Generating Physically-Realistic Tertiary Protein Structures with Deep Latent Variable Models Learning Over Experimentally-available Structures

Sophisticated deep neural networks have significantly advanced our ability to predict a native structure of a protein amino-acid sequence. However, going beyond a single-structure view remains challenging. While rapid advances are being made, fundamental questions on the ability of generative deep modeling to learn to generate physically-realistic tertiary structures remain. This paper makes two key contributions. It first extends deep convolutional variable autoencoder networks to be able to learn from experimentally-available tertiary structures of proteins of variable lengths. The presented models learn over distance matrix representations of tertiary structures. A systematic and detailed analysis demonstrates that the design of the training data is of primary importance to the ability of the proposed models to learn key characteristics of tertiary structures. The second contribution this paper makes is a careful analysis along several metrics that measure the physical realism of generated tertiary structures. The presented results are promising and show that once seeded with sufficient, physically-realistic structures, variational autoencoders are efficient models for generating physically-realistic tertiary structures.