Abstract
Using ordered and disordered regions identified either by X-ray crystallography or by NMR spectroscopy, we trained neural networks to predict order and disorder from amino acid sequence. Although the NMR-based predictor initially appeared to be much better than the one based on the X-ray data, both predictors yielded similar overall accuracies when tested on each other's training sets, and indicated similar regions of disorder upon each sequence. The predictors trained with X-ray data showed similar results for a 5-cross validation experiment and for the out-of-sample predictions on the NMR characterized data. In contrast, the predictor trained with NMR data gave substantially worse accuracies on the out-of-sample X-ray data as compared to the accuracies displayed by the 5-cross validation during the network training. Overall, the results from the two predictors suggest that disordered regions comprise a sequence-dependant category distinct from that of ordered protein structure.
