MolecularDiffusion.modules.models.loqi.graph_utils

LoQI edge-graph construction: stereo augmentation + fully-connected edges.

Two functions, both ported verbatim (NVIDIA, Apache-2.0):

• derive_stereo_edges() – data_processing/utils_data.py::add_stereo_bonds. Emits LoQI’s auxiliary edge classes 5-8, which encode E/Z and R/S. These are not bonds: they are directed (class 8 deliberately asymmetric – the direction is the R/S signal), so the platform’s symmetric, upper-triangular bond_index/bond_type storage cannot hold them. The approved plan therefore re-derives them per batch from the geometry, using the same AssignStereochemistryFrom3D that produced them upstream.

• make_graph_fully_connected() – src/megalodon/data/batch_preprocessor_conf.py:8-40. Merges the directed real-bond + stereo edge list into a fully-connected directed edge list, materializing class 0 (“no bond”) for every remaining pair. Where a stereo edge lands on a real bond, coalesce(reduce="min") on the bond list first, then reduce="max" against the zero-filled complete graph, means the bond order wins – upstream’s behaviour, preserved.

Canonical edge vocabulary here (9 classes): 0=none 1=SINGLE 2=DOUBLE 3=TRIPLE 4=AROMATIC 5=E 6=Z 7=chirality(sym) 8=chirality(directed). Class 4 never occurs for LoQI: the pipeline kekulizes.

Attributes

CHI_BONDS
N_EDGE_CLASSES

Functions

derive_stereo_edges(→ list[tuple[int, int, int]])	RDKit mol (with a conformer) -> list of (i, j, class) stereo edges.
make_graph_fully_connected(edge_index, edge_attr, batch)	Directed real+stereo edges -> fully-connected directed edges + 9 classes.

Module Contents

MolecularDiffusion.modules.models.loqi.graph_utils.derive_stereo_edges(mol, chi_bonds: tuple[int, int] = CHI_BONDS, ez_bonds: dict | None = None, from_3D: bool = True) → list[tuple[int, int, int]]

RDKit mol (with a conformer) -> list of (i, j, class) stereo edges.

Returns an empty list for an achiral, non-E/Z molecule. Duplicated entries are possible and intentional – upstream re-emits a stereocentre’s block once per incident bond, and the later coalesce deduplicates.

MolecularDiffusion.modules.models.loqi.graph_utils.make_graph_fully_connected(edge_index: torch.Tensor, edge_attr: torch.Tensor, batch: torch.Tensor)

Directed real+stereo edges -> fully-connected directed edges + 9 classes.

MolecularDiffusion.modules.models.loqi.graph_utils.CHI_BONDS = (7, 8)

MolecularDiffusion.modules.models.loqi.graph_utils.N_EDGE_CLASSES = 9