molecule-solubility-analysis

Chemistry · Data analysis · Machine learning

What makes a molecule dissolve in water?

Measured water solubility for 1,128 molecules from the ESOL dataset, what a machine-learning model predicts for each one, and the chemical families where it still gets things badly wrong.

Explore the molecules

Each dot is one molecule. Hover or tap for details. Orange means the model was off by more than 10×.

Horizontal axis
Highlight a chemical family

What the data says

Oiliness drives it

Correlation of each property with measured solubility. Negative means more of it, less soluble.

Where the model struggles

Prediction error (RMSE, log units) by chemical family. Click a family to highlight it above.

Unfamiliar molecules

Error grouped by how similar each molecule is to its closest neighbour in the data.

Open research questions

The errors aren't random. Each cluster points to something worth testing.

  1. Why are organophosphates under-predicted?

    43 phosphorus esters have ~2× the typical error, and they're always more soluble than predicted. Is the Crippen logP method mis-scoring phosphorus?

    Test: compare with measured logP, or retrain with a different logP method, and check whether the bias disappears.

  2. How much is crystal packing?

    Rigid, flat, high-melting solids like anthraquinone are less soluble than predicted. Structure alone can't see how tightly a crystal packs.

    Test: add melting point, the key term in the General Solubility Equation, and measure the error drop.

  3. Can the model say "I don't know"?

    Molecules with no close neighbour in the data (similarity < 0.4) get ~50% larger errors.

    Test: build an applicability-domain score and check that it ranks the worst predictions first.

  4. How far can more data go?

    The learning curve is still falling, from 0.89 to 0.70 as training molecules grow from 90 to 902.

    Test: rerun on a larger curated set such as AqSolDB (~10k compounds) and estimate the measurement-noise floor.

  5. What about drug-like acids?

    The dataset has no carboxylic acids and only 9 molecules above 500 g/mol, so it says little about ionisable, drug-sized compounds.

    Test: evaluate on an external set of acids and larger molecules with pH-controlled solubility.