henhaobio
Sign in

bioRxivBiophysicsPreprint

Joint protein, mRNA and DNA sequence design and optimization with nucleotide-level Potts models

Błażej BanaszewskiiD, L DornfeldiD, Dexiong CheniD, Karsten Michael BorgwardtiD, Lukas F. MillesiD

PDFDOICode
In-depth full-text digest available · Jump to digest ↓

Abstract

Protein design typically generates amino acid sequences first and then translates them into DNA post-hoc limiting control over the nucleotide sequence via codon choice at design time. We introduce NuCaliby, a structure-conditioned model that jointly designs amino acid, mRNA, and DNA sequences of a protein at nucleotide resolution. NuCaliby predicts a Potts model over a nucleotide graph, trained using amino acid supervision by marginalizing over synonymous codons. It preserves residue-level designability on par with the original Caliby model. Its energy-based formulation supports composable inference-time guidance from sequence-level objectives through proposal construction and candidate acceptance, allowing the incorporation of biological proxies into the coding nucleotide sequence. Under organism-specific tRNA adaptation guidance, NuCaliby improves adaptation to host tRNA pools beyond synonymous sequence space, trading-off self-consistency. NuCaliby also embeds specified RNA motif sequences directly into coding sequences, avoiding the combinatorial growth required for the same task at the amino-acid level. In wet lab experiments, NuCaliby produces expressible and soluble proteins on par with state of the art methods, while enabling more complex nucleotide-level control and optimization. Together, these results extend protein design beyond the amino acid level, enabling joint optimization across the central dogma of molecular biology. The code and model weights are available here: https://github.com/blazejba/NuCaliby

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (6 Oct 2026) · CC BY 4.0

A structure-conditioned nucleotide-level Potts model designs protein and coding sequences jointly, enabling tRNA adaptation and RNA motif embedding beyond synonymous codons with designability comparable to Caliby.

Protein design pipelines pick an amino-acid sequence first and optimize its DNA encoding afterward, which limits DNA- and RNA-level objectives to synonymous codons; CodonMPNN predicts codons but offers no inference-time control. NuCaliby models nucleotides directly with composable energy guidance, so codon adaptation and RNA motifs become design objectives without retraining.

Key results

  1. ClaimedSampling from either the residue-level or nucleotide-level Potts head, NuCaliby matches Caliby on sequence recovery and AlphaFold2 self-consistency on 183 monomers; ensemble conditioning improves designability further.

    Authors report that NuCaliby AA and NT sampling match Caliby on recovery and designability, with ensemble conditioning improving designability further; Table 1 values are not in the text. By contrast, CodonMPNN has TM .441 vs .509 and pLDDT 47.0 vs 54.3 relative to ProteinMPNN. · 4.4 Codon-level organization and protein…; F Effect of training data on designability

  2. ShowntAI guidance raises adaptation to the E. coli BL21(DE3) tRNA pool beyond what synonymous optimization of Caliby sequences reaches, at a growing cost to self-consistency.

    Mean tAI 0.572 for Caliby + DNA Chisel vs 0.660 at λ = 300 and 0.725 at λ = 600 for NuCaliby NT; one sequence per protein on 183 backbones; TM-score falls with stronger guidance. · 4.2 Adaptation to organism-specific tRNA pools; A.3 Adaptation to organism-specific tRNA pools

  3. ShownNucleotide-level motif guidance embeds SiRA aptamer and HIV-1 frameshifting element sequences into coding sequences, while residue-level Caliby sampling yielded no motif-containing sequence for any of the 183 proteins in the baseline.

    SiRA and FSE give 264 and 129 codon variants (68 and 31 amino-acid patterns). Caliby: no motif in any of 183 proteins at n=16 (the appendix describes n=64 baseline samples). NuCaliby: d=0 on all 183 by n=2 (one motif) and n=16 (two motifs); 91.8% exact at best-of-1 for SiRA. · 4.3 Inclusion of mRNA sequence motif; A.5 Sensitivity to sampling hyperparameters; NuCaliby

  4. ShownViennaRNA ensemble free energy used as an acceptance-only objective shifts designs toward more stable predicted mRNA folding without local substitution scores.

    At λ=10, guided designs beat the unguided baseline in predicted ensemble free energy per nucleotide for 182 of the 183 backbones; mean shift −0.07 kcal/(mol × nt). · 4.4 mRNA folding energy optimization

  5. SuggestedModerate tAI guidance raised soluble yield across all eight tested backbones, and designs were expressible without post-hoc codon optimization; motif-conditioned designs were expressible in several cases.

    Median yield rose from 0.042 to 0.052 mg/mL (about 24%) at λ = 30; exact one-sided Wilcoxon signed-rank n = 8 paired backbones, W = 36, p = 0.0039; two designs per backbone averaged. · 4.5 In vitro wet lab validation; B.1 Creating sequences

Methods, data and reproducibility

A shared structural encoder (AtomMPNN) predicts residue-level and nucleotide-level Potts parameters, trained as a masked denoiser with loss marginalized over synonymous codons. Sampling uses the DLMC discrete MCMC sampler with energy guidance via proposals and Metropolis-Hastings acceptance.

  • Training set: PDB structures from the Boltz-1 training split, temporal split, clustered at 40% sequence identity with MMseqs2, transmembrane proteins excluded · n = ~ 231k PDB structures
  • Test set: soluble monomer backbones from the Caliby benchmark (PDB IDs listed in Appendix A.6) · n = 183 backbones
  • tAI weights from tRNA gene copy numbers of E. coli BL21(DE3); CodonMPNN conditioned on species-level E. coli · Escherichia coli · NCBI tax ID 469008 (BL21(DE3)); 562 (E. coli)
  • Wet-lab backbones: four published de novo backbones and four RFdiffusion backbones · n = 8 backbones, 24 sequences per backbone per method, 2 designs tested
Methods & tools
NuCaliby, Caliby, ProteinMPNN, CodonMPNN, DNA Chisel, Protpardelle-1c, AlphaFold2 via ColabFold, TM-align, ViennaRNA, MMseqs2, AtomWorks, GenScript and Twist synthesis
Software & environment
Designability: AlphaFold2 via ColabFold, single-sequence mode, 4 recycles, model 1. Sampling: DLMC, 500 sweeps, final temperature 0.01. Single checkpoint trained on the Boltz-1 split.
Assay conditions & compute
Pre-training takes about 15 hours on 8 × NVIDIA B200 GPUs. Inference runs on a single H100, about 1-2 s per design; AtomWorks preprocessing takes 10-20 hours on a multi-CPU node.
Reproducibility gaps
Model weights are said to be available with the code; no separate data URLs or package version pins are given in the text.
Limitations (5)
  • Wet-lab assays measure expression and solubility only; translation efficiency, RNA motif activity and cellular mRNA stability were not measured, and the objectives are simplified sequence-level proxies. · stated by the authors
  • Synonymous invariance is only approximate: median per-protein mean absolute energy change is 0.085 for synonymous vs 3.54 for non-synonymous substitutions, and median normalized KL divergence from uniform is about 0.69 at T = 0.01 without guidance. · stated by the authors
  • Stronger tAI guidance reduces self-consistency and TM-score; strongly guided and CodonMPNN sequences were often too repetitive for synthesis (more than 50%). · stated by the authors
  • Wet-lab evidence is small: 8 backbones and 2 designs per method, a yield readout from size-exclusion traces, and the authors call it an initial indication. · noted in this summary
  • tAI gains are scored with the same weights used as the guidance objective, so the in silico improvement is partly by construction and not an independent measure of translation. · noted in this summary

Read this if you design proteins and want inference-time control of codon usage, embedded RNA motifs or mRNA folding energy without retraining.

Citations
0

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISShuai '25Fang '26Dai '26Rajbanshi '25Birnbaum '26Zhu '26
64% semantic matchBioengineering
Richard W. Shuai, Ling Zhou, Gina El Nesr+7 more·bioRxiv·Sep 30, 2025· 6 citations