AlphaGenome predicts molecular effects from 1Mb
- Google DeepMind’s AlphaGenome is now the clearest example of a new genomics AI class: models that read 1 megabase of DNA and predict regulatory effects. - The key trick is doing two hard things at once — long-range context and near base-level outputs across expression, splicing, chromatin, binding, and contacts. - That matters because most disease variants sit outside genes, where biology is harder to read and lab testing is painfully slow.
DNA is not just a parts list. It is also control logic — when genes turn on, where RNA gets spliced, which proteins bind, and which distant stretches of genome physically interact. That control layer is where a huge share of disease risk seems to live, but it has been brutally hard to read. AlphaGenome is a new attempt to change that: a model from Google DeepMind that takes up to 1 million DNA letters at once and predicts thousands of molecular effects from the sequence alone. ### Why is 1 megabase a big deal? Because gene regulation is often long-range. A mutation can sit far away from the gene it disrupts — sometimes tens or hundreds of thousands of bases away — inside an enhancer or other regulatory element. Older sequence models usually had to choose: either look at a long stretch of DNA with coarse outputs, or make sharp local predictions with a much shorter window. AlphaGenome’s pitch is that it does both in one system. (deepmind.google) ### What does the model actually predict? Not just “good” or “bad” variants. It predicts many molecular readouts tied to regulation — gene expression, transcription initiation, chromatin accessibility, histone marks, transcription factor binding, chromatin contact maps, splice site usage, and splice junction behavior. Basically, it tries to forecast the intermediate biology between a DNA change and a disease phenotype. That is more useful than a single risk score because researchers can inspect the likely mechanism. (deepmind.google) ### Why has this been hard before? The genome has a scale problem. Nearby motifs matter, but so do distant interactions. A model has to notice tiny sequence changes at single-base resolution while also carrying information across a region a million letters long. That is a nasty engineering tradeoff — like trying to read one typo while still understanding the whole chapter. DeepMind says AlphaGenome combines convolutional layers, transformers, and modality-specific prediction heads to hold onto both kinds of information. (deepmind.google) ### So what changed with AlphaGenome? The change is unification. Instead of separate tools for splicing, expression, chromatin, or contact prediction, AlphaGenome packages many of those tasks into one model and then uses the difference between reference and mutated sequence predictions to score variant effects. In the Nature paper and DeepMind release, the model is described as outperforming prior approaches across a range of regulatory variant benchmarks. (deepmind.google) ### Does that mean it can diagnose disease now? No — and this is the catch. These models predict molecular consequences, not clinical outcomes. A variant can alter splicing or chromatin accessibility without causing disease in a person, and lab context still matters — cell type, development stage, environment, and genetic background all shape the final effect. So AlphaGenome is best thought of as a prioritization engine, not a doctor. (nature.com) ### Why are people linking this to Illumina and NVIDIA? Because the bottleneck is shifting. Sequencing DNA is no longer the only hard part; interpreting the flood of genomic and multiomic data is the bigger challenge. Illumina and NVIDIA said in January 2025 that they were combining sequencing, analytics, and AI tools to build biology foundation-model workflows for drug discovery and clinical research. AlphaGenome fits that same direction of travel — sequence generation on one side, large predictive models on the other. (deepmind.google) ### What could this unlock if it holds up? The obvious win is variant triage. Researchers could move faster from a suspicious noncoding mutation to a mechanistic hypothesis worth testing in the lab. Drug teams could also use these models to narrow target lists, design perturbation experiments, and make sense of regulatory biology that used to look like static. But the real value will depend on external validation — especially whether predictions stay reliable across rare cell states and messy disease contexts. (investor.illumina.com) ### Bottom line? AlphaGenome matters because it attacks the hardest part of genomics — turning raw sequence into mechanism. If these 1 Mb, base-resolution models keep improving, the center of gravity in genomics may move from reading DNA to simulating what DNA does. (deepmind.google) (nature.com)
