- AlphaGenome reads one million DNA letters at single-letter precision.
- It matched or outperformed prior models in 25 of 26 benchmarks.
- Predictions for rare variants still require experimental validation before clinical use.
Žiga Avsec spent five years solving a trade-off: how to read long DNA sequences while detecting single-letter changes.
His first attempt, Enformer, could read 200,000 DNA letters at 128-letter resolution. Better than anything before it, but still too coarse to pinpoint the individual mutations that disrupt gene regulation.
In June 2025, Avsec’s team at Google DeepMind released AlphaGenome, a new AI model that predicts how DNA changes affect gene activity, along with an accompanying preprint which explained what AlphaGenome is and what it does.
The full study was published in Nature in January, formalizing the results.
AlphaGenome can read one million DNA letters at single-letter precision- about five times more sequence than its predecessor Enformer, and at much finer resolution.
It matched or outperformed existing models in 25 of 26 variant-effect benchmarks.
Context vs precision
1,000,000
DNA letters AlphaGenome reads at single-letter precision, five times Enformer's context, 128 times its resolution
Solving Gene Regulation's Length-vs-Precision Problem
Understanding disease requires reading genes and the regulatory switches that control them. While the majority of established cancer drivers still reside in protein-coding regions, regulatory variants in the non-coding genome increasingly appear in genome-wide association studies for complex diseases.
These switches - enhancers that activate genes, silencers that suppress them -can sit hundreds of thousands of DNA letters away from their targets. A mutation in an enhancer might activate an oncogene. A single-letter change in a silencer might fail to stop it.
Google's AlphaGenome scans million‑letter stretches of DNA, reading the non‑coding “dark” 98% and the tiny 2% of DNA gene segments (imagined in gold in the image) to predict how single‑letter changes can alter gene activity. (Science Reader)
Previous models faced a trade-off. Enformer's 200,000-letter context captured long-range regulation but averaged predictions across 128-letter windows, missing the single mutations that matter. Borzoi reached 524,000 base pairs at the same 128-letter resolution, extending context but not improving precision.
AlphaGenome solved both problems simultaneously.
What is single-letter precision?
DNA consists of four letters (A, T, G, C). AlphaGenome predicts how changing a single letter affects gene activity. This is the resolution needed to interpret disease-causing mutations.
What the Model Can Actually Predict
AlphaGenome outputs predictions across 5,930 human molecular tracks and 1,128 mouse tracks. These include where genes start and stop in different tissues, where RNA gets spliced, how much RNA a cell produces, which DNA regions are accessible, and which proteins bind where.
The training data came from ENCODE, GTEx, 4D Nucleome, and FANTOM5 - public consortia that experimentally measured these properties across cell types and tissues.
Training took four hours on tensor processing units, half the compute Enformer required.
In validation tests, the model's predictions matched known patterns of TAL1 gene disruption in T-cell leukemia, recapitulating observed effects from sequence alone, though not establishing causal mechanisms in patients.
The TAL1 results demonstrated what researchers had suspected since 2014: oncogenic mutations could create new regulatory switches that activate the gene inappropriately.
It's a milestone for the field. For the first time, we have a single model that unifies long-range context, base-level precision and state-of-the-art performance.
Caleb Lareau, Memorial Sloan Kettering Cancer Center
The model serves as what researchers call a powerful hypothesis-generation tool. Predictions still require experimental validation before clinical use.
For expression quantitative trait loci (variants affecting how much RNA genes produce), AlphaGenome improved accuracy 25.5% over Borzoi. For chromatin accessibility, the improvement reached 45%.
The Data Quality Bottleneck
The model faces several reliability challenges. Predictions for rare variants and very distal regulatory elements (beyond 100 kilobases) remain unreliable. Training on bulk tissue data means the model averages signals across cell populations, potentially missing cell-type-specific regulation.
The model faces several reliability challenges.
As one expert cautioned, current AI models "are still not reliable enough for patient care". Predictions can overstate risks for certain genetic changes.
"Most existing data in biology is not very suitable for AI," says Ben Lehner, head of generative and synthetic genomics at the Wellcome Sanger Institute, who tested AlphaGenome with over 500,000 new experiments. "The datasets are too small and not well standardised. The most important challenge right now is how to generate the data to train the next generation."
Related reading
Your Intelligence Comes From Genes, Experience - and Molecular Switches
Epigenetic marks guide neurons during development and keep rewiring your brain throughout life. Neither genes nor environment win.
→Yet data quality presents a deeper challenge: different experimental assays can yield conflicting results, and biological measurements carry systematic biases. More data doesn't automatically resolve which regulatory mechanisms are real. That requires experiments that test predictions in living systems.
Three directions look promising. Integrating single-cell data could capture cell-type-specific regulation that bulk measurements miss. Fine-tuning on perturbation experiments, where researchers deliberately break regulatory elements, could improve causal inference.
Developing methods to detect and correct measurement biases could make predictions more reliable.
The model is available via API for non-commercial research. DeepMind has indicated plans to extend it to additional species and refine predictions for clinical applications.
Sources
- Primary Research: Advancing regulatory variant effect prediction with AlphaGenome (Nature, January 2026)
- Additional Context:
- AlphaGenome: AI for better understanding the genome (Google DeepMind)
- Expert reaction to paper on Google DeepMind's AlphaGenome (Science Media Centre)
- Genotype-Tissue Expression (GTEx) Portal (GTEx Consortium)
- ENCODE Project (ENCODE Consortium)
- Žiga Avsec of Google Deepmind: AI allows us to grasp the complexity of the genome (Slovenia Times)
Fact Check: Claim-by-Claim Verification Verified
Limits and uncertainties
Core capabilities like long-context single-base precision and broad multimodal predictions are strongly supported by DeepMind's primary sources and peer-reviewed publication.
Article accurately notes limitations: unreliable for rare variants, distal elements, bulk-averaging effects; needs experimental validation for causality.
Expert quotes (e.g., Lareau) match blog; data quality issues highlighted are consistent with field challenges.
No hyped claims; emphasizes hypothesis-generation tool. Readers should note reliance on public consortia data and ongoing needs for single-cell/perturbation training.
Non-coding 98% focus aligns with established genomics (protein-coding ~2%).
Bottom line
Article faithfully represents AlphaGenome's advances and limits based on primary sources. Claims are well-supported for research contexts, advancing non-coding variant interpretation.
Fact-checked by Perplexity Sonar Pro on 2026-02-08
