Lab Newsletter — September 12, 2026: Writing the Messenger

The COVID-19 vaccines settled an argument that had simmered for a decade: mRNA is a medicine you can manufacture by writing a sequence. But the protein a therapeutic mRNA encodes is only part of the design. Wrapped around that coding sequence — and hidden inside its choice of synonymous codons — is a second layer of instructions that decides how much protein gets made, how long the molecule survives in a vial and a cell, and how loudly the immune system reacts. That layer is not written by nature for our purposes, and there are astronomically many ways to write it. Today’s digest is about the machine-learning tools now writing it for us — designing the message itself.

🧬 The founding move: learn the 5′UTR, then design it

The 5′ untranslated region — the stretch of RNA a ribosome scans before it starts translating — is a dial for protein output, and Sample et al. (Nature Biotechnology, 2019) learned to turn it. They “combine polysome profiling of a library of 280,000 randomized 5′ untranslated regions (UTRs) with deep learning to build a predictive model that relates human 5′ UTR sequence to translation.” A model that predicts is useful; a model you can run backwards is transformative. So they “use this same model, coupled with genetic algorithms, to engineer new 5′ UTRs that accurately direct specified levels of ribosome loading” — dialing protein output up or down on demand. Two details make it our kind of result. First, it points straight at medicine: “the same approach can be extended to chemically modified RNA, an important feature for applications in mRNA therapeutics.” Second, it reads disease: evaluating UTRs carrying human variants, they “identify 45 single-nucleotide variants (SNVs) associated with human diseases that substantially change ribosome loading.” One massively parallel assay, and the UTR becomes both designable and interpretable.

📏 Generalize the model, read the clinic

That first model was trained on fixed-length library sequences — but real UTRs come in every length, and the interesting ones are often the pathogenic ones. Karollus, Avsec & Gagneur (PLoS Computational Biology, 2021) closed both gaps. Starting again from the fact that “a model was trained on a massively parallel reporter assay to predict mean ribosome load (MRL),” they “introduced frame pooling, a novel neural network operation that enabled the development of an MRL prediction model for 5′UTRs of any length.” The reward for generalizing is clinical reach: “variant interpretation is demonstrated on a 5′UTR variant of the gene HBB associated with beta-thalassemia,” and, “released open source,” the model “could help pinpoint pathogenic genetic variants.” The design tool and the diagnostic tool turn out to be the same network.

🧊 Make it last: design the structure for stability

Getting a lot of protein is moot if the molecule falls apart first — and mRNA’s Achilles’ heel is chemistry, not biology. Wayment-Steele et al. (Nucleic Acids Research, 2021) name the problem exactly: “RNA hydrolysis presents problems in manufacturing, long-term storage, world-wide delivery and in vivo stability of messenger RNA (mRNA)-based vaccines.” Their fix is to design the shape: “a general strategy to stabilize mRNA is to redesign RNAs to form double-stranded regions, which are protected from in-line cleavage,” and they supply “a model that links the average unpaired probability of an mRNA, or AUP, to its overall hydrolysis rate.” Minimize the AUP — tuck the vulnerable bases into paired structure — and the molecule survives longer. They demonstrate it on the vaccine target itself, describing “two designs of mRNAs coding for a model protein and the SARS-CoV-2 spike protein with low AUP, which we term ‘superfolder’ mRNAs,” and conclude that “increases in in vitro mRNA half-life by at least two-fold appear immediately achievable.” Design isn’t only about expression level — it’s about the medicine reaching the patient.

⚡ The flagship: codons and structure, co-optimized at scale

Expression and stability pull in different directions, and the number of ways to encode one protein is beyond astronomical. Zhang et al. (Nature, 2023) put a number on it: “the mRNA sequence design problem is exceptionally difficult because of the exponentially large search space—for example, there are around 2.4 × 10^632 candidate mRNA sequences for the SARS-CoV-2 spike protein.” Their breakthrough is to borrow a tool from an unexpected field, “reformulating this task as a lattice parsing problem, in which the lattice represents the entire design space and parsing recursively finds the optimal design” — the same dynamic-programming idea that parses human language, now parsing genetic codes. The result, LinearDesign, “finds an optimal mRNA design for the spike protein in just 11 minutes, and can concurrently optimize stability and codon usage.” And the wet lab delivers the exclamation point: it “substantially improves mRNA half-life and protein expression, and profoundly increases antibody titre by up to 128 times in mice compared to the codon-optimization benchmark.” A hard algorithm, and a 128× vaccine at the end of it.

🅰️ A foundation model for the untranslated region

The rest of biology has been swept by pretrained language models; the messenger is no exception. Chu et al. (Nature Machine Intelligence, 2024) “introduced a language model for 5′ UTR, which we refer to as the UTR-LM.” In the now-familiar recipe, “the UTR-LM was pretrained on endogenous 5′ UTRs from multiple species and was further augmented with supervised information including secondary structure and minimum free energy.” It beats the specialists at their own tasks — “outperformed the best-known benchmark by up to 5% for predicting the mean ribosome loading, and by up to 8% for predicting the translation efficiency and the mRNA expression level” — but the payoff, again, is design: they “designed a library of 211 novel 5′ UTRs,” and “wet-lab validation showed that our top designs achieved a 32.5% increase in protein production level relative to well-established 5′ UTRs optimized for therapeutics.” Beating a UTR that was already optimized for medicine, by nearly a third. (A nice thread to yesterday’s knowledge-graph digest: Kexin Huang, an author on TxGNN, is a coauthor here — the people building biology’s reasoning layer are building its design layer too.)

🔀 RNA that computes: programmable switches

Everything so far treats mRNA as a template — a thing to be translated well and to last. But RNA can also decide. Angenent-Mari et al. (Nature Communications, 2020) applied deep learning to toehold switches, RNA elements that turn translation on only in the presence of a specific trigger sequence — the basis of RNA diagnostics and smart therapeutics. They “investigate Deep Neural Networks (DNN) to predict toehold switch function as a canonical riboswitch model in synthetic biology,” training on a purpose-built dataset: they “synthesize and characterize in vivo a dataset of 91,534 toehold switches spanning 23 viral genomes and 906 human transcription factors.” The gap over physics-based models is dramatic — “DNNs trained on nucleotide sequences outperform (R2 = 0.43–0.70) previous state-of-the-art thermodynamic and kinetic models (R2 = 0.04–0.15)” — and it stays interpretable, with a visualization method (VIS4Map) that “successfully identify[s] sequence and structural elements that correlate with toehold switch function.” Designing the message now includes designing its logic: not just how much protein, but when and whether.

🧫 Why it’s our kind of problem

Look across the six and one pattern repeats: a massively parallel assay feeds a model, and the model is then run in reverse to design new sequences the wet lab checks. Sample’s genetic algorithms, Karollus’s any-length predictor, LinearDesign’s 11-minute optimum, UTR-LM’s 211-UTR library — every one closes the predict → design → measure loop. That is precisely the build-and-measure ethos behind a self-driving lab: a model is only as good as the experiment that can falsify it, and the best designs come from tools that were trained on data and validated at the bench. The field is open and benchmarked in the way the lab likes — Karollus and UTR-LM are open source, and everyone is scored on shared measures like mean ribosome load and translation efficiency, the same publish-the-model-and-the-data spirit behind the BioImage Model Zoo and BioEngine. And it’s a concrete rung toward the virtual cell: a faithful model of a cell has to capture not only what a molecule is but how efficiently it is made and how long it survives — exactly the quantities these UTR, stability and codon models predict. mRNA design is foundation-models-for-biology meeting sequence design, with a delivered medicine as proof it matters.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits. Anchors were verified via NCBI E-utilities today, as Europe PMC was temporarily down.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.