GraphCliff: trained checkpoints for 30 MoleculeACE activity cliff targets

Trained GraphCliff model checkpoints for the 30 MoleculeACE regression targets (Ki / EC50).

GraphCliff is a graph neural network for activity cliff prediction. Each filter layer runs a short-range edge-aware path (GINE) and a long-range spectral path (Chebyshev polynomial) in parallel and combines them with a learned per-atom gate, so that a low gate suppresses long-range smoothing and preserves the local signal where a single-atom modification causes a large activity change.

 Contents

--------

 graphcliff_ckpt.tar.gz (636 MB, MD5 708921f6503864393107a332f4ac32fc)

250920_final_ckpt/<TARGET>/<TIMESTAMP>/best_model.pt 

30 checkpoints, one per target. Each file holds a single key, "model_state_dict", with 6,021,198 float32 parameters, matching the default architecture (hidden_size=256, num_layers=3, groups=4, mid_K=3).

Usage

 -----

tar xzf graphcliff_ckpt.tar.gz

mv 250920_final_ckpt ckpt

Place the extracted directory at the root of the GraphCliff repository. main.py loads ckpt/<TARGET>/*/best_model.pt automatically and runs evaluation instead of retraining.

Source code: https://github.com/dmis-lab/GraphCliff