SupraDB-CavityScore

What it is

SupraBench/SupraDB-CavityScore is a SupraBench feature dataset generated from the SupraEngineering compute pipeline. Pipeline position: Phase 1 docked-subset GLIDE feature computation by compute_all_features; Phase 2 publish split.

The table is designed to join with the Phase 0 SupraDB-GEOM identity table and the other feature datasets through inchikey.

Schema

Column	Dtype	Units	Meaning
inchikey	string	none	Primary InChIKey join key shared across SupraDB-GEOM, LigandScore, CavityScore, and PoseFeat.
name	string	none	Guest name from the canonical SupraDB-GEOM identity table.
smiles	string	none	Canonical largest-fragment SMILES from the canonical SupraDB-GEOM identity table.
source	string	none	Winning source from the canonical SupraDB-GEOM identity table after priority deduplication.
docked	bool	none	Boolean flag emitted only with --attempted for CavityScore and PoseFeat; false rows were submitted for docking but absent from the docked pickle and have empty feature values.
batch	string	none	Optional publish batch identifier emitted for CavityScore and PoseFeat when --batch is provided; every row from the publish run receives the supplied value.
S_occupancy	float32	unitless score	Cavity occupancy score from the hydrophobic occupied volume of the selected docked pose.
S_portal	float32	unitless score	Portal compatibility score from cation-portal distance, charge accessibility, hydrogen bonding, and orientation.
S_accessibility	float32	unitless score	Charge accessibility score from solvent-accessible positive atoms in the selected docked pose.
S_orientation	float32	unitless score	Orientation score measuring whether the positive center points from the cavity center toward the near portal.

When --attempted is used for CavityScore or PoseFeat, docked marks whether a submitted guest is present in the docked pickle. Rows with docked=False were attempted but missing from the pickle, so their feature columns are empty/NaN. This flag is not emitted without --attempted and is never emitted for LigandScore.

When --batch <id> is used with --attempted, batch is emitted immediately after docked for CavityScore and PoseFeat and contains the supplied id for every row from that publish run, including both docked rows and docked=False no-pose rows. This column is not emitted without --batch and is never emitted for LigandScore.

Schema meanings are summarized from SupraEngineering/src/features_lib.py and SupraEngineering/src/constants.py. Local feature-code docstrings: Feature computation for CB[7]-guest: the 13 mechanism scores and the 24-dim pose pose_features: Returns (vec24 in POSE_FEATURES order, derived dict for mechanism scores). finalize_pose_vec: Fill PC-dependent + tpsa-dependent fields and the PoseScore (Pose_Energy).

Join key

inchikey is the sole join key for SupraDB-GEOM, SupraDB-LigandScore, SupraDB-PoseFeat, and SupraDB-CavityScore. Downstream loaders should join on this column and treat the feature values as produced by the pipeline order in constants.SCORE_NAMES and constants.POSE_FEATURES.

Provenance

• Pipeline position: Phase 1 docked-subset GLIDE feature computation by compute_all_features; Phase 2 publish split.
• Source pickle: data/dock_b4/dock/features/scores.pkl.
• Computation environment: CRC.
• Docking/software context: GLIDE 2025u2 / aISS fallback as documented in the integration spec.
• Pose collapse: pose-collapse=highest Boltzmann weight; PoseFeat keeps the real pose at np.argmax(boltz) and records its boltzmann_weight and delta_e.
• Row count: 3183.
• Exact publish.py command: /users/tma2/workspace/SupraDashboard/.venv/bin/python engineering/src/publish.py --dock-scores data/dock_b4/dock/features/scores.pkl --pose-feats data/dock_b4/dock/features/pose_feats.pkl --identity data/pool_full/guests.csv --attempted /tmp/b4_docked.csv --batch batch4 --out data/publish_b4.

Regeneration

Regenerate this dataset by rerunning SupraEngineering/src/publish.py with the same pipeline pickle input and --out target. Use --push only in an authenticated environment with HF_TOKEN set; local generation is fully offline by default.