Artificial intelligence methods for protein folding and design

Machine learning has revolutionized protein structure prediction and design. This review discusses current methods for protein folding and inverse folding challenges. Models like AlphaFold2 (AF2), RoseTTAFold, and ESMFold excel at leveraging evolutionary information to accurately predict protein structures while still struggling to capture the physics of protein folding. Their repurposing for protein design has led to innovations such as RFdiffusion, AF2-design, and relaxed sequence optimization. ProteinMPNN and ESM-IF design sequences based on structure, so they are frequently referred to as “inverse folding’ methods. By examining the potential and limitations of current protein design methods and metrics, we provide perspectives on developing models that fully characterize energy landscapes associated with amino acid sequences. Such advances would enable more accurate structure prediction and the design of proteins with specified conformational dynamics, potentially transforming our ability to engineer novel proteins for biotechnological applications.