Applying supervised and unsupervised learning methods to discover regularities in protein sequences and structures

The gap between the number of known protein sequences and the number of known protein structures is rapidly increasing. Computer-based protein structure prediction aims at reducing this sequence-structure gap. Establishing connections between sequence similarity and structural similarity is a prerequisite in the development of protein structure prediction methods. The automated discovery of these connections is an excellent field for applying machine learning techniques. The paper presents an approach to identify long range relationships between protein sequence patterns and structural motifs by varying the granulation of the structure description. The accuracy of predicting hexameric constituency in structural classes from sequence information provided the basis for evaluating and optimizing various classification and pattern recognition schemes. About 10 structural states seem to be the optimum to take into account subtleties of the protein structure on the one hand and to allow the discrimination of its sequential equivalents on the other hand.