Learning Secondary Structure of Proteins

Accurate prediction of protein secondary structure is an important step in understanding protein folding. Most of the knowledge of the secondary structure of proteins comes from the X-ray diiraction patterns of crystallized proteins. However, these crystallographic analyses are very expensive, laborious and time consuming. An alternative is to use the known structures to predict the secondary structure of new proteins. Therefore it is very important to nd good prediction models for this problem. Given a data base of known proteins, and the primary structure of a target protein, we want to predict its secondary structure. The problem is more diicult if the data base does not contain proteins that are homologous to the target one. 1107 We present a PAC-learning model and its implementation to study this problem. Our implementation uses the same set of proteins as presented by Qian and Sejnowski in their paper, and we pay special attention to the case when there are no homologous proteins between the training and testing set.