Fragment Extraction#
enerzyme extract ranks structures by local force uncertainty and extracts capped substructures for QM relabeling. Algorithm: enerzyme/tasks/extractor.py (RDKit).
Command#
enerzyme extract -c extract.yaml -o out/ -m model_dir/ -mc config.yaml [-s]
-s skips prediction if pickles already exist in output_dir.
Configuration#
Datahub:
data_path: candidates.pkl
features:
N:
Q: Q
Ra: Ra
Za: Za
neighbor_list: full
preload: true
Extractor:
reference_mol_path: cluster.mol
fragment_per_frame: 6
local_uncertainty_radius: 5.0
fragment_radius: 4.0
n_centers: 2
Trainer:
inference_batch_size: 4
Parameters#
reference_mol_pathMOL/SDF defining connectivity. One molecule applies to all frames; multiple entries align by index.
local_uncertainty_radiusNeighborhood (Angstrom) for aggregating atomic force variance (J. Chem. Inf. Model. 2024, 64, 6377-6387).
fragment_radiusAll heavy atoms within this radius of the uncertain center are included.
fragment_per_frameTop-N fragments per trajectory frame.
n_centersNumber of uncertainty centers per frame.
Prerequisites#
Uncertainty-capable model (shallow ensemble or committee)
Prediction step on the same
Datahubfeatures as trainingConsistent charge and topology with reference molecule
Pipeline position#
Typical Enerzymette round: MD trajectory → predict → extract → annotate → train.
Outputs#
Fragment SDF (e.g. FF02-SpookyNet-19_fragments.sdf) fed to annotate via Supplier.path.