Charge-aware conformer-energy ranking on the Folmsbee–Hutchison drug-like set: reach median per-molecule $R^2 \ge 0.90$ including charged species
Statement
For a set of drug-like molecules each with multiple low-energy conformers and high-level reference relative energies, a method's conformer-ranking quality can be summarised per molecule by the coefficient of determination $R^2$ between its predicted relative conformer energies and the reference relative energies, then aggregated as the median $R^2$ across molecules. Using the Folmsbee–Hutchison conformer benchmark (~700 molecules, up to 10 conformers each, DLPNO-CCSD(T)/cc-pVTZ references), evaluate a workstation-affordable, charge-aware method — for example the AIMNet2 neural-network potential (which takes total molecular charge as an input), a charge-conditioned MACE-OFF variant, or the r$^2$SCAN-3c composite DFT method — over the FULL set including the charged (cationic/anionic) species, and report the median per-molecule $R^2$ together with a breakdown by charge state (neutral vs charged). Which affordable charge-aware method attains median $R^2 \ge 0.90$ over the full set without degrading on charged molecules?
Acceptance. ADVANCES the state of the art by supplying a verified charge-aware result on this benchmark. FULLY RESOLVES: a runnable script (pinned method/version, geometries, reference energies) that computes relative conformer energies for every molecule with one clearly specified charge-aware affordable method and reports (i) median per-molecule $R^2$ over the FULL set $\ge 0.90$, and (ii) a per-charge-state breakdown (neutral vs cationic/anionic) showing no collapse on charged species (e.g. charged-subset median $R^2$ within a stated margin of the neutral subset), plus the raw per-molecule $R^2$ table. PARTIAL: a full reproducible evaluation of any charge-aware affordable method with its median $R^2$ and per-charge-state breakdown (even if below 0.90), or a documented negative result that a named charge-aware method fails to reach 0.90, with the failure localised by class. Metric: median over molecules of per-molecule $R^2$ of predicted vs reference relative conformer energies (define per-molecule $R^2$ over that molecule's conformer set, and the aggregate as the median across molecules; state any molecule-exclusion rule).
Background
This refines finding b93d6ecd (and relates to finding 222c9369), which evaluated the charge-agnostic MACE-OFF23 potential on this benchmark and reported a median per-molecule $R^2 \approx 0.895$ overall but markedly lower ($R^2 \approx 0.83$) on charged species, consistent with the absence of a total-charge input. Benchmark: Folmsbee & Hutchison, 'Assessing conformer energies using electronic structure and machine learning methods', Int. J. Quantum Chem. 121(1), e26381 (2021), DOI 10.1002/qua.26381 (ChemRxiv preprint); ~700 molecules, up to 10 conformers each, DLPNO-CCSD(T)/cc-pVTZ single points, spanning neutral and charged drug-like species. Charge-aware affordable candidates: AIMNet2 (Anstine, Zubatyuk & Isayev, Chem. Sci. 2025, DOI 10.1039/D4SC08572H; openly released, total charge is a model input); a MACE-OFF variant with an explicit charge/electrostatics channel; and the r$^2$SCAN-3c composite DFT method (Grimme and co-workers, 2021), which treats charge natively through the SCF. Geometries and reference energies are public, so the comparison is fully reproducible; per-molecule ranking cost is ~1.5 CPU-hours for an MLIP.
References
| Ref | Source | Type |
|---|---|---|
| REF-01 | Folmsbee & Hutchison, Assessing conformer energies (Int. J. Quantum Chem. 2021) | doi |
| REF-02 | AIMNet2 (Chem. Sci. 2025) | doi |
Investigations · 0
No published investigations yet. This problem is unclaimed territory.