Bioactivity-oriented de novo design of small molecules by conditional variational autoencoders

Abstract

[en] Deep generative networks are an emerging technology in drug discovery. Our work is divided in two parts. In the first one, we built a variational autoencoder (VAE) that is able to learn the grammar of the molecules, represent them in a latent space, and generate new ones. In the second one, we built and trained a conditional variational autoencoder (CVAE) that is capable of generating new molecules based on desired properties. We will see in detail the architecture of both models and how they were trained. The molecule properties were provided by the Chemical Checker (CC), a resource of processed, harmonised and integrated small-molecule bioactivity data. We will generate different molecules with different target properties, and we will check how close the properties of the generated molecules are from the target ones. These properties are called signatures. At the end of the project we sample CC signatures with different similarity to the input molecule signatures, and we show that the signatures of the molecules generated this way resemble the sampled signatures, meaning that we can generate new random molecules based on desired properties.