Learning Conformation Rules

Protein conformation problem, one of the hard and important problems, is to identify conformation rules which transform sequences to their tertiary structures, called conformations. Our aim of this work is to give a concrete theoretical foundation for graph-theoretic approach for the protein conformation problem in the framework of a probabilistic learning model. We propose the conformation problem as a learning problem from hypergraphs capturing the conformations of proteins in a loose way. We consider several classes of functions based on conformation rules, and show the PAC-learnability of them. The refutable PAC-learnability of functions is discussed, which would be helpful when a target function is not in the class of functions under consideration. We also report the conformation rules learned in our preliminary computational experiments.