Fig. 1: Standard autoencoder architecture (left) and an example DiffNet architecture (right).

Autoencoders have an encoder that compress the input data to a bottleneck, or latent, layer and a decoder that expands the latent representation to reconstruct the original input. The DiffNet adds a classification task to the latent space. In the example shown, the input is split into two encoders. One is a supervised encoder that operates on atoms near the mutation (cyan) and must predict the biochemical property associated with a structure. The second encoder is unsupervised and operates on the rest of the protein (blue). The latent layers from these two encoders are concatenated and trained to reconstruct the original input.