AlphaGenome

What is AlphaGenome?

AlphaGenome is a deep learning model from Google DeepMind that predicts how genetic variants affect gene regulation. Its RNA-seq output contains 667 tracks spanning 285 unique biosamples at single base-pair resolution.

The model processes up to 1 million base pairs of DNA sequence and outputs predictions for gene expression (RNA-seq), splicing patterns, chromatin accessibility, and other regulatory features. Unlike models that focus on protein-coding regions (about 2% of the genome), AlphaGenome specializes in the non-coding 98% where most disease-associated variants reside.

AlphaGenome builds on DeepMind's earlier Enformer model and complements AlphaMissense, which predicts effects of variants within protein-coding regions.

How to use AlphaGenome online

Submit one VCF-style variant and a personal DeepMind API key to run AlphaGenome RNA-seq predictions online. ProteinIQ returns run details and track metadata for inspection, plus the paired reference and alternate full-resolution arrays for download without local Python setup.

AlphaGenome on ProteinIQ uses a Bring Your Own Key model. The DeepMind key supplies access to the model API. ProteinIQ credits still apply for job orchestration, result processing, and storage; the run form shows the current credit estimate before submission.

Why do you need your own key?

AlphaGenome is licensed by DeepMind for non-commercial research use only. This means ProteinIQ cannot host the model on its own servers or provide a shared API key. Doing so would violate the commercial use restrictions. By using a personal API key, researchers access AlphaGenome directly through DeepMind's infrastructure under their own non-commercial research agreement. Users are responsible for ensuring their use complies with DeepMind's Terms of Use.

How to get an API key

1. Visit deepmind.google.com/science/alphagenome
2. Sign in with a Google account
3. Accept the Terms of Use (non-commercial research only)
4. Copy the generated API key

The API key is free for academic and non-commercial research. Organizations seeking commercial access should contact DeepMind through their dedicated inquiry form.

API key security

ProteinIQ does not store API keys in its database. Keys are stripped from job records before storage and are only used transiently to communicate with DeepMind's API. However, the key does pass through ProteinIQ's processing servers during job execution.

Inputs

Settings

Output settings

Results

The output includes:

• Run details: Requested output, interval, variant, ontology filter, track count, and track resolution
• Track metadata: Downloadable JSON plus an in-app preview of the returned RNA-seq tracks, including ontology and biosample annotations
• Full resolution data: Downloadable .npz file containing the reference and alternate RNA-seq arrays plus interval, variant, and track metadata payloads

The downloadable .npz archive is intended for downstream analysis in Python, while the JSON metadata export helps map each returned track back to its ontology and biosample labels.

How AlphaGenome works

AlphaGenome uses a hybrid architecture combining convolutional neural networks and transformers:

1. Convolutional layers detect short sequence patterns (motifs) in the input DNA
2. Transformer layers propagate information across all positions, capturing long-range regulatory interactions
3. Output heads convert learned representations into predictions for different functional modalities

The model was trained on thousands of experimental datasets measuring gene expression, chromatin accessibility, histone modifications, and transcription factor binding across diverse cell types. Training completed in four hours on TPUs—half the compute of the earlier Enformer model.

For variant effect prediction, AlphaGenome generates paired predictions for the reference and alternate sequence. ProteinIQ preserves both arrays so downstream analyses can calculate the effect measure appropriate for each track.

Limitations

• Research use only: Predictions are not validated for clinical diagnostics
• Long-range interactions: Accuracy decreases for regulatory elements more than 100,000 bp from the variant
• Non-coding focus: For variants in protein-coding regions, AlphaMissense may be more appropriate
• Window size constraints: Only specific window sizes are supported (16K, 128K, 512K bp)
• Rate limits: The API is designed for small to medium analyses (thousands of predictions), not genome-wide scans requiring millions of predictions