Fig. 3

Comparison of three neural networks with identical architecture on EPC-map_test (long-range contacts). The baseline network (square marker) uses the full feature set and is trained on 657 proteins. The training proteins are a mix of EPC-map_train and MetaPSICOV proteins. The square marker denotes the neural network that is trained without the amino acid composition but on the same data set. The second network (circle marker) shows the performance of the neural network after increasing the training set size from 657 to 1479 proteins, which was possible because dropping the amino acid composition reduced the dimensionality of the learning problem. Note here that most of the new proteins are much more complex and may not be helpful for predicting proteins in EPC-map_test