Google has introduced AlphaGenome, an advanced open-weights model that interprets non-coding DNA, which comprises 98 percent of the human and mouse genomes and is crucial for regulating gene expression. By utilizing a combination of convolutional neural networks and transformers, AlphaGenome can predict gene properties and assess the impacts of genetic mutations with remarkable accuracy, surpassing previous models in most evaluations. This innovative tool not only aids in understanding how genetic variations influence diseases, such as T-cell acute lymphoblastic leukemia, but also offers free access to its API, model weights, and inference code for noncommercial research, promoting further exploration in biomedical studies.

Google: Google operates DeepMind, an AI research lab advancing solutions for complex biological problems through machine learning. DeepMind recently released AlphaGenome to decode non-coding DNA functions and mutation effects. This positions Google as a key player in AI-driven genomics research featured in the news.
AlphaGenome: AlphaGenome is Google DeepMind’s open-weights deep learning model that interprets non-coding regions of human and mouse genomes to predict gene properties like transcription start sites, RNA expression, and splicing patterns. It was pretrained on public genomic datasets and distilled from multiple specialist models for broad applicability. The news highlights its role in accelerating biomedical research by modeling mutation impacts on disease mechanisms.

Architecture: Combines convolutional neural network encoder for sequence embeddings, transformer for distant relationships, and CNN decoder for property predictions.
Accessibility: Provides freely licensed API, model weights, and inference code for noncommercial research.
Mutation Prediction: Superiorly models effects of genetic variations, including those linked to diseases like T-cell acute lymphoblastic leukemia.