Jennifer Listgarten, Hanlun Jiang
AI-based generative models and representation learning are revolutionizing the design-evaluation cycle of protein engineering, and this paper provides a review that integrates them through statistical interpretation.
Protein engineering requires searching vast sequence space to design proteins with desired functions (therapeutics, diagnostics, agriculture, etc.). Traditional computational models and high-throughput experiments have limitations, and AI offers efficient search methods.
The paper categorizes AI-based protein engineering into four main streams: (1) generative modeling of sequences, backbone structures, and atoms; (2) tailored generative models for specific properties; (3) protein representation learning and scoring candidate sequences; (4) synthesis-aware library design. These are unified from a statistical perspective.
AI methods have significantly improved the accuracy and efficiency of protein design, particularly generative models and representation learning reducing experimental costs and expanding design space. This review provides a framework integrating diverse approaches, guiding future research directions.