Lab Newsletter — August 20, 2026: The Harder Fold

Two days ago this digest celebrated the proteins evolution never wrote — generative design so good the 2024 Nobel sealed it. RNA is the humbling counterpoint. It is the molecule in the middle of the central dogma, and predicting its 3D shape from sequence is harder than for proteins: RNA is floppier, its folds hinge on subtle non-canonical base pairs, and there are far fewer experimental structures to learn from. So the RNA story splits cleanly in two — a structure problem that is honestly not yet solved, and a design capability that is already changing medicine. Both are worth the lab’s attention.

🧬 The harder fold

For proteins, AlphaFold turned structure prediction from a grand challenge into a solved-enough tool. AlphaFold3 (Abramson et al., Nature, 2024; senior authors Demis Hassabis and John Jumper) pushed the same machinery beyond proteins to nucleic-acid complexes, ligands and modified residues, reporting “much higher accuracy for protein–nucleic-acid interactions” than specialized predictors. RNA-specific efforts went deeper: RhoFold+ (Shen et al., Nature Methods, 2024; senior author Yu Li) is a language-model-based predictor of single-chain RNA structure, its RNA language model pretrained on ~23.7 million sequences, reported to surpass existing methods — “including human expert groups” — on the RNA-Puzzles and CASP15 benchmarks and to generalize across RNA families.

And yet the field is refreshingly honest about where it stands. The community assessment of CASP15 (Das et al., Proteins, 2023; senior author Eric Westhof) concluded flatly that “the prediction of RNA three-dimensional structures remains an unsolved problem” — and, strikingly, that at that bake-off “predictions from deep learning approaches were significantly worse than these top ranked groups, which did not use deep learning.” This is the rare frontier where AI has not yet won, and the assessors say so out loud. RNA folding is unfinished business.

💉 The messenger we can already write

Design, remarkably, outruns prediction — and that’s where the human impact is immediate. You don’t need a perfect structure to build a better mRNA. LinearDesign (Zhang et al., Nature, 2023; project led by Liang Huang, a Baidu Research–Oregon State–StemiRNA–Rochester collaboration) jointly optimizes an mRNA’s codon usage and structural stability at once — out of some 2.4 × 10⁶³² possible sequences for the SARS-CoV-2 spike protein, it finds an optimal design in just 11 minutes, and its designs raised antibody titres up to 128× in mice versus the standard codon-optimization benchmark (for both COVID-19 and varicella-zoster vaccines). An algorithm, quietly, at the heart of a pandemic-era medicine.

The read-and-write toolkit is filling in around it. Sample & Seelig (Nature Biotechnology, 2019) trained deep learning on 280,000 randomized 5′ UTRs to predict — and then design — how efficiently an mRNA is translated, and flagged 45 disease variants that shift ribosome loading. RiNALMo (Penić et al., Nature Communications, 2025; senior author Mile Šikić) is a 650-million-parameter RNA language model pretrained on 36 million non-coding RNA sequences that generalizes to unseen RNA families — an open, general-purpose RNA foundation model of the kind the protein and genome worlds already take for granted.

🧭 Why it matters for the lab

RNA sits at the center of the cell — transcription, splicing, translation, regulation, structure — so any virtual cell or foundation-model-for-biology program needs an RNA module, not just proteins (Aug 18) and DNA (Aug 19). And RNA hands us a pointed lesson: the reason its structure prediction lags is data scarcity — far fewer experimental RNA structures to train on — which is the strongest argument yet for open, shared, standard-format data and callable models, the BioEngine and Model Zoo ethos, applied to a field that badly needs it. Best of all, the prove-it discipline is baked right into RNA’s own scorecard: CASP15 tells us plainly that a predicted RNA fold is still a hypothesis, and even a beautifully designed mRNA is one until it’s expressed and assayed. Design already saves lives; prediction still has to earn its place. That gap — honest, measurable, open — is exactly the kind of problem the lab is built to work on.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (X/Twitter sweep was skipped today — our news API is out of credits; a Grok-based replacement is wired and awaiting credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.