pancreatic-lipase-pro-docking

Docks small molecules against human pancreatic lipase (PDB 1LPB, lipase+colipase+Ca2+) across 5 validated sites with AutoDock Vina — pH 7.4 protonation, tautomer/stereoisomer enumeration, multi-seed consensus, native re-dock RMSD gate, and calibration drift detection. Runs locally or on free Kaggle CPU kernels. Use when the user asks to dock, screen, or rank compounds against pancreatic lipase, PNLIP, hPL, or for anti-obesity/lipase-inhibitor virtual screening.

Deterministic Vina screening against human pancreatic lipase. Scores come from Vina, never from a language model.

Rule 0 — do not invent numbers

Report only scores that appear in a generated CSV. If a tool did not run, say so. No estimated affinities, no remembered values, no "typical" numbers. Every claim traces to dock_results/results_all_sites.csv or it is not made.

If vina is missing, the run fails loudly — that is intended. Do not substitute a guess and do not describe a dry run as if it were docking.

Decide the route first

Run command -v vina micromamba conda first, then:

Situation	Route
vina present, or micromamba/conda available to build the env	Local (below) — build the env, do not skip to Kaggle
No local toolchain and no conda/micromamba, or >200 ligands	Kaggle → references/kaggle.md
Something is broken	references/debugging.md

Local run

bash

# 0. env (once). Python must be <=3.11.
micromamba create -p plenv -c conda-forge python=3.11 rdkit meeko vina gemmi openbabel pytest
export PATH="$PWD/plenv/bin:$PATH"

# 1. preflight — MUST print RESULT: OK before docking
python docking_professional_stack/multi_site_docking.py --check

# 2. dock (ligands.csv needs columns: name,smiles)
python docking_professional_stack/multi_site_docking.py \
  --ligands ligands.csv --precision balanced --workers 4 --outdir dock_results

# 3. gate the output — non-zero exit means do not report the numbers
python docking_professional_stack/validate_results.py \
  --results dock_results/results_all_sites.csv

Tiers: fast = ex4/1 seed (triage) · balanced = ex8/1 (default) · max = ex24/3 seeds + consensus (publication).

Read dock_results/results_all_sites.csv — one row per (ligand, site): score is the best affinity in kcal/mol (more negative = better); treat unstable(sd>0.5) rows as unranked.

The 5 sites

catalytic_triad (Ser152-Asp176-His263) · oxyanion_hole (backbone N of Phe77+Leu153) · lid (~239-259) · hydrophobic_pocket (hydrophobes ≤8 Å of Ser152-OG) · colipase_interface (41 cross-chain contacts ≤5 Å).

Centres are computed from the real coordinates at run time. Quote catalytic_triad for inhibitor ranking unless asked otherwise.

Trust the numbers before reporting them

bash

# protocol gate: re-dock the co-crystallized MUP inhibitor (run once per receptor change)
python docking_professional_stack/validate_native.py        # PASS <=2 A, WARN <=3 A

# drift gate: docks ibuprofen+caffeine and compares against calibration/baseline.json
python scripts/selfcheck.py run --python plenv/bin/python   # exit 0 = calibrated, exit 1 = DRIFT

validate_native.py: RMSD ≤2 Å PASS · ≤3 Å WARN (proceed, disclose it) · >3 Å FAIL — stop, do not dock, do not report scores. The protocol is wrong.

selfcheck.py run performs the anchor check for you; you do not dock the controls by hand. Exit 1 means do not publish until the drift is explained. Its stored anchors at the catalytic site: ibuprofen −7.3 to −7.5, caffeine −6.6 to −6.8 kcal/mol.

Vina's own reproducibility band is ~0.5 kcal/mol. Differences smaller than that between two ligands are not a ranking. Say so instead of ranking noise.

Interpreting honestly

• Coarse filter, not binding free energies: rank order within a congeneric series is meaningful, absolute numbers are not. No score proves inhibition.
• Ser152 is a nucleophilic serine — covalent/mechanism-based inhibitors (e.g. orlistat's real mechanism) are not modelled by Vina. Say so.

What this skill does to the machine

Declared up front so nothing below is a surprise.

Capability	When	Scope
Extract archive	restore_and_run.sh	unpacks the bundled stack into docking_professional_stack/ beside the skill. Inspect it first: unzip -l the payload block, or read the extracted .py files.
Install packages	only the micromamba create line you run yourself	into the plenv/ prefix in your cwd (~2 GB, persists until you delete it). conda-forge packages, not hash-pinned. No sudo, no shell-profile edits.
Optional installers (opt-in)	only if you run them	setup_mamba.sh downloads micromamba into $HOME/micromamba when no conda is present. arena_auto_run.py can pip install --user --upgrade 8 packages into your account-wide site-packages — this is disabled by default and requires HPL_ALLOW_PIP_BOOTSTRAP=1. Prefer the plenv/ route above; neither is needed.
Write files	docking	--outdir (default dock_results/), calibration/, and selfcheck_work/. Nothing outside the working directory.
Execute binaries	docking	vina, and obabel/meeko for ligand prep. No other subprocesses.
Network — PubChem	only resolve_names.py	sends the compound names you pass it. Skip it and supply SMILES to stay fully offline.
Network — Kaggle	only scripts/kaggle_dock.py	uploads your ligand CSV and the stack to a Kaggle kernel under your account, and the kernel downloads micromamba + conda-forge packages at run time. Kernels are private by default (--public opts out). --extra is allow-listed and the kernel runs the docking command without a shell. Do not use this route for confidential structures.
Network — none else	—	docking, scoring, prep and validation are fully local and involve no LLM calls.

Not used: no sudo, no system package manager, no writes outside the working directory, no credentials read except the Kaggle token you point at, no telemetry.

Everything else

Need	File
Kaggle: run in the cloud, API keys, debugging kernels	references/kaggle.md
Failures, error messages, environment problems	references/debugging.md
Full CLI surface, every flag, output schemas	references/reference.md
Report generation, name→SMILES, large screens	references/workflows.md

Authorized research and education only.