MolBasis
A.IMPROVEfamilyBradley Merck melting-point database · n=1183

Bradley · Tm

Melting point · K

closed-form kernel
ŷ = β₀ + β₁ · z(g_polar^0.204) + β₂ · z(donor^0.520) (family: size × polarity-spread)
0.655
nested-CV r
in-sample, winner's-curse-safe
0.622
scaffold-blind r
25% held-out OOD
+38.9%
blind MAE-vs-NULL lift
gate = +30%
38
MAE (K)
grouped-LOO

In-sample → out-of-distribution

How the correlation holds when moving to unseen scaffolds.

nested-CV r (in-sample)0.655
scaffold-blind r (OOD)0.622
Blind MAE-vs-NULL lift38.9%

Vertical marker = the +30% production gate.

Blind calibration slope

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

0.91
slope (ideal 1.0)
Blind lift holds (≥30%)

What it is

Melting point from a size × polarity-spread FAMILY of exponents (not a single unique pair). The headline self-correction: the published r=0.826 was inflated by a hydrogen-bond classification bug; the honest value is r=0.655.

The physics

Lattice cohesion scales with molecular size and electrostatic complementarity. The data identifies the family (many exponent pairs within ±0.01 r) but not one specific pair — so it ships as a family predictor, not a mechanistic law.

upper / lower bound

Learning curve is flat (adding 1831 more anchors would buy ≈+0.003 r, below noise) → data-saturated, feature-limited. Ship as a predictor; drop any mechanistic story about the exact exponents.

caveat

🚨 CORRECTED 2026-06-06. The v1 published r=0.826 was partially inflated: a 1.2σ H-bond donor threshold miscounted aromatic π-clouds as false donors (~45% false-positive rate on aromatics), and the Bradley cohort is aromatic-rich, so the spurious donor signal leaked into the fit. After the v2 σ-degree atom-role fix, the honest r is 0.655. The other 7 production kernels do NOT select donor/acceptor and are unaffected.

How the methods compare

Melting point is genuinely hard — it needs the crystal lattice, which no cheap method captures. ML wins on MAE (~26–31 K), but by memorising a large training set. This kernel is honest about the ceiling: a size × polarity family at ~38 K MAE, with no training and full interpretability.

Reported error

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

ML (ASNN / DNN / GNN)28 K · MAE
Group contribution (Joback; Marrero–Gani)35 K · MAE (~30–40 K)
Bradley · Tm · this kernel38 K · MAE

Not charted (no single reported figure): Crystal structure prediction.

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.

Group contribution (Joback; Marrero–Gani)2D structure (pencil / inference)
ML (ASNN / DNN / GNN)2D structure (pencil / inference)
Bradley · Tm · this kernelone QM calculation (SCF)
Crystal structure predictionMD 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
Bradley · Tm · this kernel1
Group contribution (Joback; Marrero–Gani)0
ML (ASNN / DNN / GNN)1
Crystal structure prediction1

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
Bradley · Tmthis kernelclosed-formPearson r 0.655 (nested-CV) · 0.622 scaffold-blind · MAE 38 Kone DFT SCF σ-profile · no training set · no MDMF-FQSL (this lab)
Group contribution (Joback; Marrero–Gani)group-contributionMAE ≳ 30–40 K on diverse sets; fails for rigid/H-bonded solids2D functional-group countsJoback & Reid 1987; Marrero & Gani 2001
ML (ASNN / DNN / GNN)ML / GNNMAE ≈ 26–31 K on Bradley / Tetko sets (SOTA, near experimental noise)2D descriptors or graph + a large labelled training setTetko et al. 2014; Coley et al. (MAE 26.2 K)
Crystal structure predictionphysicsNo reliable general route — melting needs the crystal latticepredicted crystal structure + lattice-energy calculatione.g. Price; CCDC blind tests

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

g_polarg_polar_kcal — polar screening (electrostatic solvation) energy (kcal/mol)
donorhb_donor_mass — hydrogen-bond donor surface area (σ-degree classified)

Version history

  1. 2026-05-29v1 published (SUPERSEDED)

    r=0.826 — later found inflated by the v1 H-bond π-cloud false-positive bug.

  2. 2026-06-06v2 CORRECTED

    σ-degree atom-role classification. Pass A retrobench r=0.610; Pass B (free gate) r=0.655. Form shifted to g_polar+donor. This is the honest number.