Aqueous solubility from structure

A drug that will not dissolve in water is hard to absorb and hard to dose. This benchmark asks for a compound’s aqueous solubility from its structure, over the largest curated solubility set in public use.

Task

Given a compound as a SMILES string, predict its aqueous solubility in $\log_{10}$ mol/L. The labels in the benchmark files run from $-13.2$ to $2.14$. Submissions are scored by mean absolute error, so lower is better.

Background

AqSolDB (Sorkun, Khetan and Er, 2019) merges nine public solubility sources into 9,982 curated entries with 2D descriptors, resolving conflicts between them; TDC uses it whole as the ADMET group’s largest regression benchmark. Solubility is the textbook target for structure-property prediction, which makes it a useful control: a model that cannot do solubility is unlikely to do the harder endpoints, and the dataset is big enough that data-hungry methods can show their advantage.

Data and evaluation

TDC ships the group as one archive (admet_group, Harvard Dataverse file 4426004), loaded with admet_group(path='data/') from the tdc.benchmark_group module. The group’s solubility_aqsoldb folder holds 7,985 train_val and 1,997 test rows (9,982 distinct SMILES, matching TDC’s page). This is the only dataset in the group whose licence TDC states outright, as CC BY 4.0.

The split is a scaffold split: TDC holds out a fixed 20% test set (here 1,997 of 9,982 rows in the group’s files), and the remaining train_val part is split again per seed by group.get_train_valid_split(benchmark=name, split_type='default', seed=seed) into 87.5% training and 12.5% validation. The leaderboard asks for five independent runs with seeds 1, 2, 3, 4 and 5; group.evaluate_many then reports the mean and the standard deviation of the per-seed test scores, rounded to three decimals. The test set never changes, so only the training split moves with the seed.

The metric is mean absolute error, $\mathrm{MAE} = \frac{1}{n}\sum_{i=1}^{n} |y_i - \hat{y}_i|$, computed by TDC’s Evaluator(name='mae') on the test set; lower is better.

Official evaluation code: tdc/benchmark_group/base_group.py in the TDC repository (MIT licence), whose evaluate and evaluate_many methods are what the leaderboard numbers come from. The dataset page gives the licence as CC BY 4.0.

What is known

• Best on the leaderboard: MiniMol, MAE $0.741 \pm 0.013$ (Kläser, Banaszewski et al., 2024).
• Next: Chemprop-RDKit, $0.761 \pm 0.025$, a message-passing network with RDKit descriptors (Yang et al., 2019), then DeepMol (AutoML), $0.775 \pm 0.006$.
• Eighteen entries are listed, spanning MAE $0.741$ to $1.203$. Unlike the small endpoints in this group, the ordering here is fairly stable: learned molecular representations lead, and the gaps are wider than the seed-to-seed spread.

What would count as progress

• A mean MAE on TDC’s fixed test set that beats MiniMol (MAE 0.741), averaged over the official seeds 1-5 and produced by group.evaluate_many, with the five per-seed numbers reported as well.
• The training and inference code, the environment and the fitted model, so that the five runs can be repeated end to end from the train_val part alone.
• Evidence that the gain is larger than run-to-run noise: the gap between the top two entries here is 0.02 log mol/L, so a difference smaller than the reported standard deviation is not a result.
• Solubility spans more than fifteen log units, so a submission should report error on the poorly soluble tail separately; that is the part that matters for formulation.

References

1. Sorkun, M. C., Khetan, A. and Er, S., "AqSolDB, a curated reference set of aqueous solubility and 2D descriptors for a diverse set of compounds", Scientific Data 6 (2019), 143. doi:10.1038/s41597-019-0151-1
2. Kläser, K., Banaszewski, B. et al., "MiniMol: A Parameter-Efficient Foundation Model for Molecular Learning", arXiv:2404.14986 (2024). arxiv.org/abs/2404.14986
3. Yang, K. et al., "Analyzing Learned Molecular Representations for Property Prediction", Journal of Chemical Information and Modeling 59 (2019), 3370–3388. doi:10.1021/acs.jcim.9b00237
4. Huang, K. et al., "Artificial intelligence foundation for therapeutic science", Nature Chemical Biology 18 (2022), 1033–1036. doi:10.1038/s41589-022-01131-2

Source

Imported from TDC ADMET benchmark group, Solubility_AqSolDB, curated by Therapeutics Data Commons (Harvard). Open according to the source as of 2026-10-20. The statement above is written in our own words; see the source for the original wording, history and discussion.