HalluDesign leverages the hallucination capabilities of AlphaFold3-style structure prediction models to enable fine-tune-free, forward-pass-only iterative sequence-structure co-optimization that successfully optimizes existing protein designs, rescues failed binders, and generates diverse de novo biomolecular structures—including protein binders for small molecules, metal ions, proteins, RNA, and post-translationally modified peptides—with experimentally validated high success rates, nanomolar to picomolar binding affinities, and atomic-level structural accuracy across all tested applications.

Deep learning has revolutionized biomolecular modeling, enabling the prediction of diverse structures with atomic accuracy. However, leveraging the atomic-level precision of the structure prediction model for de novo design remains challenging. Here, we present HalluDesign, a general all-atom framework for protein optimization and de novo design, which iteratively update protein structure and sequence. HalluDesign harnesses the inherent hallucination capabilities of AlphaFold3-style structure prediction models and enables fine-tune free, forward-pass only sequence-structure co-optimization. Structure conditioning at different noise level in the structure prediction stage allows precise control over the sampling space, facilitating tasks from local and global protein optimization to de novo design. We demonstrate the versatility of this framework by optimizing suboptimal structures, rescuing previously unsuccessful designs, designing new biomolecular interactions and generating new protein structures from scratch. Experimental characterization of a diverse set of proteins—including protein binders for small molecules, a metal ion and proteins; antibody design of phosphorylation-specific peptide; and monomeric proteins—revealed high design success rates and excellent structural accuracy. Together, our comprehensive computational and experimental results highlight the broad utility of this framework. We anticipate that HalluDesign will further unlock the modeling and design potential of AlphaFold3-like models, enabling the robust creation of complex biomolecules for a wide range of biotechnological applications.