MolecularDiffusion.data.component.diffint_prep¶

Build a DiffInt complex from a raw protein PDB + reference ligand SDF.

DiffInt is DiffSBDD with a CA pocket whose node set is extended by H-bond pseudo-atoms: two per detected protein-ligand hydrogen bond, placed at 1/3 and 2/3 along the donor-acceptor vector, carrying two extra one-hot channels (DD = 20, AC = 21) on top of the 20 amino-acid classes. They are conditioning – appended to the pocket, never diffused.

This module is the single implementation of that construction. Both the offline converter (docs/model_integrations/diffint/scripts/ convert_dataset.py) and the on-the-fly generation path (DiffIntPocketGenerator with pocket_pdb + ref_sdf) call complex_from_files(), so the two can never drift apart.

## Ported from test_single.py, NOT from lightning_modules.py

Upstream ships two novel-PDB paths and only one of them is correct:

lightning_modules.py:852-890 (generate_ligands / prepare_pocket) builds the 22-wide one-hot by zero-padding the 20-class CA encoding and then never adds the particles. Every DD/AC column stays zero, so the model silently degrades to plain DiffSBDD – the paper’s entire contribution is switched off, with no error.
test_single.py:141-192 (process_data + process_data_h) rebuilds the augmented pocket properly via hbond_double.hbond_create. That is what is ported here.

## Two upstream data bugs are fixed rather than reproduced

hbond_double.py:41-43 appends a dummy empty row to int_id when a complex has zero H-bonds, which produces a ragged array. int_id is only consumed by the auxiliary interaction loss, which is out of scope (it is commented out in the release and the shipped checkpoint was trained without it), so it is simply not built here.
process_crossdock.process_ligand_and_pocket derives lig_coords from every atom but lig_one_hot from a filtered atom list, and its H filter (a.element != 'H') hits an attribute RDKit’s Atom does not have. Both only matter for SDFs carrying explicit hydrogens; here hydrogens are stripped from the molecule up front, which is what upstream’s --no_H preprocessing achieved anyway.

Heavy dependencies (Biopython, ODDT/OpenBabel, RDKit) are imported inside the functions and gated by optional.require_modules, so importing this module never fails on a bare install.

Attributes¶

`CLASS_AC`
`CLASS_DD`
`DIFFINT_POCKET_VOCAB`
`NUM_POCKET_CLASSES`

Functions¶

`ca_pocket`(pdbfile, lig_coords[, dist_cutoff])	One node per CA of every standard residue within `dist_cutoff`.
`complex_from_files`(→ Dict[str, Any])	PDB + reference SDF -> the raw arrays one converted db row holds.
`hbond_particles`(→ Tuple[numpy.ndarray, numpy.ndarray])	2 pseudo-nodes per H-bond, at 1/3 and 2/3 along it.

Module Contents¶

MolecularDiffusion.data.component.diffint_prep.ca_pocket(pdbfile: str, lig_coords: numpy.ndarray, dist_cutoff: float = 8.0)¶

One node per CA of every standard residue within dist_cutoff.

Port of process_crossdock.process_ligand_and_pocket’s ca_only=True branch (process_crossdock.py:65-80). Returns (coords (N, 3) float32, classes (N,) int64 in 0..19).

MolecularDiffusion.data.component.diffint_prep.complex_from_files(pdb_file: str, sdf_file: str, dist_cutoff: float = 8.0) → Dict[str, Any]¶

PDB + reference SDF -> the raw arrays one converted db row holds.

“Novel pocket” means “novel protein + a reference ligand pose”: the SDF is both the 8 A pocket-selection reference and the ligand ODDT detects the H-bonds against. That is upstream’s own framing (test_single.py) and is unavoidable – the interactions are protein<->ligand.

Pocket nodes come back residues first, then particles, which is the order num_pocket_nodes (residues only) assumes.

MolecularDiffusion.data.component.diffint_prep.hbond_particles(protein, ligand) → Tuple[numpy.ndarray, numpy.ndarray]¶

2 pseudo-nodes per H-bond, at 1/3 and 2/3 along it.

Port of hbond_double.py:8-43. Returns (coords (2K, 3) float32, classes (2K,) int64 in {20, 21}), with the two particles of a bond adjacent, the hbond_acceptor_donor(protein, ligand) block first.

data/component/diffpharma_prep.py:hbond_particles runs the same ODDT detection but emits DiffPharma’s geometry (3 particles at 1/4, 1/2, 3/4 with a 3-class vocabulary), which is not interchangeable with DiffInt’s – different node count, different one-hot width, different checkpoint. What is genuinely shared (the np.in1d shim, the protein.protein = True / ligand.removeh() preconditions) is reproduced above rather than imported for a side effect.

MolecularDiffusion.data.component.diffint_prep.CLASS_AC = 21¶

MolecularDiffusion.data.component.diffint_prep.CLASS_DD = 20¶

MolecularDiffusion.data.component.diffint_prep.DIFFINT_POCKET_VOCAB: List[str] = ['A', 'C', 'D', 'E', 'F', 'G', 'H', 'I', 'K', 'L', 'M', 'N', 'P', 'Q', 'R', 'S', 'T', 'V', 'W',...¶

MolecularDiffusion.data.component.diffint_prep.NUM_POCKET_CLASSES = 22¶