Lab Newsletter — August 9, 2026: The Grammar of the Genome

A cell’s genome is the one molecule that is both the blueprint and, increasingly, something a model can generate. For most of the genomics era we could read genes — the roughly 2% of human DNA that codes for protein — while the other 98%, the regulatory sequence that decides when, where and how much a gene is used, stayed mostly dark. And though we learned to read, write, and edit DNA, we never learned to compose it: we could cut and paste nature’s sentences without knowing the grammar to write our own. In 2026, two landmark Nature papers moved both walls in the same season — one taught a model to read the whole regulatory genome, the other taught one to write genomes across all of life. The genome, at last, has a grammar you can model.

📖 Read the other 98%

The reading breakthrough is AlphaGenome (Hassabis, Kohli et al., Google DeepMind, Nature, 28 January 2026). It predicts regulatory features directly from DNA sequence — variant effects, RNA splice sites, chromatin accessibility, gene expression — and it does so at a resolution the field had been forced to trade away. Earlier base-resolution models like SpliceAI saw only short windows (≤10 kb) and missed distant regulation; long-context models like Enformer and Borzoi reached hundreds of kilobases but only at a coarse 32–128 bp grain. AlphaGenome takes in 1 megabase of sequence and predicts 5,930 human genome tracks across 11 output types at single-base resolution — both at once — via a U-Net-style encoder→transformer→decoder. Its most elegant trick is a 2D representation for splicing that, in the authors’ framing, “parallels AlphaFold’s pairwise amino-acid representations”: just as AlphaFold’s pair track captures which residues sit close in a folded protein, AlphaGenome’s captures which genomic positions interact, so it can weigh a single splice-site nucleotide against a regulatory element half a megabase away. On the numbers, it “matched or exceeded the strongest available external model on 24 of 26 variant-effect evaluations.” As the Francis Crick Institute’s Robert Goldstone put it, “this level of resolution, particularly for non-coding DNA, is a breakthrough that moves the technology from theoretical interest to practical utility.” Why it matters for the lab: the noncoding genome is where most disease-associated variants live and where interpretation has been hardest. A model that reads the regulatory 98% at base resolution is the sequence-to-function layer a virtual cell needs to connect genotype to what a cell actually does.

✍️ Write across all of life

If AlphaGenome reads, Evo 2 (Brixi, Durrant, Ku et al.; co-led by Patrick Hsu and Brian Hie; Arc Institute + NVIDIA, Nature, 2026) writes. It’s a genome language model at frontier scale — 7B and 40B parameters trained on over 9.3 trillion nucleotides from more than 128,000 whole genomes spanning all three domains of life, reading up to 1 million nucleotides of context with a StripedHyena 2 architecture — “the largest fully open biological AI model to date.” Trained only to predict sequence, it turns out to have learned biology: it classifies BRCA1 mutations as benign or pathogenic at over 90% accuracy with no task-specific training, and — the part that still feels like science fiction — it can design new genomes as long as those of simple bacteria. Crucially, this isn’t only in silico: of Evo 2’s designed sequences, “16 of 285 tested designs successfully propagated and inhibited growth of the appropriate bacterial strains” — composed DNA that actually worked in living cells. Patrick Hsu’s summary is the whole thesis in one line: the models “have enabled machines to read, write, and think in the language of nucleotides.” Why it matters for the lab: generative genome design plus zero-shot variant interpretation is the engine of automated discovery — and Arc frames Evo explicitly as one layer of a stack that runs up to virtual cell models. It’s the same bet as our Human Cell Simulator, and the fact that Evo ships fully open — weights, code, data — is the BioImage Model Zoo / BioEngine ethos carried from images to genomes.

🧭 The honest grammar — and the safeguard

The excitement is earned, but the frontier is not finished, and this is where the story stays useful rather than triumphant. A peer-reviewed benchmarking study (Nature Communications, 2025) finds a revealing split: on tissue-specific QTL tasks the specialized models still win, because foundation models tend to learn a general, context-free sense of what makes a sequence “broken” while specialists capture the cell-type-specific regulatory grammar that decides whether a variant is functional here, in this tissue. AlphaGenome’s own authors flag the same gap — cell-type specificity and rare variants remain unreliable — which is why KCL’s Xianghua Li cautions that “for important medical tasks, current AI models are still not reliable enough for patient care.” That’s not a knock; it’s the prove-it discipline we keep returning to — hold these models to held-out, task-relevant tests, exactly as the genomic-FM evaluation reckoning demanded. And a model that can compose biology raises a second obligation the field is, encouragingly, taking seriously: Evo 2’s team excluded human pathogens from training and red-teamed the model so it “would not return productive answers” about them. That is Design, Build, Test — and Safeguard in its sharpest form: when design becomes generative, alignment has to be built in before capability outruns it.

Read together, the two models bracket a single shift. The genome used to be a text we could spell out letter by letter but not truly read past the genes, and copy but not write. Now one model reads the regulatory 98% at single-base resolution and another writes sequences that live and function in real cells — and the sober benchmarks and biosafety guards arriving alongside them are what make the capability trustworthy rather than merely impressive. The blueprint has become a language: something we can finally read in full, and are learning, carefully, to compose.

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.