AlphaGenome

AlphaGenome predicts what non-coding DNA does. Given up to a million base pairs of sequence it returns gene expression, splicing patterns, chromatin state and contact maps — attacking the regulatory genome, which is where most disease-associated variants sit and where interpretation has been weakest.

What it does

The model is multimodal, producing several classes of functional prediction from one sequence input, at single-base resolution for most outputs. It is served as a remote API accessed through a Python SDK, free for non-commercial use with an API key. DeepMind released the source code in January 2026, which opened the model to independent scrutiny.

Strengths

• Predicts multiple regulatory modalities from a single input.
• Million-base-pair context, enough to link distal elements to targets.
• Single-base resolution for most output types.
• Free for non-commercial use, with source code now released.

Limitations

It runs as a remote API by default, so sequence leaves your machine — a consideration for confidential or patient-derived data. Non-commercial terms restrict industrial use. Regulatory prediction remains genuinely hard, and outputs are hypotheses to test with reporter assays or perturbation, not conclusions.