UTR-Diffusion

When an mRNA therapeutic is designed, the protein it encodes is usually settled before sequence work begins. What remains adjustable is the 5' untranslated region, the handful of nucleotides on either side of the start codon, and the synonymous codon choices throughout the coding sequence. Those degrees of freedom have to satisfy several quantities at once — ribosome loading, the local folding stability that governs start-codon accessibility, and codon adaptation to the host — while leaving fixed elements such as a Kozak sequence, a DRACH motif, or the encoded peptide itself untouched. Existing tools split that problem: predictor-guided search pushes a single score as high as it will go, LinearDesign jointly optimizes folding and codon usage, and DNAChisel enforces user-declared constraints. None of them hits a specified value of one objective while simultaneously preserving exact nucleotides and amino-acid identity.

UTR-Diffusion, released as a preprint in August 2026 by Chuankai Dai and Kengo Sato at the Institute of Science Tokyo, puts all three kinds of control into a single trained network and applies them at sampling time. Generation is conditioned on continuous-valued targets for mean ribosome load (MRL) and minimum free energy (MFE); sequence constraints and codon-usage bias are enforced by modifying the reverse-diffusion trajectory rather than by refitting the model. One checkpoint is therefore swept across a whole grid of design specifications, and a user who changes their MRL target does not retrain anything.

#Key Features

• Continuous target conditioning: MRL and MFE enter as real-valued scalars, embedded into a 200-dimensional vector that modulates the denoiser's intermediate features. Classifier-free guidance combines conditional and unconditional predictions to sharpen adherence to the request.
• Exact nucleotide preservation: a RePaint-inspired resampling scheme forward-diffuses the user's fixed subsequence to the current noise level and re-injects it at every denoising step, with periodic re-noising so the constrained and unconstrained regions stay globally coherent.
• Amino-acid constraints with synonymous flexibility: a constrained residue becomes a set-valued constraint over its synonymous codons, and at each step the codon closest to the current three-nucleotide state in squared Frobenius distance replaces it — identity is preserved without freezing the codon.
• Graded codon adaptiveness: a one-parameter exponential tilt of the human codon-usage distribution sets a target expectation for relative codon adaptiveness, reweighting synonymous selection so that sequence-level codon adaptation index (CAI) moves on demand.
• Junction-aware scope: the same machinery covers the 5' UTR–CDS boundary, where UTR structure and the first few codons jointly shape translation initiation.

#Technical Details

The denoiser is a U-Net operating on a 4×L bipolar one-hot matrix, with entries of +1 for the present nucleotide and −1 elsewhere, run over 200 diffusion timesteps. Timestep and condition embeddings pass through two-layer MLPs with GELU and SiLU activations respectively. Training used 260,000 unique 50-nucleotide randomized 5' UTRs from the HEK293 massively parallel reporter library of Sample et al., paired with experimentally measured MRL and RNAfold-computed MFE; 10–20% of condition labels were masked during training so the same network learns both conditional and unconditional denoising. Evaluation swept MRL from 4.0 to 8.0 in steps of 0.4 against MFE from −20 to 0 kcal mol⁻¹ in steps of 2, giving 121 joint targets with 100 sequences generated per target and scored using 5' UTR-LM for MRL and RNAfold for MFE. In the junction task — 50-nucleotide designs comprising a 26-nucleotide UTR, a fixed AUG and seven codons of a reference peptide — adaptiveness targets of 0.85, 0.90 and 0.95 produced mean CAI values of roughly 0.82, 0.88 and 0.94.

#Applications

The setting the model is built for is mRNA payload engineering with a predetermined protein: vaccine and protein-replacement constructs, where a designer needs a panel of candidate 5' UTRs at a chosen expression level rather than a single maximal one, and where a manufacturing or regulatory requirement fixes particular bases. Its ability to request an intermediate MRL also suits reporter and circuit tuning in synthetic biology, where overshooting expression is as much a failure as undershooting it.

#Impact

UTR-Diffusion's contribution is compositional: it shows that continuous multi-property targeting, hard sequence constraints and graded codon bias can share one diffusion model instead of requiring three separate tools. That distinguishes it from mRNAutilus, the closest catalog entry by task, which designs full-length transcripts with masked discrete diffusion steered by Monte Carlo tree search, and from broader RNA generative backbones such as RNAGenesis. The evaluation is entirely in silico — MRL comes from a predictor and MFE from a folding algorithm, with no reported wet-lab validation — and the work is a preprint awaiting peer review. Trained weights are distributed under an MIT license on Hugging Face; the implementation repository named in the paper's availability statement has not yet been made public.