AlphaGenome 2 aims to predict DNA changes in individual cell types

Sequencing human DNA took 13 years, but understanding what the resulting letters do has taken longer, says Hannah Fry. Google DeepMind's AlphaGenome predicts effects of non-coding DNA across tissues; AlphaGenome 2 extends that work to individual cell types.

AlphaGenome reads a stretch of DNA and predicts molecular activity at that location, including how a single-letter change might alter it. The goal is to help researchers study sequences and variants, not directly read a person's future health from their genome.

Žiga Avsec came to genomics after studying physics and discovering machine learning. His first project sought variants in patients with rare diseases; the difficulty of understanding mutations stayed with him. Adam Gayoso came from computer science and operations research, motivated in part by illness in his family, and found his way into single-cell genomics.

A genotype is the sequence of DNA letters; a phenotype can be a visible trait or a molecular feature such as gene expression in a cell. The relationship between sequence and complex outcomes like Alzheimer's or type 2 diabetes involves many steps, so the researchers focus on molecular measurements closer to the genome, where more data is available.

Some conditions can be traced to a variant affecting a particular gene, while complex diseases reflect multiple genes and environmental factors. The researchers caution that a one-variant, one-disease picture is not always appropriate.

The researchers compare proteins to building blocks. Nearly all cells carry the same DNA, but different cells need different proteins for different jobs. Roughly 2% of the genome codes for proteins; regulatory DNA helps determine when and how much of them cells produce. Different cell types use combinations of these controls.

The team began with Enformer, a model built with Calico to read DNA sequences. After AlphaFold 2, they also explored protein-coding variants: AlphaFold 2 was designed to be robust to mutations and did not predict their effects well, but its learned information could be adapted.

AlphaMissense builds on AlphaFold and is fine-tuned to predict whether protein-changing variants may be harmful or benign. The researchers say these predictions have improved, although the models still do not explain every physical mechanism and protein-coding DNA is only a small part of the genome.

Returning to non-coding DNA, the team expanded Enformer's 200,000-letter input to one million and improved prediction resolution from 128 base pairs to individual base pairs in AlphaGenome. Because running the model requires computing resources, they created an Atlas with precomputed predictions for roughly nine billion possible single-letter changes.

The Atlas combines the model's many outputs with other information, including evolutionary conservation, into a variant impact score to help researchers rank candidates. A high score suggests a variant deserves attention; it is a prioritization tool rather than a diagnosis.

The team trains on public measurements collected by universities and laboratories, such as gene expression and DNA accessibility. Those measurements are aligned to a reference genome; the model is not trained on individual genomes. It learns relationships between the reference sequence and molecular measurements, then makes predictions about altered sequences.

The ENCODE project and other community efforts helped create the experimental data used to interpret the genome. Fry compares the sequenced genome to a shelf of books whose contents still need interpreting; differences among people can help researchers connect variants to traits and rare diseases.

The variant impact score summarizes which changes may merit closer study. The researchers trained a separate model partly using whether variants are common or absent in human populations: common variants are generally benign, while absence can reflect harm or simply limited sampling. They describe this as a noisy signal and stress that the models remain imperfect.

The Atlas is a research tool, not a clinical diagnostic. A prediction may suggest that a variant changes how a gene is processed, but it is only one piece of a larger investigation. Researchers must bring together independent clues and validation before gaining confidence.

The team gave early Atlas access to Broad Institute researchers investigating unresolved rare diseases. In one case, researcher Laura Koval examined a child diagnosed with epilepsy at six weeks after years of inconclusive testing. Ranking variants put a plausible candidate at the top; its predicted effect fit the condition and earlier literature, and the researchers tested it experimentally.

The suspected variant affected RNA splicing, the process of cutting and joining RNA. Researchers inserted a DNA construct into a cell line and compared the resulting RNA processing with and without the change. The experiment supported the predicted splicing effect.

A molecular diagnosis can help guide the search for treatment and inform family planning, the researchers say. Possible approaches include short RNA-based medicines and genome editing, but they emphasize that identifying a variant remains a long way from developing an effective treatment.

Genome-wide association studies compare DNA differences among many people and look for links to traits, from how coriander tastes to disease risk. An association does not by itself identify the causal change or explain its molecular effect.

Because stretches of DNA are inherited together, an associated variant may merely travel alongside the variant that actually has an effect. Computational models and statistical methods can help distinguish candidates. Association studies generally focus on common variants and are less applicable to very rare diseases, where too few people share the same condition or variant.

AlphaGenome can systematically predict how variants affect molecular properties, then inspect the biological context in which a change appears. That context may be a particular cell type or state, such as an immune cell responding to inflammation.

For a group of variants associated with a disease, the researchers want to know which variant is causal, which gene it affects, and which cell type is involved. Earlier models often predicted measurements from mixed cells at the tissue level. AlphaGenome 2 aims to resolve effects at the level of individual cell types, drawing on single-cell data including the Human Cell Atlas.

Cell types derived from the same stem cell can respond differently to the same variant, and effects may also differ by developmental stage. The updated model is designed to make more specific predictions about those differences rather than treating an entire tissue as one mixture.

As an example, the researchers examined a region associated with inflammatory bowel disease, where no gene lies immediately nearby. AlphaGenome 2 predicted that a variant affects a gene 290,000 base pairs away. A separate study at the Crick Institute had experimentally supported that relationship; the model reproduced the finding retrospectively.

Training data are uneven: blood and immune cells are easier to sample and may be overrepresented. The team says it assembled diverse human and mouse tissue data, but rare cell states can remain difficult to cover. Researchers can also provide their own data to AlphaGenome 2 to make predictions for their particular context without necessarily fine-tuning the model.

Fry asks how computational predictions compare with experiments in a dish or a body. The researchers say their models do not fully simulate cells and miss biological exceptions. They see value in generating clues at scale and iterating quickly, but say experiments remain essential.

The researchers describe the tools as high precision but low recall: when a strong effect is predicted, they say it is often real, yet many effects are missed. Even a positive prediction is the beginning of an investigation into pathways and disease, not a completed causal explanation.

DNA folds into a tangled three-dimensional structure, so regions distant in the linear sequence can influence each other. The researchers report that their model can make good predictions for distant variants when the relevant DNA is accessible, but performs less well in inaccessible regions. They do not yet know how much reflects data gaps or additional biological mechanisms.

The AlphaGenome family is available through a public API and Python package, while the Atlas offers a no-code website for looking up variants and predicted molecular effects. The researchers also discuss using coding agents to call the API and assemble information about a genomic region.

Looking ahead, the researchers want to predict the molecular effect of every variant and eventually build a virtual cell: a computational system that predicts how a cell's state changes after a genetic alteration, a drug or another intervention. They stress that useful simulation of biology remains very far away.

The hope is that better understanding of disease mechanisms could eventually help researchers develop medicines that succeed more often in clinical trials. Gayoso says achieving a virtual cell will require better data and infrastructure; he declines to give a timetable and predicts that artificial general intelligence will arrive first.

Fry reflects that linking a genetic recipe to its biological outcome is a monumental task because biology has evolved over billions of years. She sees the work as a step toward precision medicine that could eventually help researchers understand disease at a molecular level for individuals.