets, run example/test.pdb full inference on GPU through Slurm, then add API runner and artifact packaging.

SOP recovery agent found DeepRank-Ab source at /media/nik/seagate_nik/bio_server/tools/deeprank-ab, origin https://github.com/haddocking/DeepRank-Ab.git, commit 908e411, source size 135M. Required model weight exists at src/weights/af/top_mse_ep37_mse2.3192e-02_treg_ydockq_b128_e100_lr2e-03.pth.tar and vdw_energy_bounds.json exists. Staged package path /media/nik/seagate_nik/bio_server/envs/deeprank_ab_py312_site is 12G and must be appended with site.addsitedir, not prepended in PYTHONPATH, because its numpy 2.5.0 shadows system numpy/sklearn/scipy. GPU import smoke passed: numpy 1.26.4, sklearn 1.4.1.post1, torch 2.12.1+cu130, CUDA true, runtime 5.76s. Full inference was not run because ESM2_t33_650M_UR50D model/cache files were absent and would trigger large downloads. 2026-07-07 16:05 +0800: SOP Runner Adapter Pending audit: moved from Runner Adapter Pending to Need >5GB because only GPU import smoke passed; full DeepRank-Ab inference is blocked on missing ESM2_t33_650M_UR50D/contact-regression cache and a 12GB staged package path already exists.