Abstract
Protein binders, including antibodies and diverse nonantibody scaffolds such as monobodies and designed ankyrin repeat proteins, have become central to targeted therapy, molecular diagnostics, biosensing, and synthetic biology. However, conventional binder discovery remains largely dominated by screening-based paradigms (immunization or in vitro display followed by iterative optimization), which are constrained by library size, experimental cost, and multiobjective developability requirements. Recent advances in structure prediction, protein language models, and diffusion-based generative modeling have established a standardized technology stack for artificial intelligence-driven de novo binder design. This review summarizes this stack through a practical workflow. First, we outline essential data resources and the role of complex structure prediction. Next, we detail generative backbone design and structure-conditioned sequence optimization. We then evaluate computational metrics and uncertainty management. Finally, we discuss integrating high-throughput experimental feedback into closed-loop optimization. We conclude by discussing key challenges, including induced fit, negative design, and dataset bias, as well as future directions toward end-to-end complex generation.