Fig. 3. Deep neural network model for protein fold classification.

The Network takes as input two protein structures represented by their 3DZD vectors. The encoder layer uses the three hidden layers, each with 250, 200, 150 nodes, to encode the features in the 3DZD. The encoding vector of a length of 1452 is then input into the feature extractor layer, which is used to compare the encoded feature of the two structures using four distance metrics, the Euclidian distance, the cosine distance, the Manhattan (absolute value) distance, and dot product. The FC network takes the feature extractor output and predicts the probability that the two structure belong to the same fold.