Every kernel, with its honest number
Each kernel maps mfsig σ-profile descriptors to a physical property through a closed-form ridge equation. Cards show the in-sample nested-CV r and the 25% scaffold-blind r — the commercial out-of-distribution estimate.
Production kernels
A.IMPROVE · 8MNSol · Water
Hydration free energy ΔG_hyd
MNSol · Hexadecane
Non-polar solvation ΔG_solv
Aqueous · logP
Partition coefficient logP (cosolvent)
HSP · δ_h
Hansen H-bonding parameter
MNSol · Octanol
Mixed solvation ΔG_solv
ESOL · logS
Aqueous solubility logS
CompSol · ΔG_hyd
Hydration free energy (large)
Bradley · Tm
Melting point
Honest under-fits
B.UNDERFIT · 4These clear r ≥ 0.5 but fall below the +30% MAE-vs-NULL blind-lift gate. We ship them labelled as under-fits rather than dressing them up — the physics ceiling is real, and the honest verdict is the useful one.
Sangster · logP
Octanol/water partition (large)
HSP · δ_p
Hansen polar parameter
HSP · δ_d
Hansen dispersion parameter
Llinás · logP
Partition coefficient (small, high-quality)
Pipeline & roadmap
7 targetsProperties we’ve probed that aren’t shipped — shown with their honest status so the scope is complete: what failed and why, what’s blocked on infrastructure, and what’s a data-prep away from a fit.
Polymer Tg
Glass transition temperature
The kernel fails the 5-fold CV gate. The signal is physical (donor^-0.25) but a pentamer atlas plus a 9-descriptor vector hits a chain-length ceiling.
Next: Nonamer/decamer chains + lattice-aware features — not more SVP atlas growth.
ASD miscibility χ
Drug–polymer Flory χ
Needs the hetero (Boys–Bernardi counterpoise) pipeline, whose orchestrator shipped but was never validated on a GPU pod.
Next: 1–2 smoke pairs on a pod (~$2) to confirm the code path before any cohort spawn.
PDBBind binding ΔG
Protein–ligand binding free energy
Blocked on a pocket schema regen — 305 v28 pocket fragments carry only 5/9 features (missing HOMO/LUMO/gap/g_polar) — plus hetero pipeline validation.
Next: Pocket v29 regen (~5×H100, ~3h, ~$50), then an apples-to-apples re-fit.
Surface tension γ
Liquid surface tension @ 25 °C
Data is 94% covered and the ¼-integer law predicts a clean 3/2 exponent here — but the raw file needs a SMILES column added (French headers) before audit.
Next: Data prep, then run through the standard rigor gate.
FreeSolv ΔG_hyd
Hydration free energy (independent baseline)
A valuable independent test set for the hydration kernels; needs an xlsx column fix before it can be wired in.
Next: Fix the input, use as an out-of-distribution check on MNSol_water / CompSol.
FDA drugs LogW
Drug aqueous solubility
Small, drug-relevant cohort — underpowered on its own but useful pooled with ESOL/AqCoSol.
Next: Compute the 64 missing (~$1); treat as exploratory.
Tm v1 (FDA + cosolvent)
Melting point (small cohort)
An honest but weak result on a small cohort — the Bradley cohort (n=1183) gives a far stronger Tm predictor.
Next: Deferred in favour of Bradley_Tm as the primary melting-point anchor set.
Reading the cards
- nested-CV r — in-sample Pearson r under per-fold pair selection (winner’s-curse-safe).
- blind r — Pearson r on a 25% Bemis–Murcko scaffold split held out up front, re-fit on the 75%. The OOD estimate.
- A.IMPROVE — clears the +30% MAE-vs-NULL gate. B.UNDERFIT — clears r ≥ 0.5 but not the +30% gate.
- 1-term / 2-term / family — number of descriptor terms; family means many exponent pairs tie within ±0.01 r (no single ‘magic’ pair).
12 kernels total. All numbers from internal audit docs (May–June 2026); corrected values shown where an earlier figure was superseded.