Generation of mutant protein amino acid sequence with Variational Autoencoder

Abstract

In order to improve the efficiency of functional protein development, there are widespread attempts to have amino acid sequence generation models propose promising protein candidates. However, sequences similar to known sequences output by generative models are not always promising in terms of performance. In this study, we propose a method to learn not only sequence similarity but also performance similarity at the same time. This method uses a Variational Autoencoder trained to correlate one component of a latent vector with protein performance, and generates mutant protein amino acid sequences extrapolated from known ones to enhance performance. A simulated evaluation using a performance prediction model confirms the effectiveness of the proposed method in improving development efficiency in the design of candidate proteins.