DeepMind's AlphaGenome can read a million DNA letters and predict their functions.

In a Nature paper, DeepMind introduces AlphaGenome, a deep learning model trained on human and mouse genomes that can analyze up to 1 megabase of DNA at once and predict thousands of functional genomic signals.

DeepMind researchers published a paper in Nature introducing AlphaGenome, a deep learning model for DNA sequence analysis. The team says the model can analyze up to 1 megabase (about 1 million DNA letters) at once and predict thousands of functional genomic tracks. AlphaGenome was trained on human and mouse genomes and reportedly matched or outperformed other models in 25 of 26 tests measuring variant effects. Outside researchers note the model performs well but that data and validation limits remain. Key details: - AlphaGenome analyzes up to 1 megabase (~1 million DNA letters) per input. - The model was trained on both human and mouse genomes. - It predicts thousands of functional genomic tracks and nearly 6,000 human genetic signals tied to specific functions. - In the paper, AlphaGenome matched or outperformed other existing models in 25 of 26 tests for predicting variant effects. - Independent testing at the Wellcome Sanger Institute using over half a million experiments found strong performance but highlighted limitations in available biological datasets. Summary: Researchers say AlphaGenome may illuminate functions in non-coding regions and aid the study of how genetic variants affect health, including in rare disease and cancer research. Outside scientists emphasize the need for additional validation and improved, standardized datasets before broader application. Undetermined at this time.