StoryResearch
DeepMind unveils AlphaGenome, a model that predicts how DNA changes affect gene regulation
Google DeepMind introduced AlphaGenome, which reads up to a million DNA letters and predicts thousands of molecular properties at single-letter resolution, from where genes start and stop to splicing and protein binding. It was offered through an API for non-commercial research.
- Beat the best external models on 22 of 24 sequence tasks and matched or beat them on 24 of 26 variant-effect tasks.
- Trained in about 4 hours on TPUs, using ENCODE, GTEx, 4D Nucleome and FANTOM5 data.
- Stated limits: weak on switches over 100,000 letters away; not meant for personal genome prediction.