MOUSE-III: learning rules of conformational analysis from X-ray data

MOUSE-III is learning program that finds rules of conformational analysis from raw crystallographic data. The program perceives molecular features, finds conformational classes in the data and then learns rules that link features to classes. The rules that MOUSE learns are capable of correctly assigning conformations to ring systems that were not used for training with greater than 95% accuracy, when MOUSE was presented with sufficient data. The rules also show a compression of as much as 99% when compared to the raw data. This is accomplished through abstraction and generalization. The algorithm is presented along with a carefully worked example. An example of a learned rule is also presented and analyzed. Some conclusions about the scope and limitations of the learning process are presented.