Lab Newsletter — July 30, 2026: The Messenger, Read and Written

We spend a lot of these digests at the two ends of the central dogma — the genome that a cell could build from, and the proteins and structures it actually assembles. The middle layer, RNA, has been quietly under-modeled: harder to crystallize than protein, more structurally alive than DNA, and for years without a foundation model of its own. That gap is closing fast. In the last year RNA got both a way to be read — structure and function predicted straight from sequence — and a way to be written — generative models that design functional RNAs from scratch. For a lab building toward a virtual cell, that’s the messenger layer coming online.

📖 Reading RNA: a language model, then a structure

The base layer is RNA-FM, a BERT-style RNA language model (12 transformer layers) pretrained by masked-language-modeling on ~23.7 million non-coding RNA sequences — no alignments, no evolutionary input, just sequence. Its per-nucleotide embeddings transfer to structure and function tasks, and its successor RhoFold+ (Shen et al., Nature Methods 2024) couples those embeddings to a geometry-aware network to predict RNA 3D tertiary structure end-to-end from a single sequence — reaching state-of-the-art on RNA-Puzzles and CASP-style targets in seconds. It is, roughly, the AlphaFold moment for RNA: fold prediction that used to need painstaking modeling now falls out of a pretrained model. Both are open (MIT). Why it matters for the lab: an open, benchmarked model that turns sequence into structure is precisely the kind of tool our BioImage Model Zoo culture is built to share and stress-test — now for the molecule that carries the cell’s instructions.

✍️ Writing RNA: from understanding to de novo design

Reading is half the story; the newer leap is design. RNAGenesis (~1B parameters) unifies both in one model — a bidirectional encoder for understanding, a query-based latent compression feeding a diffusion-guided decoder for generation. It ranks first on 11 of 13 tasks of the BEACON RNA benchmark and, on design tasks, outperforms prior RNA models including RNA-FM and the genomic model Evo2. Crucially it doesn’t stop at benchmarks: the authors introduce RNATx-Bench, a therapeutics-oriented benchmark aggregating over 100,000 experimentally validated RNAs across ASOs, siRNAs, shRNAs, circRNAs, aptamers and UTR variants — and they took a designed output to the bench, generating CRISPR guide RNAs validated in HEK293T cells (targeting EGFP and B2M) with editing efficiency equal to or better than wild-type guides. Why it matters for the lab: this is design-build-test on the messenger layer — a generative model that proposes a functional guide RNA and gets it working in cells. That closed loop, kept open and benchmarked, is the shape of the tools we build.

🌐 The frontier: full-length transcripts, in one context

The freshest entry pushes on the thing that has held RNA models back — context length. EVA (“Evolutionary Versatile Architect,” bioRxiv, Mar 2026; open under Apache-2.0, preprint, not yet peer-reviewed) is a ~1.4B-parameter Mixture-of-Experts generator with an 8,192-token context window — versus roughly 1,024 in earlier RNA models — so it can model full-length transcripts without truncation. It’s trained on OpenRNA v1, a curated atlas of 114 million full-length RNA sequences across structural, regulatory, coding and viral RNAs, and it conditions explicitly on RNA type and taxonomic lineage while jointly learning generation and infilling. The authors report state-of-the-art across mutation-effect prediction, conditional generation and functional design — and, like the others, released data, weights and code. (Alongside it, models like AIDO.RNA and diffusion inverse-folders like RiboDiffusion round out a suddenly crowded field.) Why it matters for the lab: long-context, full-length modeling is what lets RNA join a whole-cell picture rather than a fragment-at-a-time one — the transcript layer between genome and protein, at last legible end to end.

The pattern across all three is the one we keep betting on: open weights, shared benchmarks, and design that gets checked at the bench. RNA was the layer of the virtual cell we could least read; a year later it’s one we can read and write. The messenger is starting to speak our language back.

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.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.