QSAR

QSAR is the idea that a molecule's biological activity can be predicted from numbers describing its structure, the way a house price can be estimated from its size, location, and number of rooms. You collect a set of molecules with measured activities, turn each into descriptive numbers, and fit a model that maps those numbers to activity.

Concretely, each molecule is encoded as molecular descriptors, such as lipophilicity, polar surface area, or counts of certain features. A statistical or machine-learning model is then trained so that, given the descriptors of a new untested molecule, it predicts its likely potency or a related property. Classical QSAR used linear regression on a handful of descriptors; modern variants may use many descriptors and nonlinear learners.

QSAR is genuinely useful for prioritizing which compounds to make, but it is only as good as its training data and only valid inside the chemical space it was trained on. Pushed outside that space, into structurally unfamiliar molecules, its predictions can fail badly, a problem called the applicability domain. Honest practice means reporting how confident a prediction is and not trusting extrapolations.