Inaccuracy in decoys is predicted by their variance after structural superposition, allowing rational truncation of inaccurate regions and improving performance. (a) For target 2p5k
, the inter-decoy variance (dark trace) mirrors the mean variance between each decoy and the deposited native structure (light trace). (b) Successes and failures versus bins of Patterson correlation coefficients (PCCs) that measure the correlation between inter-decoy and native–decoy variance values. Most successes, mainly deriving from SPICKER cluster 1 (black bars) and a few from cluster 2 (dark grey), originate in clusters of decoys for which the PCCs are high. Failing targets are indicated by light grey bars and PCC values derived from their respective top SPICKER clusters: the distribution of failure is shifted leftwards, indicating that the PCCs were generally lower and local inaccuracy was thus less well predicted, impacting on the search-model processing.