AlphaGenome: AI for better understanding the genome

Source: DeepMind Date: 2025-06-25 URL: https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/

Summary

Google DeepMind released AlphaGenome, a model that processes up to 1 million DNA base pairs at single-letter resolution and jointly predicts thousands of regulatory molecular properties across hundreds of human and mouse cell types and tissues. Outperformed best external models on 22 of 24 DNA sequence prediction tasks and matched/exceeded top performers on 24 of 26 variant-effect evaluations. Training: 4 hours using half the compute of the prior Enformer model. Trained on ENCODE, GTEx, 4D Nucleome, and FANTOM5 public consortia data.

Implications

AlphaFold for regulatory genomics. AlphaFold solved protein structure; AlphaGenome attacks regulatory genomics — how DNA sequence variants alter gene regulation. This is the next layer of the molecular biology stack. The joint prediction of all modalities (not task-specific tools) is the same architectural philosophy as AlphaFold 3’s molecular scope expansion.

22/24 and 24/26 benchmark wins are strong, and the compute efficiency is remarkable. Training in 4 hours at half the Enformer compute, while outperforming Enformer on most tasks, suggests AlphaGenome isn’t brute-forcing better results — it’s a more efficient model architecture. That matters for sustainability of ongoing regulatory genomics research.

Variant-effect prediction is the clinical application vector. Understanding how DNA variants affect gene regulation is core to understanding genetic disease. At 24/26 variant-effect task performance, AlphaGenome is in the range where clinical applications (rare disease diagnosis, drug target identification) become realistic. Watch for clinical genetics partnerships similar to AlphaFold’s pharmaceutical partnerships.

Watch:

• AlphaGenome integration with AlphaFold 3 — the combined protein structure + regulatory genomics stack is the complete “digital biology” pipeline
• Pharmaceutical company adoption for non-coding variant interpretation in drug development
• Whether AlphaGenome receives a dedicated web interface analogous to AlphaFold’s public database