Aqueous solubility (ESOL RMSE)

Licence to confirm before publication on the hub. Source(s) esol declare no known licence (an unknown licence is gated until confirmed in a publication block).

What is measured. The measured aqueous solubility of a small molecule, log10 of mol/L at room temperature, from its SMILES: MoleculeNet ESOL (Delaney, J. Chem. Inf. Comput. Sci. 2004; Wu et al., Chem. Sci. 2018) as packaged by scikit-fingerprints, 1,128 molecules from -11.6 to 1.6 log mol/L. Lower RMSE is better.

Split. The published OGB scaffold split (ogb_splits_esol.json: 902 / 113 / 113), whole Bemis-Murcko scaffolds per part. atlas-full-v1 takes the whole OGB test part (113 molecules) and the whole train part (902); atlas-n100-v1 drew 100 test molecules from the test part by seed-hash rank (seed 20261006) and 500 train molecules from the train part. The valid part is not used. Scaffold disjointness of the two subsets is re-checked when they are built. 212 SMILES of the source end with a space, which is stripped; nothing else is rewritten.

Test size. The leaderboard version is atlas-full-v1 on every profile: the whole OGB test part (113 molecules) and the whole train part (902), because a SMILES probe or a small graph network costs seconds per thousand molecules. The earlier atlas-n100-v1 (100 test / 500 train, drawn as described above) stays frozen and its cells stay current; it is a separate ranking, never mixed with the full one.

Probe (rmse_probe, mae_probe, r2_probe, spearman_probe). Frozen embeddings of the SMILES (last layer, mean over tokens); a ridge head (standardised features, alpha chosen by leave-one-out over logspace(-2, 4, 13) inside train) is fit on the train molecules and predicts the test molecules.

Supervised (rmse_supervised, mae_supervised, r2_supervised, spearman_supervised). Frameworks with no pretrained weights are trained from scratch on the same train molecules and predict the same test molecules, in the same unit.

Uncertainty. 95 % bootstrap interval over test molecules, 1,000 resamples.

Caveats. The scaffold test part is shifted towards less soluble molecules (mean -3.8 against -2.9 in train), so predicting the train mean for every test molecule scores an RMSE of 2.31 on both frozen subsets (atlas-full-v1 and atlas-n100-v1), more than the 2.1 standard deviation of the dataset. ESOL measurements come from the literature and are noisy (Delaney's own model reaches about 0.6 to 1.0); differences below about 0.1 log units are within the interval. Language models pretrained on PubChem/ZINC have seen these structures without their solubility.

Data

• esol: hf_dataset scikit-fingerprints/MoleculeNet_ESOL at revision 8929f602b69f2a32ae07dd7a4015fa4cc64224ae, licence unknown, catalogued as chem-ml/moleculenet-esol

Subset

• seed 20261006
• test: 100 item(s)
• train: 500 item(s), disjoint by scaffold
• leaderboard version atlas-full-v1 on every profile: the whole eligible test and train pool (the sizes above define the n version)

Metrics

• rmse_probe — RMSE (ridge probe); protocol probe; lower is better; range [0.0, 5.0]; 95% bootstrap CI, 1000 resamples
• mae_probe — MAE (ridge probe); protocol probe; lower is better; range [0.0, 5.0]; 95% bootstrap CI, 1000 resamples
• r2_probe — R² (ridge probe); protocol probe; higher is better; range ≤ 1.0; 95% bootstrap CI, 1000 resamples
• spearman_probe — Spearman ρ (ridge probe); protocol probe; higher is better; range [-1.0, 1.0]; 95% bootstrap CI, 1000 resamples
• rmse_supervised — RMSE (trained from scratch); protocol supervised; lower is better; range [0.0, 5.0]; 95% bootstrap CI, 1000 resamples
• mae_supervised — MAE (trained from scratch); protocol supervised; lower is better; range [0.0, 5.0]; 95% bootstrap CI, 1000 resamples
• r2_supervised — R² (trained from scratch); protocol supervised; higher is better; range ≤ 1.0; 95% bootstrap CI, 1000 resamples
• spearman_supervised — Spearman ρ (trained from scratch); protocol supervised; higher is better; range [-1.0, 1.0]; 95% bootstrap CI, 1000 resamples

Three training molecules of the frozen subset (input: the SMILES; expected output: the solubility in log10 mol/L):

item_id	molecule	SMILES	label (log mol/L)
esol-00012	1-pentene	CCCC=C	-2.68
esol-00007	estradiol	CC12CCC3C(CCc4cc(O)ccc34)C2CCC1O	-5.03
esol-00006	a pentachlorobiphenyl	Clc1cc(Cl)c(c(Cl)c1)c2c(Cl)cccc2Cl	-7.32

A model that predicts -3.0, -4.5 and -7.0 for these has squared errors 0.10, 0.28 and 0.10. The leaderboard (atlas-full-v1) scores all 113 molecules of the scaffold test part (100 in atlas-n100-v1).