When researchers optimize an mRNA therapeutic sequence, most attention goes to the coding sequence: codon usage, GC content, avoidance of CpG dinucleotides, reduction of secondary structures that stall the ribosome. These are real concerns. But the untranslated regions — the 5' UTR upstream of the start codon and the 3' UTR downstream of the stop codon — govern something equally critical: how long the message survives inside the cell before degradation. An optimally translated mRNA that is eliminated within two hours produces less total protein than a moderately translated mRNA with a ten-hour half-life.
UTR engineering has historically been treated as a detail rather than a design problem. That framing is changing. Here is why it matters, what the mechanistic constraints are, and where we think computational design can add something that empirical screening alone cannot.
The 5' UTR: More Than a Ribosome Landing Pad
The 5' UTR's primary function in the context of mRNA therapeutics is to facilitate translation initiation — the cap structure and the region immediately upstream of the AUG codon together determine how efficiently the 43S preinitiation complex loads and scans to find the start site. But the 5' UTR also affects stability in ways that are distinct from its role in translation efficiency.
Cap Structure and Stability Linkage
The m7G cap added to the 5' end of mRNA protects against 5'-to-3' exonuclease degradation. In therapeutic mRNA, modified cap analogues — the ARCA (anti-reverse cap analogue) and its successors such as CleanCap — were developed to improve both capping efficiency and cap stability. A capped mRNA is resistant to XRN1, the major cytoplasmic 5'-to-3' exonuclease. Loss of the cap triggers rapid degradation. Modified caps with better resistance to decapping enzymes such as DCP1/DCP2 directly translate to longer half-life. This is one of the more straightforward stability interventions, and it is well-established in the field — most therapeutic mRNA programs now use CleanCap or similar second-generation cap analogues rather than ARCA.
Kozak Sequence and Upstream Open Reading Frames
The Kozak consensus sequence (the region surrounding the AUG start codon, typically annotated as GCCRCCAUGG where R is a purine) affects translation efficiency but has secondary effects on stability. A strong Kozak context reduces the probability of upstream open reading frame (uORF) initiation, which matters because upstream AUGs that are translated out-of-frame consume ribosomal resources and can induce NMD (nonsense-mediated decay) if the resulting peptides are flagged as aberrant. Engineering the 5' UTR to minimize adventitious uORFs — while maintaining a clean Kozak context for the intended start codon — is a constraint that should be checked computationally before synthesis.
Secondary Structure Effects on Scanning
The 43S complex scans the 5' UTR in a 5'-to-3' direction, and stable secondary structures within the UTR create kinetic barriers to that scanning. Structures with calculated free energy below approximately -30 to -40 kcal/mol in the first 50-100 nucleotides upstream of the AUG are associated with reduced translation efficiency. But the relationship with stability is more complex: some secondary structures protect the 5' end from nucleases by occluding cleavage sites. The optimization tension here is real — you may need to balance structural accessibility for ribosomes against structural protection from degradation machinery. Our computational approach to 5' UTR design uses minimum free energy (MFE) prediction alongside ribosome profiling data from human cell lines to navigate this tradeoff, though we acknowledge that in vitro MFE predictions do not perfectly capture intracellular folding in the presence of RNA-binding proteins.
The 3' UTR: Where Most Half-Life Determination Happens
The 3' UTR is the dominant determinant of mRNA stability in most contexts. This is where RNA-binding proteins dock, where microRNA target sites are embedded, where the poly-A tail is anchored, and where the major instability elements reside. Getting the 3' UTR wrong can eliminate all the gains made by careful coding sequence and 5' UTR optimization.
Poly-A Tail Length and Tail Biology
The poly-A tail at the 3' end of mRNA has two functions: it protects against 3'-to-5' exonuclease degradation (primarily by the exosome complex) and it promotes translation through interaction with poly-A binding protein (PABP), which in turn interacts with the 5' cap-binding complex to circularize the mRNA. Tail length matters: in the literature, poly-A tails of 100-150 adenosines are commonly used in therapeutic mRNA constructs, balancing stability against synthesis efficiency. There are reports of tails up to 250 adenosines providing further stabilization, but the relationship is not simply linear — very long tails can complicate synthesis and may have variable encapsulation efficiency in LNP formulations.
An important development is the recognition that poly-A tail modification — incorporating non-adenosine nucleotides or structural motifs at the 3' end — can increase resistance to deadenylation. The major deadenylases (CCR4-NOT complex, PARN) shorten the tail progressively; deadenylation is the rate-limiting step in the dominant mRNA decay pathway. Chemical or sequence modifications that slow deadenylation directly extend half-life. This remains an active area of research and the field has not converged on a single solution.
AU-Rich Elements and Destabilizing Motifs
AU-rich elements (AREs) in the 3' UTR are recognized by a family of RNA-binding proteins — HuR (stabilizing), TTP/ZFP36 (destabilizing), AUF1 (context-dependent) — that together determine whether the message is stabilized or targeted for degradation. The canonical ARE motif is AUUUA pentamer, with higher-order AUUUAUUUA nonamer repeats having stronger destabilizing activity. These elements evolved in inflammatory cytokine mRNAs to ensure rapid message clearance after acute signaling, which is exactly the opposite of what you want in a therapeutic mRNA meant to persist and translate for days.
Eliminating AREs from the 3' UTR is a standard step in therapeutic mRNA design, but it requires computational scanning rather than manual inspection — a 200-nucleotide 3' UTR may contain several ARE motifs scattered across the sequence. We have seen cases where synthetic 3' UTR designs that looked clean by eye contained buried AUUUA motifs that were not apparent until explicit screening.
Stabilizing Motifs: Learning from Globin and Other Long-Lived Endogenous mRNAs
Alpha- and beta-globin mRNAs are among the most stable endogenous messages in red blood cell precursors, with half-lives measured in hours to days. Their 3' UTRs contain binding sites for alpha-complex proteins (alpha-CP, also called PCBP proteins) that compete with destabilizing factors for occupancy. The alpha-globin 3' UTR, and elements derived from it, have been widely used in therapeutic mRNA constructs for this reason — they are borrowed biology, essentially transplanting a proven stability architecture into the transgene mRNA.
The question we are working through computationally is whether the globin-derived elements are optimal for all tissue contexts. In CNS applications specifically, the RNA-binding protein landscape in neurons is different from erythroid precursors. PCBP expression levels vary by cell type; HuR distribution and nuclear-cytoplasmic shuttling differ between proliferating and post-mitotic cells. A 3' UTR stabilization strategy borrowed from the hematopoietic literature may not perform identically in hippocampal neurons or cortical astrocytes. We don't yet have validated data on this question — it is a hypothesis we are testing in iPSC-derived neuronal models.
AI-Guided UTR Design: What It Adds
The conventional approach to UTR optimization is empirical: synthesize a panel of constructs varying the UTR sequence, transfect into cells (often HEK293T as the screening workhorse), measure expression at multiple time points, select the best performer. This works and has produced many of the UTR sequences in current clinical use. The limitation is throughput and transferability: you can test dozens or hundreds of variants, but the number of relevant sequence combinations is much larger, and the best variant in HEK293T may not be the best in your target cell type.
Computational design adds three specific things. First, it can pre-filter the design space: eliminate sequences with predicted ARE motifs, uORFs, unfavorable secondary structures, or microRNA target sites for miRNAs highly expressed in the target tissue. This reduces the empirical screening burden by removing candidates that are likely to fail before you synthesize them.
Second, it can transfer learning from existing stability datasets. The mRNA stability literature now includes datasets of thousands of sequences with measured half-lives in various cell types. Models trained on these datasets can predict half-life rankings for novel sequences with moderate accuracy — not well enough to replace experimental validation, but well enough to prioritize which sequences to test first.
Third, and this is where our work at caVos specifically applies, computational design can account for tissue-specific RNA-binding protein expression. If you have a model that takes 3' UTR sequence plus cell-type RBP expression profile as inputs and predicts binding site occupancy, you can design UTRs that are matched to the target cell type rather than generically optimized. Our predictions in this domain are preliminary — we are not claiming validated in vivo half-life extension for any specific construct — but the framework is what distinguishes AI-guided design from empirical screening with a computational annotation layer.
Half-Life Targets for CNS Applications
For the longevity-relevant CNS applications that caVos is focused on, mRNA persistence requirements are different from vaccine contexts. A COVID-19 mRNA vaccine is designed to produce antigen transiently — days of expression are sufficient for the immune response. A therapeutic mRNA targeting neuronal protein expression in an age-related neurodegenerative context likely needs to produce protein over weeks to months, which requires repeated dosing or genuinely extended single-dose expression kinetics.
Literature estimates of unmodified mRNA half-life in neurons are in the range of several hours to roughly one day, highly variable by cell type and construct. Optimized constructs using modified nucleotides (pseudouridine, N1-methylpseudouridine), engineered UTRs, and modified poly-A tails have shown extended expression windows in rodent models — some constructs produce detectable protein expression for seven to fourteen days after intracranial delivery. Whether those results translate to primates, and at what dose, is not established.
We are not claiming that UTR engineering alone can achieve therapeutic expression kinetics for CNS longevity targets. The half-life problem is real and probably requires a combination of optimized chemistry, delivery vehicle engineering, and dosing strategy. What we are saying is that UTR design is a non-trivial component of that solution, and it is one where computational guidance can reduce the empirical iteration burden if the models are trained on relevant data.
Where the Field Is Heading
Several directions are active. Circular RNA as an alternative to linear mRNA is the most discussed: cap-independent translation via IRES elements, and intrinsic resistance to exonucleases due to the absence of free ends. The translation efficiency of circular RNA constructs is currently lower than optimized linear mRNA, and the IRES elements required for translation add immunostimulatory concerns that are not yet fully resolved. Whether circular RNA ultimately displaces linear mRNA for longevity applications is an open question.
Self-amplifying mRNA (saRNA) is another direction: the mRNA encodes an alphavirus replicase that amplifies the message intracellularly, extending the effective expression window. The immunogenicity of the replicase and the complex regulatory landscape for saRNA constructs are real challenges, but the expression kinetics are attractive for applications that need sustained protein production.
The UTR engineering problem is not solved. The field has good heuristics — use modified caps, clean AREs, borrow from globin UTRs, use long poly-A tails — but the systematic design of UTRs optimized for specific target tissues and expression duration targets remains an open computational challenge. That gap is where we think there is genuine room for improvement, and it is why we are spending significant effort on the stability modeling layer of the caVos platform.
