--- license: cc-by-4.0 pretty_name: PDE5 Inhibitor Designs (PDB 2H42) task_categories: - tabular-regression - graph-ml - other size_categories: - 1K **Why this target matters.** PDE5 inhibitors are among the most widely prescribed drugs in the world, and the target remains actively pursued in pulmonary arterial hypertension and cardiac indications far beyond its original use. 2,090 de novo small-molecule designs docked into the catalytic pocket of human cGMP-specific phosphodiesterase 5 (PDE5A), using the sildenafil co-crystal [2H42](https://www.rcsb.org/structure/2H42) (PDE5 catalytic domain with sildenafil, X-ray, 2.3 Å) as the receptor. PDE5 hydrolyses cGMP; its inhibitors (sildenafil, tadalafil, vardenafil) are established drugs for erectile dysfunction and pulmonary arterial hypertension, and the target continues to draw interest in cardiac, renal and neurodegenerative indications. This is a **score-filtered** set — every design here already passed a docking-score cut, so the score range is deliberately narrow and is not a random sample of the generative run. ## Contents | File | Description | |---|---| | `designs.csv` | Index of all 2,090 designs: ID, SMILES, docking score, physicochemical properties, per-design contact residues | | `ligands.sdf` | All 2,090 docked poses as a single 3D SDF, with bond orders and formal charges from the source SMILES and coordinates from the pose | | `structures/` | Per-design ligand-only PDB (`.pdb`), 2,090 files, in the receptor frame | | `receptor_2H42.pdb` | The single shared receptor used for every docking run | The receptor is byte-identical across all 2,090 runs, so it is stored once. Ligand poses are in the receptor's coordinate frame — load `receptor_2H42.pdb` together with any ligand file or SDF record, with no superposition needed. **Receptor detail.** Chains `A`, `B`, `C` (PDE5A catalytic domain, one copy each). Docking was done in **chain A**, from which sildenafil was deleted; sildenafil (`VIA`) is still present in chains B and C, and the Zn/Mg cofactors are retained in all three. Three chain-A active-site waters were kept (HOH 75, 86, 127) and were part of the docking receptor — `HOH 86` is contacted by 37% of the designs. Every design contacts chain A only. ## `designs.csv` columns `design_id`, `smiles`, `pose_smiles`, `stereo_check`, `vina_score` (kcal/mol), `mw`, `clogp`, `tpsa`, `qed`, `hbd`, `hba`, `rotatable_bonds`, `heavy_atoms`, `rings`, `aromatic_rings`, `formal_charge`, `n_contacts`, `contact_residues` (`chain:residue`, within 4 Å of the pose, `;`-separated). Rows are sorted by docking score, best first. `smiles` is the structure as emitted by the generative run; `pose_smiles` is re-derived from the 3D coordinates of the deposited pose, so it carries the stereochemistry actually docked. See the stereochemistry note under Caveats. ## Set profile | | min | median | max | |---|---|---|---| | AutoDock Vina score (kcal/mol) | −16.2 | −13.8 | −13.5 | | Molecular weight | 373.4 | 423.5 | 496.4 | | cLogP | −2.3 | 2.3 | 4.0 | | TPSA (Ų) | 72.8 | 111.2 | 136.7 | | QED | 0.31 | 0.57 | 0.79 | | Heavy atoms | 28 | 31 | 35 | | Rotatable bonds | 2 | 4 | 5 | | Contact residues per design | 9 | 14 | 19 | 1,869 of the 2,090 SMILES are unique; the remainder are alternative poses of a shared 2D structure (design IDs ending `a`/`b` denote poses of the same molecule). **All 2,090 satisfy Lipinski's rule of five** — a consequence of the upstream filter, not of the generator. 1,004 carry a net formal charge at the modelled protonation state (878 at +1, 58 at +2, 68 at −1). The set is chemically focused rather than diverse: 1,099 distinct Murcko scaffolds, but a single benzoxazinone–pyrazolotriazine chemotype (`O=C1COc2c(cccc2-c2ccn3ncnc3c2)N1`) accounts for 6.4% of designs, and close relatives of it dominate the top of the score ranking. ## Binding site Residues contacted by ≥50% of the designs, with frequency: `Q817` 100% · `V782` 100% · `F820` 100% · `L725` 99% · `L765` 99% · `F786` 99% · `N661` 96% · `D764` 77% · `T723` 77% · `H613` 76% · `I665` 75% · `L804` 60% · `N662` 60% · `I824` 56% · `S663` 52% This reproduces the canonical PDE5 inhibitor pharmacophore. **Q817**, the invariant glutamine that makes the bidentate hydrogen bond to sildenafil's pyrimidinone, and **F820**, the hydrophobic-clamp phenylalanine that π-stacks against the inhibitor scaffold, are each contacted by 100% of the designs — the same two-point anchor used by all marketed PDE5 inhibitors. The hydrophobic pocket (V782, L765, L725, F786, I665, L804) is filled at ≥60%. Contacts at H613 and D764 reflect designs reaching toward the metal-binding site; a further 37% engage the conserved active-site water `HOH 86`. Median 14 contact residues per design. ## Usage ```python from datasets import load_dataset ds = load_dataset("Tc-43/PDE5_Inhibitor_Designs_2H42")["train"] top = ds.sort("vina_score")[:10] ``` ```python # 3D poses from rdkit import Chem mols = [m for m in Chem.SDMolSupplier("ligands.sdf") if m is not None] mols[0].GetProp("vina_score") ``` ```bash # single design in its receptor pymol receptor_2H42.pdb structures/52_2H42_II2608021409194979126914a.pdb ``` ## Caveats Docking scores are AutoDock Vina estimates from a single rigid-receptor run against one crystal conformation. They rank poses, not affinities, and the dynamic range here is narrow (−16.2 to −13.5) because the set is already score-filtered. Poses do not model induced fit, and protonation states are as assigned by the docking preparation. **Stereochemistry.** 1,699 designs carry no stereochemical assignment in their source SMILES. Of the 391 that do, the configuration read back from the docked 3D coordinates matches the declared one for 272 and **differs for 119**. The `stereo_check` column flags each row (`not_declared` / `match` / `pose_differs`). Where they disagree, treat `pose_smiles` as describing the geometry in `ligands.sdf` and `structures/`, and the declared `smiles` as the intended 2D identity — do not assume the deposited pose is the intended enantiomer. **Nothing in this set has been synthesised or assayed.** No compound has measured PDE5 inhibition, selectivity over PDE6 or PDE11 (the basis of the visual and off-target liabilities of marketed PDE5 inhibitors), or any ADME/tox data. These are computational starting points for triage and medicinal-chemistry review, not validated inhibitors. ## Citation ```bibtex @misc{pde5_inhibitor_designs_2h42, title = {PDE5 Inhibitor Designs (2H42 / GA-II)}, author = {TechnetiumTx}, year = {2026}, url = {https://huggingface.co/datasets/Tc-43/PDE5_Inhibitor_Designs_2H42} } ``` Receptor structure: PDB [2H42](https://www.rcsb.org/structure/2H42) — *Crystal structure of PDE5 in complex with sildenafil*.