Documentation / Property prediction

Property prediction

When Cheméo has no measured value for a property, it can estimate one from the molecular structure. The prediction tool takes a structure, returns estimated properties, and, importantly, tells you how far to trust each estimate.

Entering a structure

Open the prediction tool and give it a structure as a SMILES string, or draw it in the editor. For example:

CCO            ethanol
c1ccccc1       benzene
CC(=O)Oc1ccccc1C(=O)O   aspirin

Cheméo canonicalises the structure and returns its estimated properties.

Reading an estimate

Each estimated property comes with three things, and reading them together is the point of the tool:

• The value, in SI units.
• The uncertainty, an estimate of how far the value may be off. A property with a large uncertainty is a rough guide, not a number to design around.
• The method, naming the model that produced it. Some properties come from the Cheméo Relay family of graph machine-learning models (a versioned source such as chemeo-relay-1.0); others use established estimation methods.

The out-of-domain warning

A model is only reliable near the kinds of molecules it was trained on. When a structure sits far outside that experience, Cheméo says so rather than returning a confident-looking number that is really a guess. An estimate marked out of domain should be treated as a rough indication at best.

This is deliberate: a prediction tool that always answers confidently is worse than one that admits the edge of its knowledge, because the confident wrong answer is the one that costs you.

Measured values always win

Where Cheméo holds a measured value for a property, it shows that, not an estimate. Prediction fills the gaps; it never overrides data. The compound pages label which numbers are measured and which are predicted, so you always know which you are looking at.

Programmatic access

The same predictions are available over the REST API (/api/v1/predict/{smiles}) and to AI assistants over the MCP server (predict_properties), both with the same value, uncertainty and method on every estimate.