AlphaGenome 2: AI Reads the Human Genome to Pinpoint Disease-Causing Mutations

Sequencing the human genome took 13 years, and the resulting string of characters is many times longer than the complete works of Shakespeare. But reading this "book" has taken even longer. As geneticist Eric Lander, who participated in the Human Genome Project, put it: "We bought the book, but it's too hard to read." Google DeepMind is using AI to make this book gradually readable.

On the DeepMind podcast, the research team introduced the latest advances in the AlphaGenome family of models. Released in 2025, AlphaGenome can predict the function of non-coding DNA at the whole-tissue level, while AlphaGenome 2 further pushes resolution down to individual cell types, revealing how single-point mutations drive disease.

From "Junk DNA" to the Regulatory Code

To understand the value of AlphaGenome, you first need to understand the basic structure of the genome. The human genome consists of about 3 billion base pairs (A, C, G, T). Only a small fraction directly encodes proteins—these proteins are the "LEGO bricks" inside cells, building the molecular machines that perform various functions. The vast remainder was once called "junk DNA"; today we know these are regulatory regions that determine when, where, and in what quantity cells produce which proteins.

The research team used a vivid analogy: all cells carry the same set of genetic instructions, but eye cells need to detect light and muscle cells need to perform physical movement—they need different "machines," and therefore different combinations of proteins. Regulatory regions are the instruction layer that tells cells what to produce.

The design goal of AlphaGenome is precisely to read a stretch of DNA sequence and tell us what is happening at that location inside the cell, and what would change at the molecular level if one letter were altered (say, A to T).

Non-coding DNA and Regulatory Elements

About 98% of the human genome does not directly encode proteins; these regions are collectively called non-coding DNA. They are not useless—they contain various regulatory elements: promoters (located near gene transcription start sites, recruiting RNA polymerase), enhancers (which can be located hundreds of thousands of bases upstream or downstream of a gene, activating gene expression by binding transcription factors), silencers (which repress gene expression), and insulators (which define the boundaries of regulatory domains). By recruiting transcription factors and chromatin-modifying enzymes, these elements determine the cell type, developmental stage, and physiological state in which a gene is expressed. Understanding the function of non-coding regions is key to interpreting the genome, because the vast majority of disease-associated variants identified by genome-wide association studies (GWAS) lie in non-coding regions. AlphaGenome's core task is to predict the effects of these non-coding variants on molecular phenotypes (such as gene expression, splicing, and chromatin accessibility).

The Core Breakthrough of AlphaGenome 2: Cell-Type Resolution

Previous models of this kind typically made predictions at the "bulk" level—grinding and mixing a bunch of cells together to make judgments about tissues like heart or lung as a whole. AlphaGenome 2 aims to elevate model capability to the cell-type level, thanks to the large amount of single-cell data accumulated over the past decade by initiatives such as the Human Cell Atlas.

According to the two researchers, Ziga Avsec and Adam, blood contains many cell types, and the same variant may have completely different effects in T cells, red blood cells, and macrophages—even at different developmental time points. Through architectural improvements, AlphaGenome 2 significantly boosts prediction performance at the cell-type level and can determine whether a variant has an effect in a specific cell state (for example, an immune cell in a pro-inflammatory state).

A striking case in the paper involves macrophages and intestinal inflammation. In a locus associated with inflammatory bowel disease, the variant lies in a "blank" region between two genes, with no known gene nearby. AlphaGenome 2 predicted that one variant regulates a distant gene, and a team at the Crick Institute experimentally validated this prediction, confirming that the variant very likely does regulate that distant gene.

Single-Cell Sequencing and the Human Cell Atlas

Technologies such as single-cell RNA sequencing (scRNA-seq) enable researchers to measure gene expression and chromatin state at the level of individual cells, thereby distinguishing the molecular features of different cell types within the same tissue. The Human Cell Atlas is an international collaborative initiative aimed at using single-cell technologies to map reference atlases of all cell types in the human body, covering developmental, healthy, and disease states. These data provide cell-type-specific supervision signals for training AlphaGenome 2—the model needs to learn the regulatory output of the same DNA sequence in different cellular contexts. Without large-scale single-cell data, the model could not distinguish differences in the activity of the same enhancer in T cells versus macrophages. Therefore, AlphaGenome 2's breakthrough in cell-type resolution relies both on architectural innovation and on the data accumulation from single-cell atlas initiatives over the past decade.

The Atlas Platform: Computing All Possible Variants at Once

To make the model easier to use, the team built the AlphaGenome Atlas. They precomputed the effects of AlphaGenome on every possible single-letter change in the genome (about 9 billion variations) and opened it up as a website resource.

Atlas also introduces the AVI score (AlphaGenome Variant Impact Score), which compresses the model's tens of thousands of output values, AlphaMissense scores, and evolutionary conservation into a single number. The higher the score, the more attention the variant deserves. When geneticists receive a long list of suspected disease-causing variants, they can use AVI to rank them and prioritize high-scoring items for study.

The underlying logic of this approach comes from population observations: variants that appear frequently in the population are usually benign, while variants never observed in the population are more likely to be harmful. The model calibrates its scores by predicting the frequency of a variant in the population. The researchers readily admit that this is a very noisy signal, but it is enough to point the way.

Linkage Disequilibrium and Population-Frequency Calibration

Linkage disequilibrium (LD) refers to the tendency of adjacent variants on the genome to be inherited together, forming haplotype blocks. In GWAS, a marker variant statistically associated with a trait may merely be in the same LD block as the true causal variant, rather than being directly pathogenic. This makes it difficult to pinpoint GWAS signals to a single causal variant. The AVI score uses principles of population genetics for calibration: variants that appear at high frequency in the population have usually undergone purifying negative selection and are more likely to be benign, while variants never observed in the population may be harmful. The model adjusts its scores by predicting variant frequency in the population, thereby partially bypassing LD interference and directly assessing the molecular effect of the variant itself. The researchers also acknowledge that frequency signals are noisy, but as a ranking tool they are sufficient to guide the direction of experimental validation.

A Real Case: A Rare Disease Unsolved for Five Years

Laura Covell, a researcher at the Broad Institute, focuses on difficult-to-diagnose rare disease cases. One case involved an infant diagnosed with rare epilepsy at six weeks of age. Over the following four or five years, clinical geneticists ran numerous tests and confirmed a genetic cause, but could never find the specific variant.

Laura used the AVI score to rank the patient's variants. The top-ranked variant affects the function of a protein associated with epilepsy and is consistent with previously reported literature. The team then experimentally validated the change in splicing pattern, and all the puzzle pieces finally came together—the case was solved.

Finding a molecular diagnosis means more than just confirming the disease. The researchers note that knowing the root cause improves the chances of developing the right therapy; genetic medicines such as antisense oligonucleotides (ASOs) and genome editing offer potential treatment paths. It also directly helps the family with future family planning.

A Complementary Relationship with GWAS

Genome-wide association studies (GWAS) are a common tool in genetics: sequence large populations and look for statistical associations between genetic differences and traits. From "does cilantro taste like soap" to inflammatory bowel disease risk, GWAS covers an extremely wide range of traits.

But GWAS has two inherent limitations. One is linkage disequilibrium—DNA is inherited in blocks, so a seemingly important variant may merely be correlated with the true causal variant. The second is that it does not apply to rare diseases, because it requires large numbers of people carrying the same variant, and rare diseases precisely lack such samples.

AlphaGenome offers another path: feed variants directly into the model and systematically record molecular-level changes, without relying on population size. For associations found by GWAS, the model can also answer three key questions: which variant is causal, which gene it affects, and in which cell type it acts.

Limitations and Prospects: High Precision, Low Recall

The researchers are very candid about the model's limitations. They describe it as a "high-precision, low-recall" tool: when the model predicts that a variant has a strong effect, it is usually right; but it misses many. Biology is full of edge cases and exceptions, and the model cannot fully simulate processes inside the cell.

The three-dimensional structure of DNA also poses challenges. The researchers use "a bowl of spaghetti" to describe the tangled state of DNA—regions that are far apart linearly may be close in space and influence each other. Interestingly, the AlphaGenome 2 paper found that whether a variant is located in an accessible region is more predictive of model quality than its linear distance to a gene: variants in open regions can be predicted well even if far away, while variants in tightly compacted regions perform poorly.

The model is also affected by the distribution of training data, predicting more accurately for data-rich types such as blood immune cells. To address this, the team allows users to plug their own experimental data into the model, enabling variant-effect prediction for specific tissues without fine-tuning.

For users, the AlphaGenome family offers a public API and Python package that can be called for free; Atlas is a web interface that requires no programming. The researchers also mention that the rise of coding agents makes it possible to "ask questions in natural language, have an agent call AlphaGenome, and synthesize the analysis."

Three-Dimensional Chromatin Structure and Accessibility

DNA does not stretch out linearly inside the nucleus; it wraps around histones to form nucleosomes and further folds into chromatin loops and topologically associating domains (TADs). Spatial contact between enhancers and promoters is an important mechanism for regulating gene expression, and elements that are linearly distant may come close in three-dimensional space. Chromatin accessibility refers to whether a specific region is in an open state that allows transcription factors to bind, typically measured using techniques such as ATAC-seq. The AlphaGenome 2 paper found that whether a variant is located in an open chromatin region is more predictive of model performance than its linear distance to a gene: variants in open regions can be predicted well even when far from genes, while variants in compacted regions are predicted poorly. This suggests the model partially captures three-dimensional regulatory logic, but its resolution of highly folded regions remains limited by training data and technical constraints.

A Bigger Vision: Virtual Cells

The team's long-term goal is to build a "virtual cell"—given a cell state, a genome, and some perturbation (genetic variant or drug treatment), predict how the cell state will change. This is a vision Demis Hassabis often mentions. The researchers emphasize that this is very difficult, and AlphaGenome is only the first step.

What drives this vision is a real-world problem: the success rate of drug clinical trials is very low, largely because our understanding of the true causes of disease and how to intervene is insufficient. Better tools for understanding the genome could help us find the right direction.

When asked how many years virtual cells would take, Adam gave a telling judgment: "We might have AGI before we have virtual cells." Intelligence problems may be easier to solve than biological ones—perhaps the most humble footnote to the complexity of biology.

AlphaGenome has not existed for long, while life evolved over billions of years to weave such a complex regulatory network. Closing this gap is a daunting task, but if the future of precision medicine lies in pinpointing the cause of disease for each individual at the molecular level, tools like this are an indispensable foundational step.

One-line Summary

DeepMind releases AlphaGenome 2, pushing genomic variant effect prediction down to single-cell resolution, aiding rare disease diagnosis.

Paragraph Summary

On its podcast, Google DeepMind introduced the latest advances in the AlphaGenome family of models. Released in 2025, AlphaGenome can predict the function of non-coding DNA at the whole-tissue level, while AlphaGenome 2 further pushes resolution down to individual cell types, revealing how the same variant has differential effects in different cells such as T cells and macrophages. The team also built the Atlas platform, which precomputes the effects of about 9 billion single-letter variants and uses the AVI score to compress complex outputs into a single sortable metric. In a collaboration with the Broad Institute, the AVI score helped pinpoint a pathogenic splicing variant in a child with rare epilepsy that had gone undiagnosed for years, which was then experimentally validated. The researchers position the model as a high-precision, low-recall research tool, with a long-term vision of building virtual cells—though they believe this may be more difficult than achieving AGI.