Implementation of the "Composite Residues" neural network architecture used in Toward machine-guided design of proteins. The article does show a diagram of the network architecture but provides no implementation. Here I implement it in keras.
To make the parameter count match I figured that they didn't use bias nodes on the five node dense layer but did use a bias node for the output layer. This may be incorrect, but honestly I don't know why they would not use these bias nodes in the first place....