liuganghuggingface/moldata-simpoly_experiment

SimPoly PolyArena experimental benchmark

Microsoft Research reference room-temperature density and glass-transition benchmark for 130 polymers, collected from Bicerano.

Processed copy of simpoly_experiment from the molecular datasets collection. Structures are canonicalized and validated; labels are never averaged or filled. Prefer the original source for citation.

Where the data comes from

• Original data: https://github.com/microsoft/simpoly
• Original study: SimPoly: Simulation of Polymers with Machine Learning Force Fields Derived from First Principles

Data quality

Grade: caution (rough; rules on the hub page).

• info: experiment. Label origin from the source's tags.
• warn: multi-source. Compiled from several labs or publications; assay protocols can differ (Walters 2023: 45% of same-compound IC50 pairs from two papers differ by >0.3 log units).
• good: valid structures. Rows RDKit cannot parse (kept in quarantine).
• good: 100% label coverage. Median share of molecules labelled per task; missing labels are unmeasured, not negative.
• bad: 130 molecules per task. Median molecules with a label per task.
• good: 90% agree with polymetrix_tg. Polymer glass transition, experimental: on 62 shared molecules, 90% agree within ±15 K (MAE 6.13, r = 0.99).
• warn: 73% agree with openpoly. Polymer glass transition, experimental: on 48 shared molecules, 73% agree within ±15 K (MAE 18.38, r = 0.89).
• bad: 44% agree with radonpy_pi1070. Polymer density, experiment and MD: on 66 shared molecules, 44% agree within ±0.05 g/cm^3 (MAE 0.06, r = 0.98).

Tables

Config	Splits	Rows	Targets
000_simpoly_experiment	full	130	density (regression, g/cm^3 at room temperature), tg (regression, K)

data/ holds parseable rows with identical rows collapsed. observations/ holds every source row, including unparseable structures, for auditing. Added columns: SMILES_original, SMILES_canonical (RDKit; stereo, salts, and * kept), structure_key (JSON list of canonical components; use it to deduplicate or group), structure_status, wildcard_atom_count, source_row_first, observation_count, source_dataset, source_table. All original columns are strings, so censored values such as <10 survive; cast labels yourself.

Load

from torch_molecule.datasets import load_dataset
data = load_dataset("simpoly_experiment")

from datasets import load_dataset
ds = load_dataset("liuganghuggingface/moldata-simpoly_experiment", "000_simpoly_experiment")

Known issues

• Reference handbook measurements have heterogeneous experimental contexts.
• The same Bicerano sources contribute to other Tg compilations; check overlap before pooling.

License

MIT repository; underlying Bicerano handbook attribution retained

Citation

Cite the original study and data source:

• SimPoly: Simulation of Polymers with Machine Learning Force Fields Derived from First Principles