Pattern-recognition-based detection of planar objects in three-dimensional electron-density maps.

A pattern-recognition-based method for the detection of planar objects in protein or DNA/RNA crystal structure determination is described. The procedure derives a set of rotation-invariant numeric features from local regions in the asymmetric unit of a crystallographic electron-density map. These features, primarily moments of various orders, capture different aspects of the local shape of objects in the electron density. Feature classification is achieved using a linear discriminant that is trained to optimize the contrast between planar and nonplanar objects. In five selected test cases with X-ray data spanning 2.0-3.0 A resolution, the procedure identified the correct location and orientation for almost all of the double-ring and a majority of the single-ring planar groups. The accuracy of the location of the plane centres is of the order of 0.5 A, even in moderately noisy density maps.