Google DeepMind AlphaGenome 2 Doubles Single-Cell Variant Recall

(LONDON —) Google DeepMind has unveiled AlphaGenome 2, a next-generation multimodal DNA sequence model that shifts computational genomics from averaged bulk-tissue predictions to single-cell-type and cell-state resolution. Trained on more than 107 million individual cell profiles spanning 68 human and mouse tissues, the system ingests a one-million-base-pair DNA window alongside an explicit biological condition vector to predict how non-coding genetic variants alter gene regulation in a specific target cell.

The architectural pivot directly addresses a persistent bottleneck in genome-wide association studies (GWAS): disease-linked mutations frequently operate only within specialized or transient cellular states, leaving their signal diluted in whole-tissue averages. In internal evaluations reported by DeepMind, AlphaGenome 2 nearly doubled causal variant recovery compared to a single-cell fine-tuned version of its predecessor while introducing DNA methylation prediction and zero-shot adaptation to user-supplied single-cell RNA sequencing (scRNA-seq) datasets.

Technical Architecture & Engineering Breakdown

The original AlphaGenome, previewed in June 2025 and published in Nature in January 2026, operated as a multi-task parallel predictor. Given a 1,048,576-base-pair (1 Mb) sequence, it projected regulatory outputs simultaneously across thousands of fixed tracks representing bulk tissues and immortalized cell lines. According to Google DeepMind, that parallel output topology spread representational capacity too thin across heterogeneous assays, blunting the network's sensitivity to subtle regulatory divergence between closely related cell subtypes.

AlphaGenome 2 replaces fixed multi-head broadcasting with a conditional sequence-to-function architecture. Instead of emitting every tissue track in a single forward pass, the model accepts two primary inputs: the 1 Mb DNA sequence (parameterized for human or mouse) and a target biological condition representing a specific tissue, cell type, or dynamic cell state. Crucially, DeepMind injects this biological condition at every stage of the network rather than only at the final readout layer. Early and intermediate sequence representations are therefore modulated by cellular context, allowing the same distal enhancer motif to be interpreted differently in a resting T cell versus an activated gut macrophage.

Cell Encoder & Zero-Shot Generalization

To prevent the model from being locked into a static taxonomy of training labels, AlphaGenome 2 incorporates a dedicated cell encoder. Researchers can pass raw single-cell RNA-seq expression counts from their own experiments—requiring as few as a few dozen cells—into the encoder to synthesize a continuous condition embedding on the fly. Because the network maps expression profiles into a shared latent space, it generalizes to previously unseen cell states and experimental perturbations without gradient updates or fine-tuning.

This conditional design also unlocks cross-modality assay imputation. When provided with a single experimental measurement such as gene activity, AlphaGenome 2 predicts unmeasured nuclear properties for that exact cell state, including chromatin accessibility (ATAC-seq), histone modifications, transcription factor binding, RNA splicing, 3D DNA contact maps, and—new in version 2—CpG DNA methylation. By predicting chemical methylation switches alongside physical chromatin packaging at single-base resolution, the model helps wet-lab teams prioritize which high-cost epigenetic assays to run next.

The shift toward tightly conditioned simulation mirrors DeepMind's broader architectural trajectory across scientific and interactive domains, as seen in its recent work on real-time interactive environments like Genie 4 world models. Rather than generating static, unconditioned averages, frontier systems are increasingly built around state-conditioned rollouts that respond dynamically to fine-grained context.

Benchmark Evaluation & Rigorous Comparison

Google DeepMind evaluated AlphaGenome 2 across single-cell expression quantitative trait loci (sc-eQTL) fine-mapping, bulk-tissue expression, CRISPR interference (CRISPRi) enhancer-gene linking, and unseen clinical cohorts. All reported metrics stem from DeepMind's own technical evaluation and have not yet undergone independent third-party replication.

The headline gains appear on the TenK10K benchmark, a large-scale single-cell map spanning 28 primary immune cell types. Fixing precision at 90%—meaning nine out of ten positive calls are verified causal variants—AlphaGenome 2 recovered 30% of causal sc-eQTLs, nearly double the 16% recall achieved by AlphaGenome 1 fine-tuned on single-cell data (AG1-sc). On directional effect prediction (determining whether a variant upregulates or downregulates target gene expression at 90% sign accuracy), recall rose from 23% for AG1-sc to 43% for AlphaGenome 2. Conversely, roughly 70% of causal variants still evade detection at the 90% precision threshold, underscoring that sequence-to-function modeling remains an open frontier.

Distal Enhancers, Continuous Trajectories & Clinical Cohorts

On held-out ENCODE-rE2G CRISPRi perturbation data, AlphaGenome 2 reached an area under the precision-recall curve (AUPRC) of 0.78, outperforming AlphaGenome 1 (0.73) and a Borzoi ensemble (0.61). The architectural advantage was sharpest for long-range regulatory wiring 100 kb to 2,500 kb from the target promoter, where AUPRC jumped from 0.48 to 0.62. Across bulk GTEx tissues, median Pearson correlation (r) for cross-tissue expression changes improved from 0.61 to 0.71, while within-tissue explained cell-type variance for scRNA-seq climbed from 23% to 34%. In cross-modality imputation, AlphaGenome 2 outperformed the Avocado baseline on 166 of 181 held-out tracks.

To test zero-shot generalization on unseen disease tissue, DeepMind evaluated the model on brain samples (substantia nigra and cingulate cortex) from 190 donors with and without Parkinson's disease. Conditioned solely on the cohort's scRNA-seq profiles without retraining, AlphaGenome 2 recovered 41% of eQTL effect directions at 87.3% accuracy, more than doubling the 17% recall of a specialized Scooby model trained directly on that dataset.

DeepMind also highlighted three mechanistic validations. First, on inflammatory bowel disease (IBD), the model computationally reproduced a landmark Nature finding, isolating the variant rs2836882 inside a regulatory enhancer roughly 290 kb away that selectively overactivates the ETS2 gene in gut macrophages while sparing neighboring genes and other immune cells. Second, applied to a COVID-19 mRNA vaccine cohort, it separated vaccine-activated T cells from resting T cells, demonstrating that a severe-COVID risk variant exerts its disruptive effect primarily in the activated state. Third, along continuous hematopoiesis trajectories, it tracked blood stem cell maturation into red blood cells with markedly higher fidelity than fine-tuned AlphaGenome 1.

Deployment, Compute Footprint & Inference Optimization

Conditioning every layer of a 1 Mb transformer-convolutional hybrid on a single target state introduces a direct compute trade-off. Because AlphaGenome 2 evaluates one biological condition per forward pass rather than amortizing thousands of bulk tracks simultaneously, its per-query inference cost is substantially higher than that of AlphaGenome 1. Google DeepMind genomics lead Žiga Avsec noted that this compute profile made regenerating a full petabyte-scale AlphaGenome Atlas—which previously precomputed all 9 billion possible single-letter substitutions in the human genome—impractical with version 2.

To bridge that gap without forcing labs to run redundant API calls, DeepMind has published two targeted precomputed resources alongside the model: a panel of approximately 6.8 to 7 million trait- and disease-associated variants scored across 510 human cell and tissue conditions, and an enhancer-gene regulatory map spanning 640 conditions. The original AlphaGenome remains active in service for broad, unconditioned tissue sweeps and local execution (where v1 weights via Kaggle and Hugging Face recommend at least an NVIDIA H100 GPU).

AlphaGenome 2 was trained on 2,048 TPU v4 chips, and DeepMind has not yet disclosed its exact parameter count or local VRAM requirements. At launch, access is restricted to the hosted AlphaGenome API—free for non-commercial academic research with demand-variable rate limits, and gated behind a closed preview for commercial biotech and pharmaceutical teams. DeepMind indicated that open weights for non-commercial research are planned upon final peer-reviewed publication.

# Install official Google DeepMind AlphaGenome Python client
pip install alphagenome

# Query AlphaGenome 2 via hosted API using a custom scRNA-seq cell state
python3 -c "
from alphagenome.models import dna_client
from alphagenome.data import genome

client = dna_client.create(api_key='YOUR_NON_COMMERCIAL_KEY')
interval = genome.Interval(chromosome='chr21', start=39700000, end=40748576) # 1 Mb window
variant = genome.Variant(chromosome='chr21', position=40465534, reference_bases='G', alternate_bases='A')

# Condition on an unseen gut macrophage scRNA-seq profile (zero-shot)
scores = client.predict_variant(
    interval=interval,
    variant=variant,
    model_version='alphagenome-2',
    cell_state_counts='gut_macrophage_scrna.h5ad',
    outputs=['RNA_SEQ', 'ATAC', 'CPG_METHYLATION']
)
print(scores.summary())
"

Beyond direct Python API calls, Google DeepMind plans to integrate dedicated “AlphaGenome 2 Skills” into Google Antigravity, its agentic development environment. Paired with reasoning models exposed through the Google Gemini API, these skills are designed to automate multi-step target-validation pipelines—allowing autonomous research agents to pull GWAS loci, condition AlphaGenome 2 on patient-derived scRNA-seq clusters, and cross-reference enhancer-gene links without manual scripting. This mirrors a wider industry shift toward proactive agent architectures capable of orchestrating specialized domain models inside reproducible scientific workflows.

For computational biologists and drug-discovery engineers, AlphaGenome 2 establishes a clear new baseline for non-coding variant interpretation. While its 70% miss rate at high precision and lack of downloadable weights limit immediate on-premise deployment, the combination of per-state conditioning, CpG methylation readouts, and few-dozen-cell zero-shot adaptation turns static GWAS hits into testable, cell-specific mechanistic hypotheses.