SCOPE-BENCH: Scaffold-Cluster Out-Of-Distribution Performance Evaluation Benchmark

SCOPE-BENCH is a rigorous out-of-distribution (OOD) benchmark for molecular property prediction. Unlike conventional scaffold splits, SCOPE-BENCH creates structurally disjoint source and target domains by clustering molecules based on physicochemical descriptors, blocking shortcut learning, and revealing true extrapolation abilities.

📊 Dataset Splits (As used in the NeurIPS 2026 paper)

Split	Files	Clusters	Samples	Purpose
Source (training)	a0.csv, a1.csv, a6.csv, a8.csv, a9.csv, a11.csv	6	94,562	Supervised pre‑training + multi‑source adaptation pool
Validation	a10.csv	1	18,326	Hyperparameter tuning
Target (test)	a2.csv, a3.csv, a4.csv, a5.csv, a7.csv	5	19,894	Strict OOD evaluation (zero‑shot extrapolation)
Fine‑grained tasks	scaffold_datasets1/*.csv (each ≥200 samples)	16	varies	Independent zero‑shot evaluations (Table 2)

🔬 Molecular Properties

Each CSV file contains the following columns (based on the QM9 dataset):

Column	Description	Unit
SMILES	Simplified molecular input line entry system	–
HOMO	Highest Occupied Molecular Orbital energy	eV
LUMO	Lowest Unoccupied Molecular Orbital energy	eV
GAP	HOMO–LUMO gap (LUMO – HOMO)	eV
(other columns)	Additional QM9 properties (e.g., dipole moment, polarizability, etc.)	various

🗂️ File Structure

SCOPE-BENCH/ ├── a0.csv ... a11.csv # 12 cluster files ├── scaffold_datasets1/ # 16 independent target scaffolds ├── README.md ├── dataset_info.json └── LICENSE