---
license: other
pretty_name: SupraDB-GEOM
tags:
  - chemistry
  - supramolecular-chemistry
  - rdkit
  - cb7
configs:
  - config_name: default
    data_files:
      - split: train
        path: guests.csv
---

# SupraDB-GEOM

## What it is

`SupraBench/SupraDB-GEOM` is the Phase 0 guest-pool dataset for the SupraBench
compute pipeline. It contains the canonical guest identity table used as the
source of truth for downstream ligand-only scoring, docking-derived pose
features, cavity scores, and dashboard joins.

Rows are built from GEOM source SMILES, optional QM9 source SMILES, and optional
experimental CB[7] labeled guests. Each row starts from an input SMILES string,
is parsed with RDKit, desalted by keeping the largest fragment, converted to
canonical SMILES, and assigned an RDKit-derived InChIKey. Rows are deduplicated
by derived InChIKey with priority `labeled` > `geom_drugs` > `qm9`.

## Schema

| Column | Dtype | Units | Meaning |
|---|---|---|---|
| `inchikey` | string | none | RDKit InChIKey derived from the canonical largest-fragment SMILES; primary join key across all SupraDB datasets. |
| `name` | string | none | Guest name from the source CSV when present, otherwise the first 14 characters of `inchikey`. |
| `smiles` | string | none | RDKit canonical SMILES after largest-fragment desalting. |
| `logka` | float/string nullable | log10 association constant | Optional experimental CB[7] binding label from the source CSV; empty for unlabeled guests. |
| `source` | string | none | Winning source after priority deduplication: `labeled`, `geom_drugs`, or `qm9`. |
| `n_heavy` | int64 | atoms | Heavy-atom count of the canonical largest-fragment molecule used to derive `inchikey`. |

## Join key

`inchikey` is the sole join key for `SupraDB-GEOM`, `SupraDB-LigandScore`,
`SupraDB-PoseFeat`, and `SupraDB-CavityScore`. It is generated once here from
canonical largest-fragment SMILES so downstream compute, cache entries, 2D/3D
structure lookup, and dashboard records all use the same identifier.

## Provenance

- Pipeline position: Phase 0 pool builder, before CRC ligand scoring and GLIDE docking.
- GEOM source CSV: `data/geom_drugs_smiles.csv`.
- QM9 source CSV: `data/qm9_smiles.csv`.
- Labeled merge: 63 labeled rows were merged from the labeled CSV.
- Canonicalization: RDKit `MolFromSmiles` -> largest fragment by heavy atom count -> `MolToSmiles(canonical=True)` -> `MolToInchiKey`.
- Row count: 430904.
- Output rows by source:
- `labeled`: 63 rows
- `geom_drugs`: 301523 rows
- `qm9`: 129318 rows
- Downstream compute context: CRC; GLIDE 2025u2 / aISS fallback applies to later docked feature datasets, not this pool table.
- Pose collapse rule: not applicable for this dataset; later pose datasets keep the highest-Boltzmann-weight pose by default.
- Exact command: `/users/tma2/workspace/SupraDashboard/.venv/bin/python engineering/src/pool_builder.py --input data/geom_drugs_smiles.csv --qm9 data/qm9_smiles.csv --labeled engineering/data/guests.csv --out data/pool_full --workers 8`.