MolBasis
A.IMPROVE2-termCompSol · n=1394

CompSol · ΔG_hyd

Hydration free energy (large) · kcal/mol

closed-form kernel
ŷ = β₀ + β₁ · z(cavity^0.50) + β₂ · z(g_polar^0.25)
0.886
nested-CV r
in-sample, winner's-curse-safe
0.659
scaffold-blind r
25% held-out OOD
+24%
blind MAE-vs-NULL lift
gate = +30%
1.04
MAE (kcal/mol)
grouped-LOO

In-sample → out-of-distribution

How the correlation holds when moving to unseen scaffolds.

nested-CV r (in-sample)0.886
scaffold-blind r (OOD)0.659
Blind MAE-vs-NULL lift24%

Vertical marker = the +30% production gate.

Blind calibration slope

Slope of measured vs predicted on the blind set. 1.0 = no magnitude compression.

1.00
slope (ideal 1.0)
Below +30% blind gate

What it is

The largest hydration cohort (n=1394). Cavity plus polar screening — the classic solvation decomposition. Same overfitting signature as ESOL on blind held-out.

The physics

Hydration = cavity formation cost + electrostatic screening reward. Both ¼-integer exponents; physically the canonical two-term solvation model.

upper / lower bound

Published nested-CV r=0.886 is the full-data upper bound; 1-term fallback is the honest blind floor.

caveat

Same 2-term gate failure as ESOL; falls back to a 1-term cavity^0.75 form on the 75% subsample. Blind lift +24.0% (below the +30% gate).

How the methods compare

The physics gold standard — alchemical FEP — reaches ~1 kcal/mol but needs a full molecular-dynamics simulation per molecule. SMD and COSMO-RS need a DFT job. This kernel lands in the same order of magnitude from a single closed-form solve on one σ-profile: it trades a little accuracy for orders of magnitude less compute, and stays fully interpretable.

Reported error

Each method's own reported error in kcal/mol — lower is better. Different datasets and splits, so this is an orientation, not a controlled benchmark.

COSMO-RS / COSMOtherm0.50 kcal/mol · MAE (typical)
SMD (implicit-solvent DFT)0.80 kcal/mol · MUE (neutrals)
CompSol · ΔG_hyd · this kernel1.04 kcal/mol · MAE
Alchemical FEP (explicit-solvent MD)1.07 kcal/mol · MAE
D-MPNN (Chemprop)1.20 kcal/mol · RMSE

Lower error isn’t the whole story — read it against input cost and requirements below.

Input cost

Per-molecule compute/data burden to predict a new molecule — shorter is cheaper.

D-MPNN (Chemprop)2D structure (pencil / inference)
CompSol · ΔG_hyd · this kernelone QM calculation (SCF)
SMD (implicit-solvent DFT)one QM calculation (SCF)
COSMO-RS / COSMOthermone QM calculation (SCF)
Alchemical FEP (explicit-solvent MD)MD simulation / crystal prediction

What each method needs

External dependencies each method carries. An amber dot means the method requires it — fewer dots means fewer things to procure or that can go wrong.

Method3D geometryMD / samplingtraining corpusa measured valueproprietary paramsdeps
CompSol · ΔG_hyd · this kernel1
Alchemical FEP (explicit-solvent MD)2
SMD (implicit-solvent DFT)1
COSMO-RS / COSMOtherm2
D-MPNN (Chemprop)1

This kernel needs only a 3D geometry for its one SCF — no MD, no training corpus, no measured value, no proprietary software.

Competing methods

How this property is predicted elsewhere — with the input each method needs (a key differentiator) and the literature reference. Numbers are each method’s own reported figure on its own benchmark, so they are indicative, not a head-to-head on an identical split.

MethodClassReported performanceInput neededReference
CompSol · ΔG_hydthis kernelclosed-formPearson r 0.886 (nested-CV) · 0.659 scaffold-blind · MAE 1.04 kcal/molone DFT SCF σ-profile · no training set · no MDMF-FQSL (this lab)
Alchemical FEP (explicit-solvent MD)physicsMAE 1.07, RMSE 1.43 kcal/mol (GAFF/TIP3P, FreeSolv n=621)3D geometry + force field + molecular-dynamics samplingMobley & Guthrie, FreeSolv, J. Comput. Aided Mol. Des. 2014
SMD (implicit-solvent DFT)physicsMUE 0.6–1.0 kcal/mol for neutrals (6-31G*); ~2.5 kcal/mol over full MNSolDFT SCF + continuum, per soluteMarenich, Cramer & Truhlar, J. Phys. Chem. B 2009
COSMO-RS / COSMOthermphysics≈0.4–0.6 kcal/mol on well-parametrised sets (best-in-class physics)DFT σ-profile + proprietary parametrisationKlamt, J. Phys. Chem. 1995
D-MPNN (Chemprop)ML / GNNRMSE ≈ 1.2 kcal/mol on FreeSolv (MoleculeNet split)2D molecular graph + a labelled training setYang et al., J. Chem. Inf. Model. 2019

Metrics are as published by each method on its own dataset (different splits, different cohorts) — treat them as an orientation of the landscape, not a controlled benchmark. The differentiator for this kernel is the input column: a single closed-form solve from one σ-profile, with no training corpus, no MD, and no measured melting point.

Descriptors used

cavitycavity_area_aa2 — solvent-accessible surface area of the COSMO cavity (Ų)
g_polarg_polar_kcal — polar screening (electrostatic solvation) energy (kcal/mol)

Version history

  1. 2026-05-29v0.91.1 ship

    ⭐ Unique-pair winner, r=0.886.

  2. 2026-06-06Overfit diagnostic

    Same signature as ESOL — 2-term overfits on blind; 1-term fallback is honest floor.