AI-guided engineering of FAD-glucose dehydrogenase for enhanced processability and functional performance

Limited enzyme processability is a major bottleneck in biocatalysis and biosensing, particularly for aggregation-prone enzymes expressed in heterologous hosts. Although AI-guided protein engineering has advanced through physics-informed stability models, experimental demonstrations connecting stability-guided design to process-relevant outcomes remain limited. Here, we apply IFUM ( In silico evaluation of unfolding Free energy with Unfolded ensemble Modeling)–guided redesign combined with evolutionary filtering and LigandMPNN sequence optimization to improve the processability of flavin adenine dinucleotide–dependent glucose dehydrogenase (FAD-GDH), an oxygen-insensitive biocatalyst relevant to next- generation bioelectronic applications. AI-designed variants showed markedly enhanced soluble expression and purification in Escherichia coli . Notably, variant M3 yielded 7.68 mg purified protein from 0.2 L culture (38 mg L⁻¹), whereas the wild-type enzyme yielded 1.9 mg from 9 L culture (0.21 mg L⁻¹), demonstrating a substantial improvement in effective protein recovery. In a DCIP/PMS coupled assay, M3 reduced DCIP faster than the wild type even at ∼60-fold lower enzyme concentration, indicating improved functional usability under reduced enzyme dosage. Supporting in silico analyses, including structure prediction and molecular dynamics simulations, were consistent with enhanced loop stabilization and strengthened interdomain coupling. Together, these results demonstrate that IFUM-guided AI engineering can translate stability focused design into experimentally robust, process-relevant enzyme variants, providing a practical framework for engineering FAD-dependent oxidoreductases for bioprocessing and bioelectronic applications.