Fig. 3

Performance of the PRDIS specific predictors. a and c. Frequency distribution of MCC values for all the specific predictors generated in this work: (a) data for simple neural networks; (c) data for neural networks with one hidden layer and two nodes. Shown with a dashed line is 0, the MCC value for a random predictor. We see that specific predictors are systematically better than the random predictor. b and d. Contribution of the three biochemical/biophysical properties (Blosum62 elements, Shannon’s entropy and Position specific scoring matrix elements; see Materials and Methods) to improve the performance of the specific predictors. Points above the dotted line correspond to cases where use of these properties improves the performance of a specific predictor. We see that this is essentially always the case. b and d correspond to the simpler and to the one hidden layer neural networks, respectively