Generating Novel and Soluble Class II Fructose-1,6-Bisphosphate Aldolase with ProteinGAN

Abstract

Fructose-1,6-bisphosphate aldolase (FBA) is an important enzyme involved in central carbon metabolism (CCM) with promising industrial applications. Artificial intelligence models like generative adversarial networks (GANs) can design novel sequences that differ from natural ones. To expand the sequence space of FBA, we applied the generative adversarial network (ProteinGAN) model for the de novo design of FBA in this study. First, we corroborated the viability of the ProteinGAN model through replicating the generation of functional MDH variants. The model was then applied to the design of class II FBA. Computational analysis showed that the model successfully captured features of natural class II FBA sequences while expanding sequence diversity. Experimental results validated soluble expression and activity for the generated FBAs. Among the 20 generated FBA sequences (identity ranging from 85% to 99% with the closest natural FBA sequences), 4 were successfully expressed as soluble proteins in E. coli, and 2 of these 4 were functional. We further proposed a filter based on sequence identity to the endogenous FBA of E. coli and reselected 10 sequences (sequence identity ranging from 85% to 95%). Among them, six were successfully expressed as soluble proteins, and five of these six were functional—a significant improvement compared to the previous results. Furthermore, one generated FBA exhibited activity that was 1.69fold the control FBA. This study demonstrates that enzyme design with GANs can generate functional protein variants with enhanced performance and unique sequences.