Protein Engineering for Thermostability through Deep Evolution

Abstract

Protein engineering for increased thermostability through iterative mutagenesis and high throughput screening is labor-intensive, expensive and inefficient. Here, we developed a deep evolution (DeepEvo) strategy to engineer protein thermostability through global sequence generation and selection using deep learning models. We firstly constructed a thermostability selector based on a protein language model to extract thermostability-related features in high-dimensional latent spaces of protein sequences with high temperature tolerance. Subsequently, we constructed a variant generator based on a generative adversarial network to create protein sequences containing the desirable function with more than 50% accuracy. Finally, the generator and selector were utilized to iteratively improve the performance of DeepEvo on the model protein glyceraldehyde-3-phosphate dehydrogenase (G3PDH), whereby 8 highly thermostable variants were obtained from only 30 generated sequences, demonstrating the high efficiency of DeepEvo for the engineering of protein thermostability.