A protein structure predictor based on an energy model with learned parameters

Abstract

A new protein folding model for obtaining low-level structure information from sequence is constructed. Its form is related both to a parameterized energy function to represent the folding problem and a feed-back “neural network”. The values of unknown physical quantities appear as free parameters in this potential function. Ideas from the study of neural network models are used to develop a learning algorithm that finds values for the free parameters by using the database of known protein structures. This algorithm can be implemented in parallel on a multicomputer. The ideas are illustrated on a simple model of α-helix formation and prediction and used to investigate the role of hydrophobic forces in stabilizing helix hydrogen bonds.