MolecularDiffusion.modules.models.diffsmol.shape_utils¶

Molecule -> surface mesh -> point cloud -> equivariant shape latent.

Ported from DiffSMol source/utils/shape.py (the pointAE_shape path only). This is the only module in the integration that needs optional dependencies, and it is precompute-only: training and generation read a cached .pt of tensors and never import this module.

Optional extras (pip install '.[shape]' --no-deps):

scikit-image – marching-cubes molecular surface
trimesh – mesh handling, area-uniform surface sampling, volume

Neither oddt nor openbabel is required at runtime. generate_surface_marching_cubes below is a faithful inline port of oddt.surface.generate_surface_marching_cubes (oddt 0.7), which is dead in any modern env: it imports skimage.measure.marching_cubes_lewiner, removed in scikit-image 0.19, and silently degrades to unusable. The only thing it needed openbabel for was a van-der-Waals radius lookup, which is hardcoded below (OPENBABEL_VDW_RADII) straight from openbabel’s table. Modern skimage.measure.marching_cubes is the Lewiner algorithm – it became the default when the _lewiner alias was retired – so this is numerically faithful to what upstream ran.

pytorch3d is not required either: upstream used it for exactly two calls, sample_points_from_meshes (replaced by trimesh.Trimesh.sample, the same area-uniform face sampling) and Meshes.get_bounding_boxes (which only fed the gradient guidance this integration does not port).

Frame convention – this is the correctness-critical part. The returned shape_center is the centroid of the sampled surface points, not the atom centroid and not the centre of mass. Upstream subtracts exactly this from both the point cloud and the ligand coordinates (shape.py:341 / shape_mol_dataset.py:115). The training task must subtract the same shape_center from its coordinates, or the latent and the coordinates end up in different frames and conditioning silently breaks.

Attributes¶

`OPENBABEL_VDW_RADII`
`POINT_CLOUD_SAMPLES`
`ang`

Functions¶

`check_vdw_radii`(→ None)	Assert `OPENBABEL_VDW_RADII` matches openbabel's table.
`generate_surface_marching_cubes`(→ Tuple[numpy.ndarray, ...)	Marching-cubes molecular surface as `(verts, faces)`.
`get_mesh`(symbols, coords[, scaling, probe_radius])	Molecular surface mesh (`trimesh.Trimesh`) via marching cubes.
`read_xyz`(path[, with_hydrogen])	Minimal `.xyz` reader returning `(symbols, coords)`.
`shape_from_atoms`(→ Dict[str, Any])	Full chain for one molecule.

Module Contents¶

MolecularDiffusion.modules.models.diffsmol.shape_utils.check_vdw_radii() → None¶

Assert OPENBABEL_VDW_RADII matches openbabel’s table.

A no-op (returns silently) when openbabel is not importable – it is not a dependency of this module, only the provenance of the numbers.

MolecularDiffusion.modules.models.diffsmol.shape_utils.generate_surface_marching_cubes(symbols: Sequence[str], coords: numpy.ndarray, scaling: float = 1.0, probe_radius: float = 1.4) → Tuple[numpy.ndarray, numpy.ndarray]¶

Marching-cubes molecular surface as (verts, faces).

Inline port of oddt.surface.generate_surface_marching_cubes. Like upstream it ignores hydrogens – consistent with the integration’s heavy-atoms-only decision, where the surface, the point cloud and the diffused atom set must describe the identical atoms.

MolecularDiffusion.modules.models.diffsmol.shape_utils.get_mesh(symbols: Sequence[str], coords: numpy.ndarray, scaling: float = 1.0, probe_radius: float = 1.4)¶: Molecular surface mesh (trimesh.Trimesh) via marching cubes.

MolecularDiffusion.modules.models.diffsmol.shape_utils.read_xyz(path: str, with_hydrogen: bool = False)¶

Minimal .xyz reader returning (symbols, coords).

with_hydrogen=False mirrors the dataset’s H filter so the surface, the point cloud and the diffused atom set describe the identical atoms. (generate_surface_marching_cubes drops H a second time regardless, matching upstream oddt behaviour.)

MolecularDiffusion.modules.models.diffsmol.shape_utils.shape_from_atoms(symbols: Sequence[str], coords: numpy.ndarray, shape_ae, num_samples: int = POINT_CLOUD_SAMPLES, device: str | torch.device = 'cpu') → Dict[str, Any]¶

Full chain for one molecule.

Returns {"shape_emb": (128, 3), "shape_center": (3,), "shape_volume": float}, all CPU tensors/floats ready to cache.

MolecularDiffusion.modules.models.diffsmol.shape_utils.OPENBABEL_VDW_RADII: Dict[str, float]¶

MolecularDiffusion.modules.models.diffsmol.shape_utils.POINT_CLOUD_SAMPLES = 512¶

MolecularDiffusion.modules.models.diffsmol.shape_utils.ang¶