Protein-coding genes are a sliver of the genome; the rest is regulatory dark matter — and it’s where roughly 90% of disease-associated variants hide. A supervised sequence→function track learned to read it: Enformer predicts expression from ~200 kb of context, Borzoi predicts RNA-seq coverage across transcription, splicing and polyadenylation, and AlphaMissense settled the coding side (89% of missense variants classified). AlphaGenome (Nature, 2026) unifies the regulatory side — one model over 1 Mb of DNA at single-base resolution, beating the best external models on 22 of 24 sequence tasks and 24 of 26 variant-effect tests. But the honest frontier is sharp: these models capture promoters and largely miss distal enhancers, and predict personal-genome variation poorly. A prove-it story at genome scale — the module a virtual cell will need, and one that still has to earn every prediction in the lab.
