inSilico Dynamics · Computational biophysics
Carbonic Anhydrase II Virtual Screening Pipeline: A Documented Negative Result
A public CA-II screening benchmark reports weak discrimination against property-matched decoys, including preparation losses and the limits of re-docking.
Question and structure
Can a reproducible docking pipeline rank known carbonic anhydrase II (CA-II) binders above property-matched decoys? Acetazolamide is the positive control. The starting structure is human CA-II with acetazolamide, PDB 3HS4, X-ray resolution 1.10 Å. Chain A and the first alternate location were used; zinc was kept and waters/glycerol removed.
Pipeline and library
The Nextflow pipeline separates receptor preparation, ligand preparation, docking, ranking and re-docking:
- PREP_RECEPTOR: clean the chain while retaining zinc; extract reference ligand, define the box and prepare receptor PDBQT with Open Babel at pH 7.4.
- PREP_LIGANDS: prepare ligand PDBQT files from names, CIDs or SMILES, with a preparation manifest.
- DOCK: smina runs a task per ligand and records its best affinity score.
- RANK: produce ranked scores, ROC-AUC and enrichment factors.
- REDOCK / REDOCK_SUMMARY: re-dock acetazolamide per seed; compare with the crystal pose and summarize the RMSDs and pass fraction.
The library contains 568 molecules: 142 ChEMBL actives and 426 property-matched decoys. The pipeline docked 538 molecules (129 actives and 409 decoys); 30 failed preparation and were excluded from the metrics.
Results across scoring variants
| Variant | ROC-AUC (95% CI) | EF1% (95% CI) | EF10% (95% CI) | Re-dock top pose < 2 Å | Best of modes < 2 Å | Median top-pose RMSD (Å) |
|---|---|---|---|---|---|---|
| vina | 0.548 (0.491–0.606) | 1.39 (0.0–3.297) | 1.39 (0.879–1.873) | 0% | 90% | 5.38 |
| vinardo | 0.622 (0.569–0.675) | 1.39 (0.0–2.514) | 1.85 (1.266–2.317) | 90% | 100% | 1.89 |
| vinardo_like | 0.588 (0.535–0.642) | 2.09 (0.0–3.371) | 1.31 (0.754–1.736) | 0% | 100% | 5.41 |
| zn1 | 0.537 (0.485–0.591) | 1.39 (0.0–3.449) | 1.00 (0.553–1.445) | 0% | 100% | 5.41 |
| zn2 | 0.465 (0.413–0.517) | 1.39 (0.0–3.321) | 0.62 (0.241–1.09) | 0% | 20% | 6.19 |
| Score | ROC-AUC (95% CI) | EF1% | EF10% |
|---|---|---|---|
| Heavy atoms | 0.475 (0.419–0.531) | 0.70 | 0.93 |
| Molecular weight | 0.559 (0.504–0.615) | 2.09 | 1.54 |
Reading the negative result
ROC-AUC 0.5 is chance. The best variant, vinardo, reaches 0.622 (95% CI 0.569–0.675), only weakly above chance. In this setup the ranking is close to chance: docking scores barely separate known CA-II binders from property-matched decoys. EF1% confidence intervals start at 0 for every variant.
This is not a pipeline failure: the pipeline ran 538/538 ligands to a score and the metrics reproduce from the files. It is a negative finding about discrimination in this setup. A screening protocol can be measured against labelled data and the poor result reported alongside the protocol. Docking scores are not affinities.
Limitations
- Zinc coordination. Vina-type scores have no physical metal-coordination term. The sulfonamide is docked neutral although the zinc-bound form is an anion. The experimental zn1/zn2 variants add a heuristic Zn–N reward with arbitrary, unfitted weights and did not improve ranking. No Zn–N-restrained pose selection was performed.
- Close to chance. Best ROC-AUC is 0.622, and EF1% CIs include 0. This is not demonstrated screening capability.
- Size confound. Heavy-atom count alone gives AUC 0.475 and molecular weight 0.559. The gap between the latter and the best docking AUC is small and their intervals overlap.
- Preparation failures are not random. 30 of 568 ligands failed: 13 of 142 actives (9.2%) and 17 of 426 decoys (4.0%). Actives were lost more often, so the docked set differs from the designed library.
- Re-docking is circular. The box and reference both use crystal acetazolamide; box padding and scoring were chosen to re-dock that ligand. Vinardo with 4 Å padding gives 9 of 10 seeds below 2 Å, but acetazolamide itself ranks 455/538 in the screen. Re-docking one ligand says nothing about the other CA-II ligands.
- Protonation and water. Histidine protonation and hydrogens come from Open Babel at pH 7.4; the zinc-bound water is removed.
- Decoy labels. Decoys are approved drugs, not experimentally confirmed non-binders. Property matching limits bias but does not remove it.
- Docker profile not verified. No Docker was available on the development machines.
- Test coverage. CI runs the pytest suite, without Nextflow, smina or Open Babel; the workflow itself was checked by running it.
Source and reproducibility
This page summarizes the public repository README. Read the methods, derived data and limitations in the source repository.