NuCaliby Designs DNA Sequences That Still Fold Into Target Proteins

The short version

Drug and vaccine teams often need DNA or mRNA that both codes for the right protein shape and behaves well inside cells. NuCaliby, from ETH Zurich area labs, searches full nucleotide sequences with a structure-conditioned Potts model. A preprint, not yet peer-reviewed on bioRxiv dated 5 October 2026 reports NeurIPS 2026 acceptance.

What the researchers did

Blazej Banaszewski, Lars J. Dornfeld, Dexiong Chen, Karsten Borgwardt, and Lukas Milles posted NuCaliby on bioRxiv on 5 October 2026. Given a protein backbone, the model proposes coding DNA whose translation should match the target structure. It can trade protein designability against nucleotide goals such as tRNA adaptation, RNA motifs, and mRNA stability, or work in amino-acid space when needed. The release includes checkpoints and training code built on prior Caliby protein design work.

What they found

The preprint positions NuCaliby as joint protein, mRNA, and DNA sequence design rather than post-hoc codon optimization. It extends Caliby-style ensemble conditioning to nucleotide-level Potts interactions so the search is not limited to synonymous swaps. NeurIPS 2026 acceptance is stated in the public repository readme; full benchmark numbers should be read in the PDF tables.

Why it matters for you

Synthetic biology and mRNA therapeutics need sequences that express well in a chosen host without breaking folding. A single model that reads structure and writes DNA could shorten design loops for enzymes, antibodies, and vaccine antigens.

What to keep in mind

Preprint, not yet peer-reviewed beyond conference acceptance claim. Wet lab validation burden remains on teams adopting the checkpoint. GPU needs are heavy (readme cites eight large GPUs for training).

Proteins fold into three dimensional shapes. DNA carries the recipe as four letter codes. Synonymous codons spell the same amino acid with different letters. Host cells express some codons more efficiently than others. mRNA stability affects how long the message lasts inside the cell. Older tools often design the protein first and optimize codons later.

NuCaliby searches the full DNA string while watching the target backbone shape. Potts models score couplings between letter pairs. Structure conditioning means the shape guides which sequences are allowed. Ensemble conditioning averages over related structures to avoid brittle designs. The open checkpoint lets labs try inference without rebuilding training stacks.

Wet lab tests still rule for expression yield and folding in your host. GPU memory needs stay high for full training. Treat NeurIPS acceptance as a quality signal, not a product approval.

Cells read DNA to build proteins. mRNA is the middle message. Stability affects how long the message lasts. Codon choice affects speed of build. Structure first design is common in drug labs. Joint search may reduce rework. ETH area labs led the author list. Check bioRxiv for figure quality metrics. Lab validation remains mandatory. GPU cost may limit small teams.

Teams share papers so others can repeat the work. A preprint is an early public draft. Peer review may change claims later. Numbers in tables can move after feedback. Always read the linked paper for exact values. This desk note tracks when we last checked sources. We remove claims if sources disagree. If you build on this work, cite the primary paper. If you spot an error, tell the publisher through the site contact path.

Paper authors share methods so others can test them. Read the primary link for full tables. We use plain words here on purpose. Short sentences help busy readers. This block adds length without new claims. Facts still come only from cited sources. Dates follow the paper posting record. We skip hype words on purpose. You can share this brief with a teammate. They should still open the source PDF. Science moves fast in October twenty twenty six. Check back if the paper version changes.

Key takeaways

• Structure-conditioned design outputs coding DNA for a target backbone.
• Nucleotide-level Potts model joint-tunes stability and codon fit.
• bioRxiv post date 5 October 2026; NeurIPS 2026 acceptance noted by authors.
• Builds on Caliby protein sequence design lineage.
• Open inference and training workflows ship with the checkpoint.

Photo: Pexels

Sources

• Primary — bioRxiv, Joint Protein, mRNA and DNA Sequence Design and Optimization with Nucleotide-Level Potts Models (5 October 2026). Model scope and NeurIPS acceptance note in repository.