End-to-end differentiable protein design using variational AutoEncoder

Abstract

Protein is an essential polymer that performs various functions in living things. Protein design means creating proteins with novel structures or functions. Computational protein design has been tried for a long time, but there were limitations because it could only deal with limited protein space compared to massive protein space. Recently, due to the advent of a deep generative model, proteins can be designed faster than conventional methods. In this research, protein sequences and corresponding structures were mapped into latent space through the variational autoencoder (VAE) model. Then, protein sequences and dihedral angles were generated from the sampled latent variable through the decoder. Protein structures were approximated by fixed backbone bond lengths and bond angles and appended new atoms with predicted dihedral angles. End to end differentiable protein structure generative model was built. Through a case study of globin superfamily, the performance of the VAE model was examined. New globin proteins were designed through our VAE model, and its properties were analyzed. VAE model was also tried up to mainly alpha helix class level.