Most discussions of mRNA therapeutics sequence design focus on getting protein expressed at all. For vaccine applications — where you want a spike protein or antigen produced transiently in muscle cells to prime an immune response — the design problem is relatively tractable. You need reasonable expression for a few days, you can tolerate some innate immune activation because the inflammation actually helps adjuvant the vaccine response, and the dose-response relationship is forgiving: more spike protein generally means better immune priming.
Longevity-associated protein upregulation is a fundamentally different design problem. You want sustained expression over months to years, in tissues where chronic innate immune activation would be harmful, at expression levels that are physiologically meaningful but within a safe range. That last constraint — staying within a safe expression window rather than maximizing expression — is what makes the computational sequence design component genuinely difficult.
Why Upregulation Is Harder Than Knockdown
Gene silencing therapeutics (ASOs, siRNA) have a comparatively tractable safety framework. You add more drug, you get more knockdown — the dose-response is approximately monotone in the relevant range. You design for maximum knockdown with minimum off-target effects. Reversibility is built in: when drug is cleared, target expression recovers. If you silence the wrong gene, you stop dosing and the situation resolves.
Protein upregulation via mRNA does not work the same way. For many signaling proteins, the biological response is not proportional to protein level across a wide range. Klotho, for example, operates as a co-receptor in FGF23 signaling and as a shed soluble factor with distinct signaling properties — the relationship between Klotho level and downstream pathway activity is non-linear and tissue-context-dependent. Expressing Klotho at 2× endogenous levels may be beneficial; expressing it at 20× may activate compensatory feedback or cause off-target Wnt suppression at tissues where Wnt activity is needed for homeostasis.
FOXO3 is even more context-dependent. FOXO3 activity is regulated by nuclear-cytoplasmic shuttling in response to insulin/IGF-1 signaling, and its downstream effects depend heavily on which co-factors are present in a given cell type. Overexpression studies in model organisms show inconsistent results: benefit in some tissues and dose ranges, toxicity in others. The therapeutic window for mRNA upregulation of a transcription factor is genuinely narrow, and extrapolating from mouse overexpression data to human therapeutic doses requires being honest about how poorly we understand the precise quantitative relationships.
Codon Optimization: Beyond Codon Usage Tables
Codon optimization for mRNA therapeutics has evolved substantially beyond simply substituting synonymous codons to match human codon usage frequency tables. The problems with naive frequency-based optimization are well-documented: optimizing all codons to the most frequent synonymous codon can cause ribosome collisions by eliminating the "pausing" signals encoded in rare codon clusters that coordinate co-translational folding. Removing these pauses can result in misfolded protein even when overall translation rates are high.
Our sequence design model incorporates several features beyond codon frequency. We model ribosome density profiles based on codon decoding rates (which correlate with, but are not identical to, codon frequency) and try to preserve regions of slower translation speed near domain boundaries and major secondary structure transitions in the protein. The 2012 work from Dana and Tuller on co-translational folding, and more recent ribosome profiling datasets generated in human cell lines, give us empirical data on where pausing is biologically functional versus where it is neutral.
We also explicitly check for internal Kozak sequences, cryptic splice sites, RNA-binding protein recognition motifs in the coding sequence, and G-quadruplex forming sequences — all of which can interfere with translation efficiency or mRNA stability if inadvertently introduced by optimization. These are not hypothetical concerns: several published codon optimization case studies have identified silent mutations in coding sequences that dramatically altered expression by creating or destroying these regulatory features.
UTR Engineering for Half-Life Control
For therapeutic mRNA, the 5' and 3' untranslated regions are not afterthoughts. They are primary determinants of mRNA half-life in vivo and translation efficiency.
The 5' UTR must present a stable secondary structure that allows ribosome scanning without being so stable that it stalls the 43S pre-initiation complex. The optimal ΔG for 5' UTR secondary structure is generally cited in the literature as around -10 to -30 kcal/mol — stable enough to protect the cap from decapping enzymes but not so stable that translation is impaired. We use RNA structure prediction (RNAfold from the ViennaRNA package, cross-validated with RNAstructure) to evaluate candidate 5' UTR sequences for this property.
The 3' UTR is where most mRNA half-life control elements live. AU-rich elements (AREs) in the 3' UTR are recognized by destabilizing RNA-binding proteins including TTP (tristetraprolin) and the KSRP proteins, and including them is one of the fastest ways to shorten mRNA half-life. Conversely, the 3' UTR sequences derived from highly stable mRNAs — beta-globin is the classic reference — provide protection from deadenylation and 3'-to-5' exonucleolytic decay.
For longevity applications, we want mRNA half-life in vivo to be in the range of days to weeks, not hours. Modified nucleotides — N1-methylpseudouridine (m1Ψ) replacing uridine, as in current approved mRNA medicines — extend half-life partly through reduced innate immune activation (discussed below) and partly through reduced susceptibility to endonucleolytic cleavage. But UTR optimization on top of nucleotide modification can extend half-life further, and the two design choices interact: the optimal UTR sequence may differ between unmodified and fully modified mRNA.
The Poly-A Tail
Poly-A tail length directly affects mRNA stability and translational efficiency through its interaction with PABP (poly-A binding protein). Modern mRNA design typically uses encoded poly-A tails of 100–160 adenosines. However, the optimal poly-A tail length is context-dependent — longer tails may improve stability but can also trigger surveillance by the mRNA quality control machinery if the poly-A stretch is not completely processed by the exosome. There is also evidence that poly-A tail length affects LNP encapsulation efficiency, with longer tails potentially creating secondary structure interactions that affect packaging.
Innate Immune Activation: The CpG and Uridine Problem
Innate immune sensing of exogenous RNA is one of the most important constraints on mRNA therapeutic design. The primary sensors relevant to LNP-delivered mRNA are the endosomal toll-like receptors TLR3, TLR7, and TLR8, which recognize double-stranded RNA, single-stranded RNA with specific sequence features, and GU-rich RNA respectively, plus the cytoplasmic RNA sensors RIG-I and MDA5.
CpG dinucleotides in DNA are a classic PAMP (pathogen-associated molecular pattern) recognized by TLR9, but CpG-adjacent sequences in mRNA can also affect TLR7/8 recognition in ways that are not fully resolved in the literature. We apply a CpG suppression step to our candidate sequences to reduce potential TLR stimulation, though this can create tension with codon optimization since CG-containing codons are often preferred for certain amino acids.
Uridine content is a more direct concern. TLR7 and TLR8 are activated by uridine-rich single-stranded RNA. Substituting uridines with modified nucleotides — m1Ψ, which is standard in current approved mRNA medicines and the approach taken in Moderna and BioNTech COVID vaccines — abolishes most TLR7/8 activation and substantially reduces the innate immune response. This is not simply an immunological side effect concern: innate immune activation at the dose site or systemically shortens mRNA half-life through activation of RNA-degrading pathways and also causes cytokine secretion that can have direct toxic effects at higher doses.
We are not claiming that m1Ψ modification completely eliminates immunogenicity. At sufficiently high doses, even fully m1Ψ-modified mRNA in LNPs can stimulate innate immune responses via the LNP lipid components themselves (some ionizable lipids have intrinsic inflammatory activity) or via double-stranded RNA contaminants from in vitro transcription. HPLC purification of IVT mRNA to remove dsRNA contaminants is essentially a prerequisite for therapeutic-grade material, and this is a step that is often underemphasized in academic benchmarks.
Off-Target Protein Interaction Modeling
For knockdown therapeutics, off-target analysis focuses on unintended silencing of non-target genes with sequence complementarity to the antisense strand. For mRNA upregulation, the off-target concern is different: the expressed protein itself may interact with non-target binding partners, or the increased protein levels may amplify interactions that are minor at endogenous expression.
We use a combination of STRING network analysis and predicted protein-protein interaction docking to screen our candidate protein sequences for potential off-target interactions. For secreted proteins (soluble Klotho, for instance), we pay particular attention to interactions with abundant serum proteins since the expressed protein will be circulating. For transcription factors (FOXO3), we map the predicted cistrome — the set of genomic binding sites — at elevated expression levels and ask whether genes with high-affinity FOXO3 binding sites include any with potentially adverse biology in the relevant tissue context.
This modeling is genuinely predictive rather than definitive. We can flag high-probability concerns, but in vitro transcriptomics after mRNA transfection in relevant cell types is necessary to actually characterize the expression changes induced by protein upregulation. We use HEK293T cells for initial screening (fast, easy to transfect, well-characterized baseline transcriptome) and differentiated cell lines or iPSC-derived cell types for more specific evaluation. The gap between HEK293T and a primary human neuron in terms of relevant biology is large enough that we do not treat HEK293T data as definitive for CNS candidates.
The Output: A Ranked Candidate List, Not a Single Sequence
Our AI sequence evaluation pipeline does not output a single optimized mRNA sequence. It outputs a ranked list of candidates with predicted scores across the dimensions described above: ribosome density profile, UTR stability, CpG content, immune stimulation risk, off-target interaction probability. For most targets, we advance 3–5 candidate sequences to wet-lab evaluation rather than a single candidate, because the predicted rankings are not precise enough to distinguish candidates that are close in score.
The wet-lab evaluation protocol we use for candidate ranking is protein expression by Western blot and ELISA in at least two cell lines, mRNA half-life by RT-qPCR at 4h, 24h, 48h, and 72h post-transfection, and cytokine secretion (IFN-alpha, TNF-alpha, IL-6) by ELISA as an immunogenicity proxy. This is a relatively minimal set of assays compared to what a full candidate characterization requires, but it is what we can run with our current resources before making decisions about which candidates to advance to more resource-intensive work.
We want to be explicit about what the computational pipeline adds versus what it cannot replace. It reduces the number of candidates we need to test experimentally by eliminating obvious failures — sequences with high innate immune activation risk, poor predicted stability, or structural features inconsistent with correct protein folding. It cannot replace wet-lab data for the candidates that survive that initial computational filter. The biology at the cell and organism level is too complex for computational prediction alone to determine which mRNA candidate will be safe and effective in a disease-relevant model. The pipeline accelerates the experimental cycle; it does not substitute for it.
