Lab Newsletter — August 19, 2026: The Other 98 Percent

Only a sliver of the human genome codes for protein — a percent or two. Everything else, the other ~98%, was once dismissed as junk and is now understood as regulation: enhancers, promoters, splice signals, the switches that decide when and where a gene turns on. It matters enormously that we learn to read that layer, because roughly 90% of the trait- and disease-associated variants found by genome-wide association studies fall in exactly this non-coding sequence (Schipper & Posthuma, 2022). A mutation there doesn’t break a protein — it retunes a switch. The question of the decade: can a model read a stretch of DNA and tell you what it does, and what a single change would break? This is a different track from the generative genome language models we covered on Aug 9 — not writing sequence, but predicting the regulatory function it encodes.

📖 Reading what DNA does

The breakthrough was learning regulatory activity straight from sequence. Enformer (Avsec et al., Nature Methods, 2021; senior author David Kelley) was the turning point — a transformer that reads an input window of about 200,000 bases, integrating regulatory interactions from up to 100 kb away, and predicts gene expression and chromatin state across cell types, giving markedly more accurate variant-effect predictions than the convolutional models before it (benchmarked against natural variants and MPRA saturation mutagenesis). Borzoi (Linder et al., Nature Genetics, 2025; also Kelley) went further into the transcript itself, predicting cell- and tissue-specific RNA-seq coverage from sequence at base-pair resolution — unifying transcription, splicing and polyadenylation in one model and scoring how variants perturb each. Two models, one idea: the genome’s regulatory grammar is legible to a network that looks at enough sequence at once.

🔎 Where the variants hide

Reading regulation is not an academic exercise — it’s how you interpret a mutation. The coding side has come a long way: AlphaMissense (Cheng et al., Science, 2023; senior author Žiga Avsec) — an adaptation of AlphaFold, fine-tuned on population variant frequencies — classifies 89% of all possible human missense variants as likely benign or likely pathogenic, a remarkable coverage of the protein-changing mutations. But that’s the easy 1–2%. The non-coding variant — the regulatory one, sitting in an enhancer 40 kb from the gene it controls — is the harder, less-solved problem, and it’s where most of the disease signal actually lives. Which is exactly why the sequence→function models matter: to score a non-coding variant, you first have to predict what the sequence around it does.

🧭 AlphaGenome, and the honest frontier

The unifying leap arrived this year. AlphaGenome (Avsec et al., Nature, published January 2026; senior author Pushmeet Kohli; Google DeepMind) reads up to 1 Mb of DNA at single base-pair resolution — collapsing the old trade-off between long-range context and fine resolution — and predicts a broad panel of regulatory modalities in one shot: expression, splicing (down to sites, usage and junctions), chromatin accessibility, transcription-factor binding, histone marks, and 3D contact maps (DeepMind counts 11 modalities). It outperformed the best external models on 22 of 24 single-sequence prediction tasks and matched or exceeded top models on 24 of 26 variant-effect evaluations, and it’s out as a non-commercial research API preview — a through-line completed by the same researcher, Žiga Avsec, who first-authored Enformer and now AlphaGenome.

And yet the frontier is honest about itself. Scale has not closed the gap that matters most. Karollus et al. (Genome Biology, 2023, Gagneur lab) showed that state-of-the-art models capture the determinants of expression in promoters but largely ignore distal enhancers — their effective long-range reach is far shorter than their nominal window. A companion Nature Genetics study (2023, Ioannidis lab) found current models predict personal-genome expression variation poorly, sometimes even getting a variant’s direction of effect wrong. The root worry is correlative training: models learn from evolutionary differences between genes, so generalizing causally to a new mutation is never guaranteed — the prediction is strongest exactly where the biology is easiest (promoters), and weakest where we need it most (distal, personal, disease variants).

Here’s why it lands for us. A virtual cell — the horizon this digest keeps circling — needs a genotype→cell-state module, and reading the regulatory genome is that module: a foundation-model-for-biology problem in a new data type. These are large models that must be open, standard-format and callable to be useful, the same BioEngine / Model Zoo ethos, now for genomics. And they carry the same prove-it discipline the digest returns to again and again: a predicted variant effect is a hypothesis, confirmed only when an MPRA or a functional assay agrees. The models can already read the genome’s easy sentences. Teaching them the hard ones — the distal switch, the personal mutation, the one that matters in a patient — is the work that’s left, and it’s the kind the lab is built to do.

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.