Google DeepMind Unveils AlphaGenome 2 with Single-Cell Resolution for Genetic Medicine
Google DeepMind has released AlphaGenome 2, predicting gene expression and chromatin accessibility across human tissues down to single-cell resolution.

By Clara Tan, Artificial Intelligence
, updated 6 min read
Google DeepMind has introduced AlphaGenome 2, an expanded genomic foundation model that predicts gene expression and epigenetic regulation down to individual cellular states. Built as a successor to the original AlphaGenome architecture released in early 2025, the updated system maps how single-nucleotide variants alter cellular behaviour across more than 1,200 human cell types.
While earlier genomic models treated tissue samples as homogeneous bulk averages, AlphaGenome 2 explicitly models cellular heterogeneity. By resolving transcriptional outputs at single-cell fidelity, the architecture allows computational biologists to assess whether a genetic mutation disrupts transcription factor binding in a specific immune subpopulation, such as CD8+ effector memory T cells, without causing observable alterations in neighbouring tissue.
Architectural shifts and multiscale attention
The core technical advance in AlphaGenome 2 centres on a multiscale transformer architecture coupled with dilated convolutions capable of processing sequence contexts up to one million base pairs. Genomic regulation frequently relies on long-range chromatin loops where distal enhancers located hundreds of kilobases away come into physical contact with gene promoters.
To capture these interactions without quadratic computational scaling, DeepMind engineers combined sparse local attention blocks with hierarchical latent pooling. The model was trained across thousands of publicly available single-cell ATAC-seq (Assay for Transposase-Accessible Chromatin) and single-cell RNA-sequencing datasets curated from the Human Cell Atlas, ENCODE 4, and academic consortia.
During training on TPU v5p clusters, the loss function prioritised cell-type-specific chromatin accessibility and directional splice-junction usage. The resulting representations predict not only whether a genomic region remains open for transcription, but also the quantitative expression levels of downstream transcripts across defined developmental lineages.
Clinical utility in non-coding variant interpretation
More than 98% of the human genome consists of non-coding sequence, yet genome-wide association studies (GWAS) regularly identify disease-linked variants in these regions. Clinicians have historically struggled to interpret non-coding variants because traditional sequence alignment tools provide limited functional insight into regulatory disruption.
In benchmark evaluations conducted on clinical cohorts from the UK Biobank and ClinVar, AlphaGenome 2 demonstrated a 34% improvement in classifying pathogenic non-coding variants compared to ensemble baselines like Enformer and Borzoi. In autoimmune disease cohorts, the system identified distinct causal variants disrupting regulatory motifs in regulatory T cells, resolving previously ambiguous genetic linkage blocks into actionable molecular targets.
Pharmaceutical researchers at partner institutions used AlphaGenome 2 in retrospective validation trials to evaluate oncogenic promoter mutations in glioblastoma. The model predicted transcriptional upregulation of alternative kinase pathways, matching observed patient resistance mechanisms to standard tyrosine kinase inhibitors.
Open-access research deployment
Google DeepMind is providing pre-trained model weights, inference checkpoints, and an open Python software development kit to academic researchers under non-commercial research licences. Commercial biotechnology and pharmaceutical firms can access the architecture through Google Cloud Vertex AI under enterprise service-level agreements.
The release also integrates with European Bioinformatics Institute repositories, allowing researchers to query AlphaGenome 2 variant impact scores directly from Ensembl variant annotation pipelines. By standardising variant predictions into calibrated probability distributions, the platform provides geneticists with a consistent computational baseline for variant classification.
Frequently Asked Questions
What differentiates AlphaGenome 2 from AlphaFold?
AlphaFold predicts the three-dimensional folding structure of proteins from amino acid sequences. AlphaGenome 2 models raw non-coding and coding DNA sequences to predict gene regulation, chromatin accessibility, and transcriptional outputs across distinct human cell types.
Can AlphaGenome 2 identify disease-causing mutations?
The model scores how single-nucleotide mutations and structural variants alter transcription factor binding and gene expression. Geneticists use these scores to prioritise likely pathogenic variants in non-coding DNA for laboratory validation.
Which computational hardware is required to run local inference?
Local inference on sequence windows up to 250 kilobases requires an accelerator with at least 24 gigabytes of VRAM, such as an Nvidia RTX 4090 or A10G. Sequence analysis extending to the full one-megabase context window requires enterprise accelerators such as the Nvidia H100 or Google Cloud TPU v5e.
Sources
- 1Google DeepMind Researchdeepmind.google
- 2Nature Geneticsnature.com