MolBasis
B.UNDERFIT1-termLlinás logP · n=40

Llinás · logP

Partition coefficient (small, high-quality) · log units

closed-form kernel
ŷ = β₀ + β₁ · z(sigma_variance^-0.75)
0.660
nested-CV r
in-sample, winner's-curse-safe
scaffold-blind r
25% held-out OOD
blind MAE-vs-NULL lift
gate = +30%
1.06
MAE (log units)
grouped-LOO

In-sample → out-of-distribution

How the correlation holds when moving to unseen scaffolds.

nested-CV r (in-sample)0.660
Blind MAE-vs-NULL lift

Vertical marker = the +30% production gate.

Blind calibration slope

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

0.88
slope (ideal 1.0)
Below +30% blind gate

What it is

A small (n=40), high-quality closed logP set. r=0.660 at +28% MAE-vs-NULL. Any 2-term pair collapses under nested CV (winner's curse on n=40, r_cv→0.57), so it stays 1-term and ships as B.UNDERFIT.

The physics

Same polarity-spread physics as the other logP kernels; the cohort is simply too small to support a 2-term fit.

caveat

B.UNDERFIT despite a respectable r — the honest call given n=40 and a modest %-vs-NULL.

How the methods compare

logP is largely solved for interpolation — atom-additive (Wildman–Crippen) and GNN models are cheap and accurate on drug-like space. This kernel's value isn't beating them on RMSE; it's a physically-grounded logP from the very same σ-profile that predicts the other ten properties, with no logP-specific training.

Reported error

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

clogP (Leo–Hansch fragments)0.45 log units · typical
D-MPNN (Chemprop)0.48 log units · RMSE
COSMO-RS / COSMOtherm0.57 log units · MAE (SAMPL7)
Wildman–Crippen (atom contributions)0.68 log units · σ (RMSE)
Llinás · logP · this kernel1.06 log units · MAE

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.

Wildman–Crippen (atom contributions)2D structure (pencil / inference)
clogP (Leo–Hansch fragments)2D structure (pencil / inference)
D-MPNN (Chemprop)2D structure (pencil / inference)
Llinás · logP · this kernelone QM calculation (SCF)
COSMO-RS / COSMOthermone QM calculation (SCF)

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
Llinás · logP · this kernel1
Wildman–Crippen (atom contributions)1
clogP (Leo–Hansch fragments)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
Llinás · logPthis kernelclosed-formPearson r 0.660 (nested-CV) · MAE 1.06 log unitsone DFT SCF σ-profile · no training set · no MDMF-FQSL (this lab)
Wildman–Crippen (atom contributions)empiricalσ (RMSE) 0.68, r² 0.918 (n=9920)2D atom types (68 contributions)Wildman & Crippen, J. Chem. Inf. Comput. Sci. 1999
clogP (Leo–Hansch fragments)empirical≈0.4–0.5 log on drug-like chemistry (fragment library)2D fragment decomposition + correction factorsLeo, Chem. Rev. 1993; BioByte clogP
COSMO-RS / COSMOthermphysicsMAE 0.57, RMSD 0.78 (SAMPL7 blind logP, n=22)DFT σ-profile + parametrisationLoschen & Klamt, SAMPL7 logP, J. Comput. Aided Mol. Des. 2021
D-MPNN (Chemprop)ML / GNNconsensus RMSE ≈ 0.48 log (lipophilicity benchmark)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

sigma_variancesigma_variance — spread of the σ-profile (polarity dispersion)

Version history

  1. 2026-05-31v0.91.1 B.UNDERFIT

    r=0.660, +28% MAE-vs-NULL. 2-term collapses (winner's curse); 1-term ships.