mRNAutilus: Multi-Objective mRNA Design

• mRNAutilus is a multi-objective framework that generates full-length mRNAs with tailored therapeutic properties by optimizing both coding sequences and untranslated regions.
• It employs a masked discrete diffusion model integrated with Monte Carlo Tree Search to guide the design process under multiple functional objectives.
• Benchmark evaluations show enhanced mRNA stability, translation efficiency, and protein yield, outperforming established baselines in various therapeutic contexts.

mRNAutilus is a multi-objective, sequence-based generative framework for the discrete generation and optimization of full-length messenger RNAs (mRNAs) with tailored therapeutic properties. The system integrates masked discrete diffusion modeling with Monte Carlo Tree Search (MCTS) guidance, enabling the simultaneous optimization of coding sequences (CDS) and untranslated regions (UTRs) under multiple functional objectives. mRNAutilus is designed to address complex design constraints in therapeutic mRNA engineering, encompassing transcript stability, translation efficiency, and protein yield, and extends applicability across diverse biological and clinical payloads, including protein reporters, antigens, genome editors, and modulators of intracellular pathways (Patel et al., 29 May 2026).

1. Model Architecture and Masked Discrete Diffusion

mRNAutilus employs a masked discrete diffusion model in which mRNA sequences are represented as $\mathbf{x} = (x^1, \dots, x^L) \in \mathcal{V}^L$, for vocabulary $\mathcal{V}$ combining codons (CDS) and nucleotides with IUPAC degenerate symbols (UTRs), plus special tokens. Sequence generation is formulated as a continuous-time Markov process, progressively introducing and then removing masking noise. The forward masking process for a token is specified by

$$p_{t|0}^{\rm mask}(x_t \mid x_0) = \mathrm{Cat}(x_t; t x_0 + (1-t)m)$$

for $t \in [0,1]$, with $m$ as the [MASK] vector. The reverse process retains unmasked tokens:

$$p_{s|t,0}(x_s\mid x_t, x_0) = \begin{cases} (1-\frac{s}{t})x_0 + \frac{s}{t}m & x_t = m \ \delta(x_s - x_t)&x_t \neq m \end{cases}$$

A denoising Transformer network with 150M parameters, SwiGLU feed-forward layers, 20-headed attention, rotary position encodings, and FlashAttention-2 performs imputation. Hybrid tokenization provides codon-level representation for CDS and nucleotide-level for UTRs. The model is trained with a continuous-time negative log evidence lower bound (NELBO) objective,

$$\mathcal L(\theta) = \mathbb E_{t \sim \mathcal U(0,1), x_0 \sim p_{\rm data}, x_t \sim p_{t|0}} \left[ -\frac{1}{t} \log \langle x_0, x_0^\theta \rangle \right]$$

where $x_0^\theta$ is the categorical prediction over $\mathcal{V}$.

2. Monte Carlo Tree Search (MCTS) for Multi-Objective Sequence Guidance

Sampling from the generative model is orchestrated using MCTS to explore feasible unmaskings, integrating model predictions and property scores at each node. Each node corresponds to a partially unmasked $\mathbf{x}_t$ at diffusion step $\mathcal{V}$0. An iteration consists of:

1. Selection: Traverse the tree via a vectorized Upper Confidence Bound (UCB) criterion, derived for each child as

$\mathcal{V}$1

using exploration constant $\mathcal{V}$2.

1. Expansion: Candidate unmaskings are drawn from the denoiser logits (with Gumbel sampling) for $\mathcal{V}$3 mask positions.
2. Rollout: Complete unmasking by ancestral sampling generates a proposed full sequence $\mathcal{V}$4.
3. Backpropagation: Rewards (based on Pareto dominance across objectives) propagate up the tree.

The global Pareto set $\mathcal{V}$5 of non-dominated, fully unmasked sequences is maintained and updated in each iteration, and rollouts are rewarded by the degree of Pareto front expansion for individual objectives:

$\mathcal{V}$6

3. Multi-Objective Property Prediction and Pareto Optimization

Predicted functional properties for each candidate mRNA are provided by three XGBoost regression models trained on pooled model embeddings:

• Objectives: Predicted log$\mathcal{V}$7 half-life (Agarwal & Kelley, Genome Biol 2022), translation efficiency (TE; polysome profiling, Zheng et al. Nat Biotech 2025), and log protein abundance (Eichhorn et al. Mol Cell 2014).
• Performance: For validation sets, $\mathcal{V}$8: half-life 0.551, TE 0.716, abundance 0.565.
• Optimization: The system expands the Pareto set,

$\mathcal{V}$9

for dominance defined as $$p_{t|0}^{\rm mask}(x_t \mid x_0) = \mathrm{Cat}(x_t; t x_0 + (1-t)m)$$0 when $$p_{t|0}^{\rm mask}(x_t \mid x_0) = \mathrm{Cat}(x_t; t x_0 + (1-t)m)$$1 for all $$p_{t|0}^{\rm mask}(x_t \mid x_0) = \mathrm{Cat}(x_t; t x_0 + (1-t)m)$$2 and strictly for at least one $$p_{t|0}^{\rm mask}(x_t \mid x_0) = \mathrm{Cat}(x_t; t x_0 + (1-t)m)$$3. No scalarization is performed; sampling directly targets the Pareto frontier in multidimensional property space.

4. End-to-End Generative Workflow

The sequence generation process consists of:

1. Tokenization: Input templates map into masked token arrays $$p_{t|0}^{\rm mask}(x_t \mid x_0) = \mathrm{Cat}(x_t; t x_0 + (1-t)m)$$4 for diffusion initialization. CDS masks are constrained to preserve amino acid output.
2. MCTS-guided Diffusion: Multiple MCTS iterations are executed with $$p_{t|0}^{\rm mask}(x_t \mid x_0) = \mathrm{Cat}(x_t; t x_0 + (1-t)m)$$5 reverse diffusion steps to iteratively unmask the sequence.
3. Pareto Selection: The process returns the final non-dominated set $$p_{t|0}^{\rm mask}(x_t \mid x_0) = \mathrm{Cat}(x_t; t x_0 + (1-t)m)$$6 of fully unmasked, property-optimized mRNA sequences.

A detailed pseudocode is presented in Algorithm B.1 of the reference, with explicit procedures for selection, expansion, rollout, and backpropagation steps (Patel et al., 29 May 2026).

5. Empirical Evaluation and Performance Metrics

mRNAutilus was benchmarked in silico and in vitro against both random design and state-of-the-art baselines (GEMORNA AR, GenScript):

• In Silico: For 250 firefly luciferase (Fluc) mRNAs, mRNAutilus outperforms controls on coding features (CAI, MFE, optimal-codon frequency, GC%, U-content) and UTR thermodynamic stability.
• Luciferase Reporter: Top design (ABW2) in HEK293T confers $$p_{t|0}^{\rm mask}(x_t \mid x_0) = \mathrm{Cat}(x_t; t x_0 + (1-t)m)$$7400$$p_{t|0}^{\rm mask}(x_t \mid x_0) = \mathrm{Cat}(x_t; t x_0 + (1-t)m)$$8 luminescence at 48h over wild-type and surpasses machine design baselines in multiple cell lines.
• SARS-CoV-2 Spike: Zero-shot designs outperform BNT162b2 (Pfizer), GenScript Spike v2, and GEMORNA GMR-CV-F2 in A549, with matched or better durability.
• PEMax Prime Editor: ABW-PEMax-2 demonstrates highest T→A edit rate at the HEK3 locus, exceeding commercial GenScript mRNA.

The following table summarizes key quantitative performance relative to wild-type and baseline controls:

Gene	half-life (design − WT)	TE (design − WT)	abundance (design − WT)
Fluc	+0.537	+0.283	+0.769
Spike	+1.074	+0.236	–0.727
MUC1	+0.593	+0.092	–0.906

For library statistics (Table 5.4):

Target	Baseline HL	Designed HL (mean±SD)	Baseline TE	Designed TE (mean±SD)
Fluc	0.681	0.951±0.058	0.104	0.283±0.056
Spike	0.412	0.673±0.059	0.058	0.264±0.020
PEMax	0.593	0.586±0.033	0.022	0.026±0.011
uAb	0.060	0.462±0.027	–0.551	0.024±0.009

6. Applications in Therapeutic mRNA Engineering

• Prime Editing (PEMax): Zero-shot mRNAutilus constructs expressing Cas9-RT fusions (ABW-PEMax) achieve expression and functional genome editing efficiency (T→A at HEK3) matching or exceeding commercial benchmarks in vitro.
• Programmable Proteome Modulation (Ubiquibodies): mRNAutilus-optimized mRNAs encoding β-catenin-targeting peptide-guided E3 ligases (uAbs) confer 3–5$$p_{t|0}^{\rm mask}(x_t \mid x_0) = \mathrm{Cat}(x_t; t x_0 + (1-t)m)$$9 transcript enhancement over α-globin controls, suppress Wnt/β-catenin signaling in reporter assays, and trigger proteasome-dependent endogenous β-catenin degradation in DLD-1 cells.

7. Limitations and Prospective Developments

The reliability of regressor models for functional property prediction poses a constraint on overall guidance, owing to limited scope and heterogeneity in public fitness datasets for mRNA (cell-type bias, scale). Future improvements may involve integrating state-of-the-art discrete diffusion guidance frameworks such as TR2-D2, MOG-DFM, or AReUReDI for increased sampling efficiency and solution diversity. Expanding uniform, large-scale design-build-test-learn datasets across key axes (half-life, TE, abundance, immunogenicity, tissue specificity) is anticipated to enhance both the core generative capacity and property prediction accuracy. Anticipated applications include mRNA vaccines, protein replacement therapeutics, gene/base/prime editing, and broader proteome editing architectures (uAbs, duAbs).

mRNAutilus represents the first end-to-end, multi-objective-guided masked discrete diffusion approach for full-length mRNA co-optimization, achieving state-of-the-art zero-shot design performance across diverse therapeutic classes and establishing a foundation for programmable mRNA therapeutics by direct sequence-level generative modeling (Patel et al., 29 May 2026).