Applicability domains are intended to delimit the region in which a quantitative structure–activity relationship model has defensible predictive support. In practice, however, domain assessment is often compressed into an administrative decision: a query compound is classified as inside or outside a boundary, after which the binary label is allowed to stand in for prediction reliability. This Perspective argues that this compression obscures several analytically distinct quantities. Chemical-space coverage, local training support, response-surface regularity, model uncertainty, distribution shift, calibration quality, and decision consequence need not change together. A compound may therefore occupy apparently familiar chemical space while remaining difficult to predict, whereas a geometrically unusual compound may sometimes receive a defensible prediction from a well-calibrated model. The article develops an evidence-bounded distinction between domain membership and prediction confidence, examines why static boundaries can fail under realistic molecular distribution shifts, and prepares the basis for a proposed risk-calibrated applicability framework. The central claim is not that applicability domains should be abandoned, but that they should be interpreted as one component of a broader reliability assessment whose evidential meaning is explicitly validated and reported.
IJPRAS
