The datalake

Every dataset, one provenance-stamped atlas

All kernels are fit against a single atlas of σ-profiles, generated with one recipe applied identically to every molecule. Here is the complete census — every experimental dataset across every property category, with honest coverage.

6,876

production σ-profiles

v0.91.1 SVP atlas

experimental datasets

across 11 property categories

16d3ca90

recipe SHA

pinned in every provenance

3,140

SMILES still to compute

Tier A: 942 (~$9.42)

Atlas composition

6,876

production v0.91.1 SVP

reference solvents

v0.92.0 TZVPP smoke

hetero v3 (legacy)

6,903

unique molecules

Recipe: wB97M-V / def2-SVP / C-PCM (ε=80, Bondi cavity, Klamt-purge) · vendor SHA 56c80cea · regenerates for ~$10–21 compute, <2 days from raw data.

Solvent atlas: 45/49 reference solvents computed (water, the 6 MNSol solvents, alcohols, alkanes, chlorinated, polar aprotic…). Missing: tetrahydrofuran, formic acid, hexafluoroisopropanol, 2-octanol. These feed the bilinear solvent-partner kernels.

Hydration & solvation

8 datasets

Free energies of transfer into water and organic solvents — the home turf of σ-profile thermodynamics.

MNSol · water→

Marenich/Cramer/Truhlar — Minnesota Solvation Database

ΔG_hydration · kcal/mol

MNSol (solvent = octanol)

ΔG_solvation · kcal/mol

MNSol (solvent = hexadecane)

ΔG_solvation · kcal/mol

108 mol

93%

shipped

MNSol · chloroform

MNSol (solvent = chloroform)

ΔG_solvation · kcal/mol

60 mol

98%

98% covered; kernel not yet fitactive

MNSol · cyclohexane

MNSol (solvent = cyclohexane)

ΔG_solvation · kcal/mol

57 mol

98%

98% covered; kernel not yet fitactive

MNSol · carbon-tet

MNSol (solvent = carbon tetrachloride)

ΔG_solvation · kcal/mol

47 mol

98%

98% covered; kernel not yet fitactive

CompSol 2024→

Moine/Trinh/Ramos 2024

ΔG_hydration · kcal/mol

35,624 raw

29%

Largest solvation gap pool; kernel fit on n=1394.shipped

FreeSolv

Mobley FreeSolv (raw .txt + .xlsx variant)

ΔG_hydration · kcal/mol

—

Independent hydration baseline; needs an xlsx column fix before audit.potential

Aqueous solubility (logS)

5 datasets

Log aqueous solubility — cavity desolvation cost vs H-bond hydration reward.

ESOL (Delaney 2004)→

Delaney 2004, J. Chem. Inf. Model.

Aqueous cosolvent set

logS / logP · log

323 raw

100%

100% atlas coverage.100% covered

FDA drugs · LogW

FDA Orange Book + literature

logW (solubility) · log

71 raw

10%

Small (n=71), only 7 in atlas; exploratory.potential

Wiki_pS0

Wikipedia pS0 solubility compilation

logS0 + melting point · log / K

—

Pooled with ESOL + AqCoSol + FDA in the logS and Tm benchmarking runs; not a standalone atlas cohort.pooled

Lipophilicity (logP)

1 dataset

Octanol/water partition — driven by σ-profile polarity spread.

Sangster logP→

Sangster logP compilation

logP · log

23,512 raw

100%

Largest logP cohort tested; ships B.UNDERFIT (physics ceiling, slope 0.999).shipped

Melting point (Tm)

2 datasets

Lattice cohesion as a size × polarity-spread family.

Bradley DPG→

Bradley Double-Plus-Good MP

Tm (melting point) · K

3,041 raw

39%

Gold standard; kernel corrected to r=0.655 (2026-06-06).shipped

FDA drugs · MP

FDA Orange Book + literature

Tm (melting point) · K

72 raw

10%

Subsumed by the much larger Bradley cohort; Tm v1 (FDA+cosolvent) fit was only r=0.32.deferred

Hansen parameters & miscibility

2 datasets

Hansen solubility parameters (δ_d/δ_p/δ_h) and drug–polymer miscibility χ.

HSP dataset→

Hansen Solubility Parameters

δ_d / δ_p / δ_h · MPa^½

1,206 raw

60%

One source → 3 kernels: δ_h (A.IMPROVE), δ_p & δ_d (B.UNDERFIT).shipped

ASD miscibility χ

MolForge ASD — 9 drugs × 10 polymers

Flory χ interaction · dimensionless

90 raw

—

Blocked on hetero (Boys–Bernardi) pipeline pod validation. Bilinear v2 variant staged.blocked

Surface tension (γ)

1 dataset

Interfacial cooperative regime — the ¼-integer law predicts a 3/2 exponent here.

CRC / Lange 25 °C

CRC + Lange's Handbook

γ surface tension · mN/m

34 raw

94%

94% covered but flagged POTENTIAL — raw data needs a SMILES column added (French headers).potential

PC-SAFT parameters

1 dataset

Equation-of-state segment number m — a bridge to process thermodynamics.

Gross–Sadowski 2001

Gross & Sadowski 2001

m (segment number) · dimensionless

24 raw

96%

96% covered (THF missing); kernel exploratory.active

Polymer glass transition (Tg)

1 dataset

Amorphous-polymer Tg — hits a chain-length ceiling with a pentamer atlas.

Polymer Tg

Polymer Tg compilation

Tg · K

1,440 raw

FAIL — pentamer ceiling (r_cv≈0.29). Needs nonamer/decamer chains, not more SVP atlas.fail

Protein–ligand binding (ΔG)

1 dataset

Binding free energy from pocket-fragment σ-profiles — awaiting a schema regen.

PDBBind v2020 (refined)

PDBBind v2020 refined set

ΔG_bind · kcal/mol

1,310 raw

70%

Blocked on pocket v28→v29 regen (305 fragments have only 5/9 features) + hetero pipeline validation. Legacy baseline: MAE 1.72, r 0.56 (n=119).blocked

Dimer BSSE (reference)

2 datasets

Counterpoise-corrected dimer interaction energies — v0.92.0 TZVPP smoke sets.

Homodimer BSSE

MolForge v0.92.0 TZVPP smoke

interaction energy · kcal/mol

31 raw

—

Reference/validation only; not a kernel input yet.reference

Heterodimer BSSE (CL3)

MolForge v0.92.0 TZVPP smoke

interaction energy · kcal/mol

31 raw

—

Reference/validation only.reference

Binding-pocket fragments

1 dataset

Fragmented protein pockets at pH 7.4 — input source for the binding kernel.

v3 pocket manifests (pH 7.4)

MolForge pocket fragmenter

pocket σ-descriptors · 9-d vector

1,778 raw

—

Awaiting v29 regen — the old v28 fragments carried only 5/9 features and were deleted.blocked

The compute gap

3,140 unique SMILES would fully cover every kernel dataset. We prioritise by leverage, not volume — and we explicitly reject fills the learning curve says won’t move the needle.

Tier A — leverage fill

EXECUTE

942 compounds · ~$9.42

Compounds appearing in 2+ datasets — each unlocks 2–6 kernels at once.

Tier B — Bradley_Tm full

REJECTED

1,831 compounds · ~$20

Learning curve is flat (+0.0007 r per +500 anchors) → data-saturated.

Tier C — HSP full

DOWNGRADED

474 compounds · ~$5

Polarizability feature test failed the +30% gate; HSP near its physics floor.

Add your molecules to the lake

Upload a set of mfsig σ-profiles with measured property values and we can fit and benchmark a kernel against it under the same rigor gate.

Upload your mfsig dataset