Machine Learning of Amino Acid Composition Models for Protein Redesign

Abstract

Proteins from thermophiles can preserve their basic structures and original functions at high temperature. However, most of the mesophilic proteins are vulnerable to such extreme condition due to their different amino acid composition. Improve the thermostability of thermophilic proteins will reduce the cost of from storage and production in industrial process. Nowadays, machine learning becomes a powerful method for data-intensive computation. This work provides a well-tuned network, called Thermalizer, which applies Recurrent Neural Network (RNN) to encode the mesophilic proteins to thermostable proteins which is predicted to be able to perform expected function in higher temperature. The toolkit also provides workflow from gene and amino acid sequence preprocessing to encoder and decoder construction, model training and evaluation, and translation window. The project is accessible on GitHub: https://github.com/BeckResearchLab/thermalizer.