AI Tools · Oct 7, 2026
DeepMind AlphaGenome Explained: Predicting What Single-Letter DNA Changes Do
By ChannelGlance Editorial · Approx. 5 min read

AlphaGenome is a machine-learning model from Google DeepMind that predicts what happens when a single letter changes in the roughly three-billion-character genome — specifically, how that one-letter change affects molecular phenotypes such as RNA splicing and gene expression. In an interview with Two Minute Papers, DeepMind’s Pushmeet Kohli described the model as the first to combine base-pair resolution with a one-million-base-pair context window, and DeepMind reports that it matched or beat the strongest comparison model on 25 of 26 variant-effect benchmarks. The model is free for non-commercial research, its weights are available for fine-tuning, and it ships alongside the AlphaGenome Atlas, a petabyte-scale precomputed catalog covering all roughly nine billion possible single-letter variants.
What AlphaGenome Predicts, in Plain Terms
The human genome is a code written in about three billion characters. Swap a single character — an A becomes a G at one position — and the molecular consequences can be negligible, significant, or simply unknown. Answering that question has traditionally fallen to lab experiments. AlphaGenome is built to forecast the answer computationally, and its outputs are concrete molecular phenotypes: whether a gene becomes more or less active, and whether the RNA transcribed from that gene is spliced — cut and stitched together — differently because of the change.
Those two readouts matter because they sit upstream of much of what genetics researchers study. A model that precomputes them turns variant interpretation into a triage problem: instead of testing blindly across billions of possibilities, a lab can start from a ranked shortlist of variants whose predicted effects look consequential.
Why Base-Pair Resolution and a One-Million-Base Window Matter
The headline technical claim, as Kohli presented it, is that AlphaGenome is the first model to combine base-pair resolution with a context window spanning one million base pairs. Both halves of that sentence do real work.
Base-pair resolution means the model can localize effects to individual positions in the sequence rather than averaging across coarse blocks. Its predecessor, Enformer, worked at coarser resolution over a smaller window, which meant it could miss effects driven by long-range interactions — cases where a regulatory element physically folds into contact with a gene hundreds of thousands of bases away. DNA does not behave like a straight line; it packs into three-dimensional structures, so a variant far from a gene can still change that gene’s behavior. The one-million-base-pair window lets AlphaGenome see those distant relationships, and single-base resolution lets it pinpoint where the relevant change sits.
What the Reported Benchmarks Show — and What They Do Not
DeepMind reports that AlphaGenome matched or beat the strongest comparison model on 25 of 26 variant-effect benchmarks. The company also reports that predicted effect sizes for gene-activity mutations correlate with observed effect sizes at roughly rho 0.5, meaning the model often gets the magnitude of an effect approximately right rather than merely its direction.
Two caveats keep those numbers honest. First, these are DeepMind-reported results; the interview did not include independent verification. Second, a correlation of about 0.5 supports prioritization, not diagnosis. It is strong enough to help decide which variants deserve experimental attention, and it is not a final answer about any individual variant.
The AlphaGenome Atlas: Nine Billion Variants, Precomputed
Running a frontier-scale model billions of times is impractical for most labs, so DeepMind precomputed the answers. The AlphaGenome Atlas catalogs the predicted effects of all roughly nine billion possible single-letter variants in a petabyte-scale dataset — reported to be about 30 times the size of the AlphaFold protein-structure database, a resource DeepMind notes is used by more than four million researchers across 190 countries. Each entry carries an AVI impact score that helps triage which variants matter most.
For a research team, a practical workflow follows directly from the design:
- Query the Atlas for a variant or region of interest instead of running the full model from scratch.
- Use the AVI impact score to rank variants and cut the list down to those with the largest predicted effects.
- Run the full AlphaGenome model — fine-tuned on your own data if needed — only on the variants that survive triage.
Free for Non-Commercial Research, With Fine-Tunable Weights
AlphaGenome is free for non-commercial academic use, and the model weights are available for fine-tuning. That matters for AI researchers as much as for biologists: instead of a fixed oracle, labs get an adaptable model they can train further on their own data and tasks.
The training recipe points the same direction. AlphaGenome was trained on human and mouse genomes, which Kohli framed as a way to push the model toward general principles of genome function rather than memorization of human-specific patterns. If a model has to explain two species at once, it has stronger incentives to learn shared biology than to parrot one training set.
Where AlphaGenome Stops: Molecular Effects Are Not Disease Risk
The limits Kohli acknowledged deserve as much weight as the benchmarks. A prediction about splicing or gene expression is a molecular-level result. Connecting that molecular effect to whether an organism — a person — actually faces altered disease susceptibility is a separate research problem, and Kohli was clear that this link still requires further work.
An interpretability question also remains open. The model’s outputs are biophysically grounded molecular phenotypes that can be checked in the lab, yet how the model arrives at its predictions — mechanism versus correlation — is only partially understood. DeepMind frames cell-type context awareness as a next step, a reminder that the current model does not fully condition its predictions on the specific cell type involved.
The balanced read: by DeepMind’s own reported numbers, AlphaGenome looks like a genuine step forward in predicting the molecular consequences of genetic variation, and it is packaged unusually openly for a system of its kind. What it is not is a disease-risk oracle. Researchers who respect that boundary get a powerful triage and hypothesis-generation tool; anyone who crosses it is reading past the evidence.
Frequently asked questions
What does AlphaGenome predict?
It predicts how a single-letter change in the roughly three-billion-character genome affects molecular phenotypes, mainly gene expression and RNA splicing. DeepMind presents these as molecular-level predictions, not disease diagnoses.
How is AlphaGenome different from Enformer?
DeepMind describes AlphaGenome as the first model to combine base-pair resolution with a one-million-base-pair context window. Its predecessor Enformer worked at coarser resolution over a smaller window, which could miss long-range interactions between distant DNA regions.
Is AlphaGenome free to use?
It is free for non-commercial academic research, and the model weights are available for fine-tuning, so labs can adapt the model to their own data rather than using it only as a fixed tool.
What is the AlphaGenome Atlas and the AVI score?
The Atlas is a petabyte-scale, precomputed dataset covering the predicted effects of all roughly nine billion possible single-letter variants — reported to be about 30 times the size of the AlphaFold database. The AVI impact score helps researchers triage which variants matter most.
Can AlphaGenome predict disease risk?
No. DeepMind acknowledges that linking predicted molecular effects to whole-organism disease susceptibility still requires research. The model indicates how variants change splicing and gene activity, not whether a person will develop a disease.