HalluDesign-NA

Functional nucleic acids — an aptamer that binds a chosen protein, a structured RNA that holds a defined fold — are still overwhelmingly found rather than designed. SELEX and related selections sift enormous random libraries for sequences that happen to work, and the winner's three-dimensional structure is usually never determined. Protein design left that regime years ago. Nucleic acid design has not, largely because the structure predictors that make protein hallucination work were, until recently, unreliable on DNA and RNA.

HalluDesign-NA closes that gap by substitution rather than by training something new. Its predecessor, HalluDesign, exploits the hallucination effect of the generative diffusion module in AlphaFold3-style models: freeze the predictor, hand it a sequence, and iteratively rewrite that sequence until the predictor folds it confidently into the intended shape, with a sequence-design network proposing each rewrite. HalluDesign-NA changes exactly one component of that loop — ProteinMPNN and LigandMPNN are replaced by NA-MPNN, a message-passing model from the Baker lab trained to design DNA and RNA sequences — and leaves the rest of the framework unchanged. The structure oracle can be AlphaFold 3 itself or Protenix, its open reproduction.

The work is from Minchao Fang, Zhe Wang and Longxing Cao at Westlake University, posted to bioRxiv in June 2026. Its evidence is entirely computational: the benchmarks report model-confidence scores rising over design cycles, and no designed nucleic acid has been experimentally characterized. The authors state that experimental characterization of the designed aptamers is in progress.

#Key Features

• NA-MPNN in the design seat: The nucleic acid chains are the ones rewritten each cycle, while protein chains are held fixed by default — the arrangement needed for aptamer design against a given target.
• Two interchangeable structure oracles: Separate AlphaFold 3 and Protenix backends implement the same design cycle, so the loop runs on either an in-house AlphaFold 3 installation or a fully open predictor.
• Optional joint protein–nucleic acid design: A --joint_design mode has NA-MPNN update the DNA/RNA chains first, then redesigns selected protein chains with ProteinMPNN or LigandMPNN using the nucleic-acid-updated structure as context.
• Symmetry and residue-level control: Identity tying across chains or across repeated segments within one chain produces symmetric assemblies, and explicit fix/redesign residue lists restrict sampling to chosen positions.
• Random-sequence initialization: No starting structure is required. A chain specification such as A:100-120,B:20-40 sets initial random sequence lengths, and the first cycle runs as a pure structure prediction whose output seeds the loop.

#Technical Details

Each cycle copies the current best structure and runs NA-MPNN on its nucleic acid chains. From a configurable cycle onward, several NA-MPNN candidates are generated and ranked by sequence-only prediction with AlphaFold 3 or Protenix; the top candidate's packed coordinates are then fed back as reference-guided input to an optimization prediction that continues denoising for a set number of steps, and the result is recycled into the next cycle. Nothing is trained and no weights are updated: the repository holds design code and thin wrappers over two external models, with no training script and no checkpoints of its own. The repository's example commands run ten design cycles, one of them twenty. On 50–80 nucleotide ssDNA and ssRNA monomers started from random sequence, pLDDT and pTM rise steadily under both backends; long monomers of 200–300 nucleotides reach pLDDT up to 80, as do symmetric dimers at 150 nucleotides. For DNA aptamers of 30–60 nucleotides targeting IL-6 with the protein sequence held fixed, pLDDT and ipTM improve across cycles, and the designed binding site largely overlaps the aptamer site seen in the crystal structure PDB 4NI7.

#Applications

The intended users want a structured nucleic acid for a named target without running a selection campaign: DNA and RNA aptamers for diagnostics, capture reagents and therapeutic leads, structured monomers and symmetric assemblies for nucleic acid nanotechnology, and protein–nucleic acid complexes where both partners are optimized together. Because the loop wraps published checkpoints, a new target requires only a template specification rather than a design dataset. The output is a ranked set of high-confidence candidates for synthesis and binding assays, not validated binders.

#Impact

HalluDesign-NA is a compact demonstration that the hallucination recipe developed for proteins transfers to nucleic acids as soon as a competent nucleic-acid sequence-design network exists, and that all-atom predictors are now good enough on DNA and RNA to serve as the oracle. The limits are real and the authors name them: resolved nucleic acid structures, especially of nucleic acid–protein complexes, are scarce compared with protein structures, and nucleic acid secondary and tertiary structure is sensitive to temperature, buffer and ionic strength in ways structure predictors do not model. Improving pLDDT and ipTM is not evidence of binding, and the framework's value will be settled by the experimental characterization the authors describe as ongoing. The code is public but carries no LICENSE file of its own; its README declares the code MIT and points at the parent repository's license.