The workflow was executed on Linux x86_64 CPU using the public 1EVE/E20 example. GNINA 1.3.3 generated three poses, PoseBusters 0.6.5 checked all three, and PLIP 2.4.0 analyzed all three with matching chemical identity. This is computational workflow validation; no experimental or biological validation has been performed.

Purpose and mechanism

PLIP needs a protein–ligand complex, not an isolated SDF. The corrected assembler reserves LIG:Z:1, gives TER its own serial, offsets ligand serials accordingly and writes repeated CONECT targets for kekulized bond orders. It checks atom counts and coordinates; PLIP then independently confirms chemical identity. The original reference conversion exited zero but had a different InChIKey because TER/CONECT numbering was inconsistent; that report was rejected and preserved. PLIP 2.4.0 emits report.xml/report.txt under the selected directory. It has no working standalone --version option; help, package metadata and XML record the version. Contact counts are rule-based descriptions, not additive binding energies.

Explicit input preparation

These checks use the actual three-pose GNINA SDF, the coordinate-matched E20:A:2001 reference and the repaired, hydrogenated A-chain receptor. PDBFixer 1.12.0 and OpenMM 8.6.1 preserve the observed heavy-atom frame; neutral CCD E20 and omitted water/glycans are explicit choices. Preparation details are described in the GNINA guide.

Observed results

Pose index minimizedAffinity CNNscore PoseBusters PLIP total
0 -10.83538 0.9389097691 checks_passed 14
1 -10.87303 0.8397080898 validation_failed 9
2 -10.0518 0.8295055628 validation_failed 9

Pose index 0 passed all 28 enabled PoseBusters checks. Indices 1 and 2 failed only rmsd_≤_2å; the other enabled checks were true. This distinction matters: their failure is a reference-pose criterion, not an assertion that all geometry is invalid. All three were sent to PLIP; indices 1 and 2 retain validation_failed. PLIP confirmed LIG:Z:1, the same E20 InChIKey and 28 heavy atoms for each. No pose or failure was silently excluded.

Tested commands and file handoff

PROJECT_DIR=/path/to/molecular-docking-validation
RUN_DIR="$PROJECT_DIR/runs/example-run"
INPUT_DIR="$PROJECT_DIR/data/prepared"
cd "$PROJECT_DIR"
mkdir -p "$RUN_DIR"/{dock,config,validation,complex,interactions}
plip -f "$RUN_DIR/complex/pose_001_complex.pdb" -x -t \
  -o "$RUN_DIR/interactions/pose_001"

The commands below retain the parameters of the recorded calculation, with example paths. Replace /path/to/molecular-docking-validation with your own directory. Prerequisites are installed tools with their runtime libraries available on PATH and the prepared receptor.pdb, independently generated ligand.sdf, and coordinate-matched reference.sdf in data/prepared/. PoseBusters additionally needs its installed redock.yml copied to runs/example-run/config/redock.yml; PLIP needs each receptor–pose PDB complex already assembled and checked for atom serials, TER/CONECT records, bond orders, coordinates and chemical identity. This article does not distribute a complete project, prepared data or assembly scripts, so these are conditional command examples rather than a self-contained copy-and-run package. Use fresh output directories and preserve the input files, parameters and logs.

Failures and acceptance

Actual negative runs covered missing/invalid receptor, empty/invalid SDF, missing PLIP file, receptor-only input and a translated zero-contact control. GNINA and bust can exit zero for empty or bad ligand files while producing no valid result, so require file existence, parsed records and full coverage. A PLIP receptor-only XML has no binding site; it differs from a recognized, identity-matched ligand with zero contacts. The recorded calculation retained the original errors and corrected results; those records are not distributed with this article. GNINA emitted a space-groups.txt warning, and its SDF header caused RDKit to infer 3D from nonzero z coordinates; raw files and warnings were preserved, while all three structures passed finite-coordinate and chemical-identity checks.

Environment and scope

Tested host: Ubuntu 24.04.2, Linux x86_64, two vCPUs, about 8 GiB RAM, Python 3.12.15. The project has its own environment and versioned GNINA binary; the SHA-256 is 3340c1f49cd3c7c84d8699182a1c6af13c7fa2a22448d1204640446106f72172. Although execution uses --no_gpu, this binary needs CUDA/cuDNN loader libraries. They are installed only in the project environment and available through the isolated environment, together with the Open Babel Xrender dependency. No system CUDA driver or global Python modification was needed. The measured GNINA command took about 45 seconds on this host; this is one observation, not a runtime guarantee. The recorded installation and shared caches occupied about 8.3 GiB; this is an observed footprint, not a minimum storage requirement. Downloads require network access; no privacy audit is claimed.

Docking/CNN scores are not measured affinity, IC50 or efficacy. PoseBusters checks configured plausibility and reference criteria; PLIP reports rule-based contacts in supplied coordinates. This one known-pocket fixture does not establish general docking accuracy, cross-platform compatibility or biological activity. No independent prediction RMSD calculation was performed: the report retains PoseBusters' configured RMSD-threshold booleans. Neutral ligand choice, stripped water/glycans and unmodeled termini constrain interpretation.

Published resources

See the companion Recipe for the workflow sequence and prerequisites.

Primary references