Virtual screening and molecular-property models increasingly produce ordered candidate lists, yet the numerical precision of a prediction is often treated as if it were equivalent to decision certainty. That interpretation is unsafe because prediction error can arise from noisy reference data, model uncertainty, weak chemical-space support, local structure–activity discontinuities, model disagreement, and miscalibration. This Decision-Science Perspective argues that uncertainty should be represented before compounds are ranked rather than appended as a confidence note after ranking. The central distinction is between a predicted value and the evidential conditions under which that value deserves decision weight. The article develops a proposed risk-adjusted prioritization logic in which chemical-space support, model disagreement, calibration, and decision utility remain analytically visible until the point of candidate selection. No universal uncertainty metric, penalty function, or acceptance threshold is assumed. Instead, uncertainty is treated as a structured decision variable that can justify promotion, down-weighting, deferral, or targeted experimentation. The framework is intended for prospective evaluation against score-only ranking under future-like chemical and data shifts.
IJPRAS
