Registry

A living archive of tools Lazarus has brought back from the dead — each revived from its source repo into a callable, containerised brick with a verified sanity check, and (where a benchmark exists) a reproduced paper number.

6 revived tools. Pull any of them: lazarus pull <name>.

Tool	Domain	Era / stack	Result	From a URL
Basset	Genomics — chromatin accessibility	2016 · Lua Torch7	0.894 vs 0.895 (mean AUROC (164 targets))	✅
DiffDock	Molecular docking	2023 · PyTorch diffusion · ESM-2 · GPU	0.375 vs 0.4 (top-1 success rate (<2Å))	✅
dMaSIF	Protein interface (surface, GPU)	2021 · Py3.6 · torch cu111 · PyKeOps · GPU	smoke ROCAUC ≥ 0.65	—
fpocket	Druggable pocket detection	2010 C · built on modern GCC	smoke pockets ≥ 1	—
MaSIF-site	Protein interaction sites	2020 · Py3.6 · TF 1.12 · MSMS/APBS	0.82 vs 0.85 (median ROC-AUC)	—
ScanNet	Protein binding sites	2022 · Py3.6 · TF 1.14 · Keras	smoke ROC_AUC ≥ 0.7	—

---

Basset basset_predict

Predict DNaseI-hypersensitivity across 164 cell types from a 600 bp DNA sequence.

• Source: davek44/Basset · see source repo
• Stack: 2016 · Lua Torch7
• Sanity check: min_perseq_std ≥ 0.01 · reproduced the paper: mean AUROC (164 targets) 0.894 vs 0.895
• Revived: 48 autonomous agent-turns · from a bare URL (Scout-planned)
• Paper: Kelley et al., Genome Research 2016 — Basset

lazarus pull basset_predict

ℹ️ The pinned image lazarus/basset:site-ready isn't published yet — pull fetches the contract (API + CLI + Dockerfile + smoke test) so it can be rebuilt.

DiffDock diffdock_blind_docking

Blind docking — protein + ligand → ranked, confidence-scored 3D poses (diffusion model).

• Source: gcorso/DiffDock · MIT
• Stack: 2023 · PyTorch diffusion · ESM-2 · GPU · GPU
• Sanity check: rmsd < 2.0 · reproduced the paper: top-1 success rate (<2Å) 0.375 vs 0.4
• Revived: 57 autonomous agent-turns · from a bare URL (Scout-planned)
• Paper: Corso et al., ICLR 2023 — DiffDock

lazarus pull diffdock_blind_docking

dMaSIF dmasif_site

Differentiable molecular-surface interface prediction, built and run on GPU.

• Source: FreyrS/dMaSIF · CC BY-NC-ND
• Stack: 2021 · Py3.6 · torch cu111 · PyKeOps · GPU · GPU
• Sanity check: ROCAUC ≥ 0.65
• Revived: 51 autonomous agent-turns
• Paper: Sverrisson et al., CVPR 2021 — dMaSIF

lazarus pull dmasif_site

ℹ️ The pinned image lazarus/dmasif:site-ready isn't published yet — pull fetches the contract (API + CLI + Dockerfile + smoke test) so it can be rebuilt.

fpocket fpocket2

Detect and rank druggable pockets on a protein structure (Voronoi / alpha-spheres).

• Source: https://fpocket.sourceforge.net · MIT
• Stack: 2010 C · built on modern GCC
• Sanity check: pockets ≥ 1
• Revived: 32 autonomous agent-turns
• Paper: Le Guilloux et al., BMC Bioinformatics 2009 — fpocket

lazarus pull fpocket2

MaSIF-site masif_site

Predict per-residue protein-interaction-site probability from a molecular surface.

• Source: LPDI-EPFL/masif · Apache-2.0
• Stack: 2020 · Py3.6 · TF 1.12 · MSMS/APBS
• Sanity check: roc_auc ≥ 0.8 · reproduced the paper: median ROC-AUC 0.82 vs 0.85
• Revived: 18 autonomous agent-turns
• Given back: masif PR #93
• Paper: Gainza et al., Nature Methods 2020 — MaSIF

lazarus pull masif_site

ScanNet scannet_ppi_binding_sites

Per-residue protein–protein binding-site probability from one structure (structure-only, no MSA).

• Source: jertubiana/ScanNet · Apache-2.0
• Stack: 2022 · Py3.6 · TF 1.14 · Keras
• Sanity check: ROC_AUC ≥ 0.7
• Revived: 19 autonomous agent-turns
• Given back: ScanNet PR #16
• Paper: Tubiana et al., Nature Methods 2022 — ScanNet

lazarus pull scannet_ppi_binding_sites