On June 25, 2025, Google DeepMind introduced AlphaGenome, an AI model that reads up to 1 million DNA base pairs and predicts thousands of molecular properties that describe how genes are regulated, at the resolution of individual DNA letters. It can also score how a single genetic variant changes those properties, a central question in understanding disease. DeepMind released AlphaGenome as a preview API for non-commercial research, and in January 2026 it published the work in Nature and released the model code.

Key Facts#

  • Announcement: June 25, 2025, with a preprint and a preview API for non-commercial research.
  • Input and output: DNA sequences of up to 1 million base pairs in, predictions at single base-pair resolution out for most output types.
  • Predicted properties: gene expression, RNA splicing (including splice junctions), chromatin accessibility and features, protein binding and 3D contact maps across hundreds of human and mouse cell types and tissues.
  • Performance claims: DeepMind says AlphaGenome beat the best external models on 22 of 24 single-sequence evaluations, and matched or exceeded them on 24 of 26 variant-effect evaluations.
  • Efficiency: training a single model took four hours and half the compute budget of DeepMind's earlier Enformer model, according to DeepMind.
  • Terms: the API is free for non-commercial use. Outputs are for non-commercial use only and must not be used to train other machine learning models.
  • Follow-up: in January 2026, DeepMind published the research in Nature and released a JAX implementation of the model and its weights through a research repository.

What Happened#

AlphaGenome takes a long stretch of DNA and predicts where genes start and end in different cell types, where they are spliced, how much RNA is produced, and which bases are accessible, close to one another or bound by specific proteins. To score a variant, it compares predictions for the mutated sequence with predictions for the reference sequence, and DeepMind says it can summarize the impact across all modalities in about a second.

The architecture combines convolutional layers, which detect short sequence patterns, with transformers that pass information across the full 1-million-letter window, followed by output layers for each modality. Training data came from large public consortia, including ENCODE, GTEx, 4D Nucleome and FANTOM5. DeepMind highlighted a new capability: modeling the location and expression level of RNA splice junctions directly from sequence, which matters for rare diseases caused by splicing errors, such as spinal muscular atrophy and some forms of cystic fibrosis.

As a demonstration, DeepMind applied AlphaGenome to mutations previously observed in patients with T-cell acute lymphoblastic leukemia. The model predicted that the mutations activate the nearby TAL1 gene by introducing a MYB DNA-binding motif, which reproduced the known disease mechanism.

Background#

Only about 2 percent of the human genome codes for proteins. DeepMind's earlier AlphaMissense model focused on variants in those protein-coding regions. The other 98 percent, the non-coding regions, controls when and where genes are active and contains many variants linked to disease, but its effects have been much harder to predict.

AlphaGenome builds on Enformer, DeepMind's earlier gene-expression model. Previous sequence models had to trade context length against resolution: they could look far along the DNA or see individual letters, but not both. AlphaGenome's main technical advance is doing both at once without a large increase in training cost, which lets one model cover many modalities that previously required separate specialized models.

DeepMind was also clear about the limits. Capturing the influence of regulatory elements more than about 100,000 letters away remains difficult, cell-type and tissue-specific patterns need further work, and the model was not designed or validated for predicting an individual's personal genome. It predicts molecular outcomes, not the full path from a variant to a complex trait, and it is not intended for clinical use.

Why It Matters for Developers#

Scientific foundation models are arriving as APIs and research code, and developers will integrate them into larger systems. The AlphaGenome client is a Python package, so a .NET platform in healthcare or life sciences will typically call it from a small Python service. .NET Aspire can orchestrate that service next to your .NET APIs through its Python hosting integration:

C#
// AppHost: a .NET API calls a FastAPI service that wraps the AlphaGenome Python client.
var builder = DistributedApplication.CreateBuilder(args);

var apiKey = builder.AddParameter("alphagenome-api-key", secret: true);

var genomics = builder.AddUvicornApp("genomics", "../genomics-service", "main:app")
    .WithUv()
    .WithEnvironment("ALPHAGENOME_API_KEY", apiKey);

builder.AddProject<Projects.VariantApi>("variant-api")
    .WithReference(genomics)
    .WaitFor(genomics);

builder.Build().Run();

Design around the service's constraints. DeepMind says query rates vary with demand, that the API suits analyses needing thousands of predictions rather than more than a million, and that a precomputed "Atlas" of variant-effect scores offers higher throughput. Cache results, batch requests and prefer the precomputed data where it fits. Our caching guide covers patterns that apply directly.

Treat the terms of use as requirements. Because outputs are for non-commercial use and may not be used to train other models, record provenance for every prediction you store, and keep AlphaGenome outputs out of training datasets. Because the predictions are not validated for clinical decisions, any product that shows them to clinicians needs clear labeling and human review, which our guide to responsible AI in .NET discusses.

What's Next#

At launch, DeepMind said it planned to release the full model so that scientists could fine-tune it on their own data, and that extending the training data could add species and modalities. It delivered the first part in January 2026, when the Nature paper appeared alongside the alphagenome_research repository, which contains a JAX implementation of the model, the API implementation with variant scorers and a training-data loader.

The open questions are scientific rather than technical: how well predictions hold up in wet-lab validation across cell types, and whether the model's representations generalize to rare variants in under-studied populations. For developers, AlphaGenome is a template for how domain AI will be consumed: a specialized model behind an API, strict usage terms and a surrounding application that handles orchestration, caching, compliance and presentation. For a broader view of the options, see our overview of AI in .NET.

Sources#