MolecularDiffusion.modules.models.ipdiff¶
IPDiff: TargetDiff’s denoiser conditioned on a pretrained binding prior.
Only two genuinely new pieces live here – BAPNet (the frozen
“IPNet” interaction prior) and IPDiffScorePosNet3D (a thin subclass
of KGDiff’s ScorePosNet3D adding prior conditioning and prior shifting).
The backbone (uni_transformer.py) and its building blocks (common.py)
are reused by import from modules/models/kgdiff/, having been verified
byte-identical to IPDiff’s copies modulo import paths and docstrings.
See docs/model_integrations/ipdiff/INTEGRATION_PLAN.md.