MolecularDiffusion.runmodes.analyze.druglike

Drug-likeness descriptors for generated molecules.

Backs MolCraftDiff analyze metrics --metrics druglike. Everything here is computed from the molecule alone – no reference molecule, no receptor – which is why it is a separate set from similarity3d.

Columns follow what the pocket-conditioned literature reports:

  • RDKit descriptors – QED, SA, LogP, fsp3, MW, HBD, HBA;

  • lipinski – how many of the five Lipinski rules a molecule obeys (others/targetdiff/utils/evaluation/scoring_func.py:obey_lipinski);

  • pains_pass – free of PAINS-A substructures;

  • ring_filter_pass – no ring larger than 6 that is not aromatic-fused, the 2D ring sanity filter from DiffLinker;

  • ring statistics – counts plus which ring sizes are present;

  • rdkit_rmsd_* – distance from the generated pose to UFF-optimised RDKit conformers. Expensive (embeds n_conf conformers per molecule), so it is opt-in via --rdkit-rmsd.

Attributes

RING_SIZES

Functions

compute(mol[, with_rdkit_rmsd, n_conf])

All drug-likeness columns for one molecule.

rdkit_rmsd(mol[, n_conf, random_seed])

RMSD from the generated pose to n_conf UFF-optimised conformers.

Module Contents

MolecularDiffusion.runmodes.analyze.druglike.compute(mol, with_rdkit_rmsd=False, n_conf=20)

All drug-likeness columns for one molecule.

mol must already be sanitized. Returns a flat dict ready for a DataFrame row.

MolecularDiffusion.runmodes.analyze.druglike.rdkit_rmsd(mol, n_conf=20, random_seed=42)

RMSD from the generated pose to n_conf UFF-optimised conformers.

Mirrors others/targetdiff/utils/evaluation/scoring_func.py:get_rdkit_rmsd. Returns (min, median, max), or (None, None, None) when embedding fails.

MolecularDiffusion.runmodes.analyze.druglike.RING_SIZES