Generative Protein Design: From Deep Learning Algorithms to Translational Applications

Deep learning has transformed protein design from a field long dominated by explicit energy-function optimization into one dominated by probabilistic generative modeling. In this review, we summarize the protein representation algorithmic basis for this transition, from sequence-centered encodings to geometric graph representations and, more recently, SE(3)-equivariant structural manifolds that directly respect three-dimensional symmetry. We classify current approaches into three methodological paradigms according to how sequence and structure are related during design: sequence-structure decoupled design, hybrid approaches, and sequence-structure co-design. For decoupled workflows, we discuss hallucination, backbone generation, and backbone-conditioned sequence design. For hybrid approaches, we examine integrated two-stage architectures and predictor-driven iterative co-refinement. For co-design, we review explicit joint generative formulations in which sequence and structure are treated as a coupled design state throughout generation. Additionally, we summarize evaluation principles for assessing the design results, such as physical validity, folding consistency, and design coverage, and then introduce some important applications in several fields. Taken together, these developments indicate that generative protein design is making progress from structure generation toward the programmable engineering of complex biological function.