Despite the essential roles of proteins in biological systems, optimizing them to meet multiple functional requirements, such as thermal stability, binding affinity, and expression yield, remains challenging due to structural complexity and the resource-intensive nature of traditional methods. To address this, we propose an iterative machine learning (ML)-guided approach for protein engineering that efficiently explores the protein sequence space while reducing reliance on costly characterization. Our method uses ML models to predict protein properties and guide the search for optimal sequences. To improve model accuracy, we adopt an iterative process in which a subset of predicted sequences is experimentally validated, and the resulting data are used to finetune the models. We validated this approach using glutamine binding protein (QBP) as a model system, targeting improvements in structural stability, ligand binding energy, and shape complementarity. A genetic algorithm, directed by the ML models, effectively identified mutant sequences with superior performance compared to those from conventional approaches. With each iteration, the ML models improved in predictive power, enabling the discovery of novel QBP variants with enhanced properties. This study demonstrates the potential of integrating ML and iterative optimization for efficient and scalable protein engineering.
Ko et al. (Sat,) studied this question.