The prediction and modeling of protein structure is a central problem in bioinformatics. Neural networks have been used extensively to predict the secondary structure of proteins. While significant progress has been made by using multiple sequence data, the ability to predict secondary structure from a single sequence and a single prediction network has stagnated with an accuracy of about 75%. This implies that there is some limit to the accuracy of the prediction. In order to understand this behavior we asked the question of what happens as we change the target function for the prediction. Instead of predicting a derived quantity, such as whether a given chain is a helix, sheet or turn, we tested whether a more directly observed quantity such as the distance between a pair of α-carbon atoms could be predicted with reasonable accuracy. The α-carbon atom position is central to each residue in the protein and the distances between them in sequence define the backbone of protein. Knowledge of the distances between the α-carbon atoms is sufficient to determine the three dimensional structure of the protein. We have trained on distance data derived from the complete protein structure database (pdb) using a multi-layered perceptron feedforward neural network with back propagation. It shows that the root of mean square error is 0.4 Å with orthogonal coding of protein primary sequence. This is comparable to the experimental error in the structures used to form the database. The effects of exploring other encoding schemes, and different complexities of neural networks as well as related target functions such as distance thresholds will be presented.
|