mRNAutilus: Multi-Objective-Guided Discrete Generation of mRNA with Optimized Therapeutic Properties

Abstract: Therapeutic mRNA design requires coordinating multiple interacting sequence features across the full transcript, where codon usage, untranslated regions (UTRs), and their coupling jointly determine stability, translation efficiency, and protein expression. Here, we present mRNA generation via unrolled trajectories and informed latent updates (mRNAutilus), a framework for simultaneous codon optimization and de novo UTR design directly from sequence. mRNAutilus combines a masked discrete diffusion model trained on millions of full-length mRNAs with Monte Carlo Tree Guidance to generate Pareto-efficient sequences under multiple functional objectives, using lightweight regressors over model embeddings to predict half-life, translation efficiency, and protein abundance. Unlike recent methods that design coding sequences and UTRs separately or rely on post hoc assembly and screening, mRNAutilus generates complete transcripts in a single process optimized across properties. Across diverse targets, zero-shot mRNAs encoding P. pyralis luciferase achieve over 400-fold higher expression than wild-type and outperform commercial and machine learning-designed baselines, including zero-shot generative approaches. Zero-shot SARS-CoV-2 Spike mRNAs exceed clinically used and commercial constructs and match or surpass lab-optimized designs with improved durability. We further demonstrate generality in therapeutic settings, including prime editing (PEMax) and programmable proteome modulation, where mRNAutilus-designed constructs enhance expression of peptide-guided E3 ligases (uAbs) for beta-catenin degradation. These results establish a sequence-based, multi-objective framework for generating functional mRNAs tailored to diverse biological applications.

• The paper introduces mRNAutilus, a framework that combines masked discrete diffusion modeling with Monte Carlo tree guidance for multi-objective optimization of therapeutic mRNAs.
• It employs innovative hybrid tokenization with codon-level and nucleotide-level encoding to accurately model CDS and UTRs, ensuring realistic biological sequences.
• Experimental validations reveal up to 400-fold increased expression and superior performance compared to baseline designs, underscoring its therapeutic potential.

Multi-Objective Guided Generation of Therapeutic mRNAs with mRNAutilus

Model Architecture and Pretraining

mRNAutilus implements a BERT-style masked discrete diffusion model (MDM) for full-length mRNA transcript generation. The encoder leverages hybrid tokenization, encoding CDS regions at codon-resolution and UTRs at nucleotide-resolution, supporting simultaneous and independent optimization. The model was trained on 14.2 million vertebrate mRNAs from Ensembl, filtered for valid CDS and UTR composition. Rotary Positional Embeddings and FlashAttention-2 are incorporated for efficiency and long-range contextual modeling. The vocabulary encodes standard nucleotides, codons, IUPAC degenerate nucleotides, and structural motifs, supporting realistic biological sequence representation.

Unconditional generation trials confirm that mRNAutilus produces transcripts resembling natural distributions in sequence length, GC content, Kozak motif frequency, minimum free energy (MFE), and sequence diversity metrics—demonstrating the prevention of token collapse and capturing long-range dependencies (Figure 1).

Figure 1: Distributions of GC content, length-normalized MFE, Kozak motif frequencies, sequence diversity, and Shannon entropy for unconditional mRNAutilus generations, closely mirroring natural Ensembl data.

Embedding-Based Property Prediction

Internal representations extracted from the MDM are highly discriminative, enabling downstream functional prediction. mRNAutilus embeddings, projected to two principal components, distinguish mRNAs from ncRNAs with high fidelity (Figure 2). Lightweight XGBoost regressors trained on these embeddings for half-life, translation efficiency, and protein abundance achieve validation $R^2$ scores up to 0.716, outperforming KNN baselines and most competing nucleic acid LLMs except Evo-2 (Table S8). These regressors serve as reward functions for generative guidance.

Figure 2: PCA demonstrates separability of mRNAutilus embeddings for mRNA and ncRNA classes.

Multi-Objective Sequence Optimization via Monte Carlo Tree Guidance

Generation is steered through Monte Carlo Tree Guidance (MCTG), which unrolls discrete diffusion trajectories and selects parent and child sequences across multiple objectives: half-life, translation efficiency, and protein abundance. MCTG exploits Pareto-optimality by expanding and rolling out candidate sequences, scoring and updating nodes based on dominance relations and cumulative reward—enabling simultaneous codon and UTR optimization in a single framework (Figure 3).

Figure 3: Schematic of mRNAutilus workflow: MDM pretraining and generation, embedding-based property regression, MCTG-guided Pareto optimization, and experimental validation.

Optimization curves demonstrate rapid and sustained increases in all functional property scores for de novo design tasks, including out-of-distribution gene targets (P. pyralis luciferase, SARS-CoV-2 Spike, and human MUC1). Guided generations yield marked improvements in predicted half-life, translation efficiency, and protein abundance over wild-type and baseline designs, with Pareto fronts exhibiting robust stability across hyperparameters (Figures 2, 10, 11).

Figure 4: Validation and training correlations for property regressors alongside optimization trajectories for guided luciferase generation.

Figure 5: Optimization trajectories for guided conditional generation of MUC1 and SARS-CoV-2 Spike mRNAs.

Figure 6: Ablation analysis reveals MCTG robustness to iteration and Pareto front size variation.

In silico benchmarking against GEMORNA and random design baselines further confirms superior codon adaptation index, optimal codon frequencies, uridine content, and thermodynamic metrics for both CDS and UTR regions, indicative of enhanced translational potential and sequence diversity (Figure 7).

Figure 7: Benchmarking of in-silico luciferase CDS and UTR designs by mRNAutilus against GEMORNA and random baselines on key translation metrics.

Experimental Validation and Functional Efficacy

Experimental evaluations establish mRNAutilus-generated mRNAs as potent therapeutic candidates. Zero-shot designs were synthesized for Fluc, SARS-CoV-2 Spike, PEMax prime editor, and designer uAbs for β-catenin degradation, then assayed in diverse cell contexts.

For P. pyralis luciferase, mRNAutilus designs showed 400-fold higher expression than wild-type in HEK293T and consistent outperformance across Jurkat, A549, and HepG2—superseding commercial and ML-generated sequences in absolute and normalized expression (Figure 8).

Figure 8: Experimental workflow and quantitative expression results for luciferase, Spike, and PEMax mRNAs, showcasing dramatic improvement from mRNAutilus designs over commercial and baseline sequences.

Zero-shot Spike mRNAs reached or exceeded levels produced by clinical (BNT162b2) and commercial benchmarks, with some designs rivaling lab-in-the-loop optimizations in expression, stability, and durability. PEMax designs exhibited significantly higher editing efficiency at HEK3 compared to commercial and ML-generated controls.

Optimization of designer ubiquibodies (uAbs) via mRNAutilus resulted in higher transcript abundance and robust Wnt/β-catenin suppression as measured by qPCR, TOPFlash luciferase assays, and quantitative immunoblotting. Proteasome-dependent β-catenin degradation was confirmed and quantified, validating both expression and functional targeting (Figure 9).

Figure 9: Architecture of uAb degradation system, transcript expression, reporter assay outcomes, and protein knockdown efficacy for mRNAutilus-optimized therapeutic mRNAs.

Sequence and Structure Analyses

Generated UTRs—crucial for translation initiation and mRNA stability—exhibit GC-rich compositions, favorable MFE, and highly diverse secondary structure relative to natural UTRs, underscoring productive exploration of novel functional sequence space (Figure 10). Pretraining metrics further indicate broad coverage of mRNA features across vertebrate taxa (Figure 11).

Figure 10: Comparative analysis and clustering of generated vs. natural 5' and 3' UTRs for luciferase, evidencing structural and compositional diversity.

Figure 11: Pretraining corpus statistics: length, GC content, frequency of Kozak motifs, UTR length distributions, and species coverage.

Implications and Future Directions

mRNAutilus represents a unified generative solution for multi-property optimization in therapeutic mRNA design, integrating masked diffusion with Pareto-guided search. Its applicability to antigen, gene editor, and degrader payloads across multiple cell types without lab iteration demonstrates broad generalizability and practical utility, particularly under limited data regimes. The approach is gradient-free and robust to small regressor training sets—a major advantage given the sparsity and heterogeneity of public mRNA datasets.

Algorithmic advances such as multi-objective discrete flow matching and trajectory-aware discrete diffusion (TR2-D2, MOG-DFM, AReUReDI) could further augment performance. The principal limitation is regressor accuracy, which is dictated by the scale and fidelity of available experimental datasets. The systematic collection of multi-property mRNA measurements across cell types, tissues, and functional endpoints will be essential to refine the guiding reward signals and the scope of generative control.

In terms of theoretical impact, this work codifies Pareto-optimality in full-length biomolecular sequence generation, moving beyond isolated component design to encompass synergistic region coupling via discrete generative modeling and reward-driven tree search. Practically, it provides a tractable blueprint for programmable sequence design in mRNA therapeutics, including mRNA vaccines, gene editing reagents, transient protein replacement therapies, and peptide-guided proteome modulators.

Conclusion

mRNAutilus introduces an effective multi-objective-guided framework for discrete generation and optimization of full-length therapeutic mRNAs. Masked diffusion modeling combined with Pareto-efficient tree search enables joint codon and UTR programming for enhanced stability, translation, and expression, supported by embedding-based regressors. Experimental validation across diverse modalities demonstrates functional superiority to commercial, wild-type, and baseline ML-driven designs, with future potential for broader therapeutic applications contingent on improved property prediction and larger mRNA datasets.