Compose & reproduce¶

Compose — a pipeline from revived bricks¶

Because every revival speaks the same contract, revived tools compose regardless of domain, language, or era. A pipeline is a small YAML; one command runs it, passing file artifacts between steps on the host you choose.

Two pipelines, one question — is this target a druggable small-molecule cleft, or a flat protein-protein interface? — reaching opposite, both-correct verdicts. PD-L1 comes back a biologic target. BRD2 comes back a small-molecule one, cross-validated sub-Å against the drug already sitting in its crystal structure.

PD-L1 → a flat interface¶

examples/pipelines/binder_triage.yaml assembles three methods that were each individually unrunnable a week ago:

structure ─▶ ScanNet ─┐
          ─▶ dMaSIF ──┼─▶ consensus ─▶ interface residues that also line a druggable pocket
          ─▶ fpocket ─┘

lazarus run examples/pipelines/binder_triage.yaml \
  --input structure=4ZQK.pdb \
  --registry examples --registry components \
  --docker-host ssh://you@your-x86-gpu-box

Run live on PD-L1, it concluded: 27 interface residues clearly localized, but 0 druggable pockets among them → "a flat protein-protein interface: an antibody / biologic target, not a small-molecule one." That's textbook immuno-oncology (PD-1/PD-L1 is an antibody target), reproduced from dead code.

BRD2 → a confirmed drug pocket¶

pipelines/target_dock_consensus.yaml asks the same question of a different target and takes it a step further — it doesn't just locate the site, it docks a drug into it. Four tools spanning 2010→2023, C→Python, CPU→GPU, geometry→learned→generative, feeding one consensus adapter:

        ─▶ fpocket  (2010 C · geometry)       ─┐
complex ─▶ ScanNet  (learned · PPI-site)      ─┼─▶ consensus ─▶ CONFIRMED small-molecule site
        ─▶ DiffDock (2023 · generative · GPU) ─┤
        ─▶ EquiBind (generative · CPU)        ─┘

dock_prep splits the crystal complex into receptor and reference ligand; the JW4 SMILES feeds the two dockers.

lazarus run pipelines/target_dock_consensus.yaml \
  --input complex=6moa.pdb --input ligand_smiles=jw4.smi \
  --registry examples --registry pipelines \
  --docker-host ssh://you@your-x86-gpu-box

Run live on BRD2's BD2 bromodomain (PDB 6MOA, 109 residues, chain A) — a validated BET oncology target — the four methods reinforce rather than cancel. fpocket's top pocket scores druggability 0.93 over 21 lining residues; ScanNet, asked the PPI question across those same residues, stays quiet (mean 0.26) — the signature of a small-molecule cleft, not an interface. Both dockers land inside that pocket: DiffDock's top pose sits 0.20 Å (centroid) from the crystal ligand, 90% of its contacts in the pocket; EquiBind independently lands at 0.84 Å centroid / 1.30 Å RMSD, 100% in-pocket (atom order preserved, so that RMSD means what it says); and the two generative dockers agree with each other to 1.02 Å. Decisively, both converge on Asn429 (the conserved acetyl-lysine anchor) and Trp370 (the WPF-shelf tryptophan) — the real BET pharmacophore, not merely somewhere on the protein. Verdict: "CONFIRMED small-molecule site" — the same question PD-L1 answered, with the opposite result, this one cross-validated sub-Å against the drug already in the crystal. One command, run on a remote A4500 GPU box (Bertha) via --docker-host ssh.

The whole run is committed — every output file, the exact command, and an honest note on what reproduces exactly and what doesn't — at pipelines/sample_output_6MOA/. fpocket, ScanNet and EquiBind are deterministic and repeat exactly; DiffDock is a generative model, so its pose shifts slightly between runs (a second independent run gave 0.22 Å / 91% / 0.96 Å). The verdict, the pocket, and the Asn429 + Trp370 convergence are stable across both.

Reproduce — the trust layer¶

A smoke test proves a method runs; a benchmark proves it's the method. A contract's benchmark field emits a REPRODUCE.md certificate with a PASS/OFF verdict.

Method	Paper	Lazarus	Verdict
MaSIF-site transient PPI benchmark	0.85	0.82	reproduced (±0.05)
Basset mean AUROC over 164 cell types	0.895	0.8944	reproduced

Basset's reproduction is what exposed a silent bug: the naive run scored mean AUROC 0.675 because half the genome's soft-masked (lowercase) bases fell through the one-hot encoder. Reproducing the paper — not merely executing the code — is what caught it. Full story: the hard problems it solved.

Give back¶

For the genuinely-abandoned repos, Lazarus prepares maintainer-ready PRs — the real fix plus a CI smoke test so the method can't silently rot again:

MaSIF — PR #93: the rotted PDB download, fixed.
ScanNet — PR #16: library_folder='' auto-detects the repo root.