Qm9 Property Prediction Eval

qm9-property-prediction-eval

Molecular Representations in Implicit Functional Space via Hyper-Networks — Wang et al. (2026) (arXiv:2601.22327, 2026)

What this evaluates

Assesses a model's ability to predict scalar quantum chemical properties from molecular structures. It evaluates accuracy on five key electronic and vibrational targets using mean absolute error and average ranking.

Datasets

• QM9 — total 130000; splits: train (100000), val (18000), test (13000)

Metrics

• MAE (primary) — range: other
  • Mean Absolute Error between predicted and true scalar values for each of the five properties (HOMO, dipole, R2, polarizability, ZPVE). Lower is better.
• A.R. — range: other
  • Average Ranking across all five properties, computed by ranking methods per property and averaging ranks (ties receive averaged ranks). Lower is better.

Input / output format

Input: Molecular graph/structure with atomic coordinates and types.

Output: Scalar predictions for HOMO energy, dipole moment, electronic spatial extent, polarizability, and zero-point vibrational energy.

Scoring recipe

mae = np.mean(np.abs(pred - true))
ranks = [rankdata(-mae_per_property) for property in properties]
ar = np.mean(ranks)

Common pitfalls

• MAE units vary significantly across properties (meV, mD, mα₀², etc.), making cross-property comparison difficult without normalization.
• Average Ranking (A.R.) is computed by ranking methods per property and averaging; ties receive averaged ranks, which can skew results if many methods cluster.

Evidence (verbatim from paper)

Five scalar targets are considered: ε_HOMO, μ, R^2, α, and ZPVE, with mean absolute error (MAE) as the evaluation metric.

Citation

@misc{wang2026molfield,
  title={Molecular Representations in Implicit Functional Space via Hyper-Networks},
  author={Wang et al. (2026)},
  year={2026},
  note={arXiv:2601.22327}
}

• arXiv: 2601.22327