HalluDesign

HalluDesign is a general, all-atom framework for protein optimization and de novo design that iteratively updates a protein's structure and sequence by harnessing the "hallucination" capability of AlphaFold3-style structure prediction models. It was developed at Westlake University's State Key Laboratory of Gene Expression (School of Life Sciences), with Minchao Fang as first author and Longxing Cao as corresponding author, and posted to bioRxiv in November 2025.

The central idea is that modern all-atom structure predictors such as AlphaFold3 — and its open reproduction Protenix — implicitly encode a rich understanding of sequence-structure relationships, including for small molecules, metal ions, and modified residues. HalluDesign exploits this by running the predictor's forward pass and using its confidence and structural output to drive iterative co-optimization of sequence and backbone, without any task-specific re-training or fine-tuning. This places it alongside other inference-time design frameworks (most directly ByteDance's PXDesign) that turn a pretrained structure predictor into a design engine, rather than training a dedicated generative model.

What distinguishes HalluDesign is its breadth: because it operates in the native all-atom representation of AlphaFold3, a single framework spans protein-protein, protein-small molecule, and protein-metal binder design, phosphorylation-specific antibody design, and monomeric protein generation. The authors report experimental validation across each of these categories.

#Key Features

• Fine-tune-free, forward-pass-only design: HalluDesign requires no re-training or weight updates; it drives design entirely through the forward pass of a frozen structure predictor, making it immediately compatible with new AlphaFold3-class checkpoints.
• All-atom generality: Operating in an all-atom representation lets one framework design binders against small molecules, metal ions, and proteins, rather than being limited to protein-only interfaces.
• Antibody and PTM-specific design: The method was applied to design antibodies that recognize phosphorylation-specific peptides, a notoriously difficult post-translational-modification target class.
• Pluggable structure backends: Designs can be driven by AlphaFold3, by the openly licensed Protenix, or in a cross-model configuration; sequence steps use LigandMPNN or ProteinMPNN.
• Companion nucleic-acid interface: A separate repository (HalluDesign_NA) extends the same iterative loop to DNA/RNA sequence optimization using NA-MPNN with AF3/Protenix self-consistency checks.

#Technical Details

HalluDesign wraps a pretrained all-atom structure predictor in an iterative optimization loop: it proposes a structure, regenerates sequence with an MPNN-family model (LigandMPNN or ProteinMPNN), re-evaluates with the predictor, and recycles the result into the next iteration, using diffusion-based optimization with configurable step counts (the repository uses roughly 50 steps for local optimization and 150 for global de novo design). Crucially, there are no HalluDesign-specific trained weights — the framework is pure inference over external checkpoints. AlphaFold3 weights must be obtained separately under DeepMind's terms and placed manually, while Protenix checkpoints download automatically on first run. The code is released under the MIT license (~80 GitHub stars) and is implemented predominantly in Python. The preprint reports experimental validation across binder design (small molecules, metal ions, and proteins), phosphorylation-specific antibody design, and monomeric protein design, demonstrating that a single hallucination-driven loop generalizes across these tasks.

#Applications

HalluDesign targets protein engineers and structural biologists who need to design binders or de novo proteins against structurally defined targets, including challenging non-protein targets such as small molecules and metal ions that template-based or protein-only methods handle poorly. Because it is fine-tune-free, groups can apply it to new targets — or swap in newer structure-prediction checkpoints — without assembling training data or retraining a generator. Concrete use cases highlighted by the authors include phosphorylation-specific antibody design for signaling research and diagnostics, metal- and ligand-binding protein design, and generation of novel monomeric folds, with the companion HalluDesign_NA extending the approach to nucleic-acid optimization.

#Impact

By showing that the hallucination behavior of AlphaFold3-style predictors can be redirected into a single, retraining-free framework spanning small-molecule, metal, protein, and PTM-specific design, HalluDesign broadens the reach of inference-time design methods beyond the protein-protein binder setting that dominates tools like RFdiffusion and PXDesign. Its forward-pass-only formulation lowers the barrier to adopting future structure predictors and makes all-atom design accessible to labs without large training pipelines. Important caveats apply: as a bioRxiv preprint (distributed CC BY-NC) the results await peer review and independent reproduction; performance depends on the quality of the underlying predictor's confidence signal; and use of AlphaFold3 weights is gated by DeepMind's separate, non-commercial access terms, with Protenix serving as the openly licensed alternative.