This is not a molecule! Learn how to improve molecular autoencoders by using heteroencoders

In a new blog post I explain a bit about to improve properties of molecular autoencoders by trans coding between different formats or versions of representations of the same molecule with a chemical heteroencoder. Instead of encoding the representation, the artificial neural networks are instead forced to encode the underlying common information between the representations. This gives improvements for the smoothness of the latent space and makes it more relevant for use in QSAR models. Additionally, the change in training regime leads to more creative de-novo molecular generation.

As it's really easy to use SMILES enumeration techniques during training, the take home message is: Use chemical heteroencoders, not autoencoders.

Read the summary on the blog post, or go directly to the details in the preprint: Improving Chemical Autoencoder Latent Space and Molecular De novo Generation Diversity with Heteroencoders