DeepMind's genome model passed peer review. It reads the part that isn't genes

Most of a human genome does not code for protein. More than 98 percent of observed genetic variation in humans lies outside protein-coding regions, and interpreting what those variants do has been one of the hardest problems in genomics [s1].

On January 28, Nature published the peer-reviewed description of AlphaGenome, a deep-learning model from Google DeepMind built for that problem [s1].

What the model does

AlphaGenome takes one million base pairs of DNA as input and predicts thousands of functional genomic tracks at up to single-base-pair resolution [s1]. A genome track is a data format that assigns a value to each base pair — a read count or signal derived from an experimental assay performed in a cell line or tissue.

The modalities span gene expression, transcription initiation, chromatin accessibility, histone modifications, transcription factor binding, chromatin contact maps, splice site usage, and splice junction coordinates and strength [s1].

Žiga Avsec, the paper's lead author and a Google DeepMind researcher, describes the model as predicting "gene expression, DNA accessibility, histone modifications, transcription factor binding, and even folding structure of the genome," with high accuracy in DNA sequences up to a million base pairs long [s2].

To predict what a variant does, the model compares its predicted tracks for a reference sequence against its predictions for the altered sequence. The difference is the predicted molecular consequence.

The trade-off it claims to break

Existing sequence-to-function models have faced a structural trade-off between how much sequence they can take in and how finely they can resolve the output, which limits both the range of modalities they can address and their performance [s1]. Long-range regulatory interactions and base-pair resolution have been difficult to have at once, largely for computational reasons [s1].

AlphaGenome's central claim is that it does not require that choice: one megabase of input, thousands of tracks, single-base-pair resolution [s1].

The result

Trained on human and mouse genomes, AlphaGenome matches or exceeds the strongest available external models in 25 of 26 evaluations of variant effect prediction [s1].

The paper also reports a mechanistic demonstration: the model's ability to score variant effects across all modalities simultaneously accurately recapitulates the mechanisms of clinically relevant variants near the TAL1 oncogene [s1].

That second result is the more interesting one for anyone asking whether the model is doing biology or curve-fitting. Benchmark wins show that a model reproduces measurements; recovering a known regulatory mechanism at a specific locus shows that the internal representation tracks something about how the regulation actually works. One locus is not a general proof, but it is the right kind of evidence.

What this is not

It is worth being precise about the boundaries.

A benchmark is not a clinic. Twenty-five of 26 evaluations is a strong result against other computational models [s1]. It is not evidence that the model's predictions are accurate enough to guide interpretation of a patient's variant, which requires validation against clinical outcomes rather than against assay measurements.

Trained on human and mouse. The model was trained on those two genomes [s1]. Its behaviour on other organisms is outside what the paper establishes.

Prediction is not causation. The output is a predicted molecular consequence — a change in expression, accessibility, or splicing. The path from a molecular consequence to a disease phenotype involves cell type, developmental timing, genetic background, and environment, none of which the model scores.

Why publication now, when the model shipped in June

AlphaGenome has been publicly available for non-commercial use since June 2025, when a preprint describing it was released, with the latest version available on GitHub [s2].

That gap between release and peer review is now common in machine learning and worth naming as a change in how this kind of science reaches the field. Researchers used the model for seven months before the reviewed paper appeared. Pushmeet Kohli, vice president of research at Google DeepMind, says that early-release feedback allowed researchers to pinpoint mutations in cancer genomes that drive proliferation, among other uses [s2].

Kohli describes the model as "our solution to deciphering the complex regulatory code" that makes up most of a genome [s2].

The upside is faster iteration and real-world testing before publication. The cost is that the scientific record trails the deployment — claims circulate, and get built on, before reviewers have examined them.

Where it sits in a lineage

The same lineage produced AlphaFold, whose creators were awarded the 2024 Nobel Prize in Chemistry, and AlphaMissense, released in 2023 to predict the effects of mutations in the protein-coding regions of the genome [s2].

AlphaMissense addressed the roughly 2 percent of the genome that codes for protein. AlphaGenome is aimed at the rest.

What to watch

Whether independent groups reproduce the benchmark results on data the model did not see, and whether performance holds on variant classes and cell types under-represented in training data.

Whether the predictions prove useful in the setting where the need is sharpest: the interpretation of variants of uncertain significance found in clinical sequencing.

C&EN's own conclusion is the appropriately cautious one — that only time will tell how much impact AlphaGenome has on genome research [s2].