3 days ago
AlphaGenome: New Google AI reads DNA mutations, predicts molecular consequences
In a big leap for genomics, Google on Wednesday unveiled a powerful AI model that predicts how single DNA mutations affect the complex machinery regulating gene activity.
Named AlphaGenome, the tool covers both coding and non-coding regions of the genome, offering a unified view of variant effects like never before.
It brings base-resolution insight to long-range genomic analysis, decoding the impact of mutations with speed, scale, and unprecedented depth.
The model processes up to 1 million base pairs in a single pass and predicts thousands of molecular properties, including gene expression, splicing patterns, protein-binding sites, and chromatin accessibility across diverse cell types.
It's the first time such a wide range of regulatory features can be modeled jointly using one AI system.
AlphaGenome's architecture first uses convolutional layers to spot short patterns in the DNA sequence, then applies transformers to share information across the entire stretch of genetic code. A final set of layers converts these learned patterns into predictions across various genomic features.
During training, all computations for a single sequence are distributed across multiple interconnected Tensor Processing Units (TPUs), enabling efficient large-scale processing.
A single model was trained in just four hours, using half the compute budget required for its predecessor, Enformer.
Built as a successor to Enformer and complementing AlphaMissense, AlphaGenome is the only model that can jointly predict all evaluated molecular modalities, outperforming or matching specialized models in 24 of 26 benchmark tests.
It was trained on massive public datasets including ENCODE, GTEx, 4D Nucleome, and FANTOM5.
Unlike earlier models that traded sequence length for resolution, AlphaGenome handles both with precision. It captures long-range genomic context and offers base-level predictions, unlocking insights across disease biology, rare variant research, synthetic DNA design, and more.
One standout feature of the new model is its variant scoring system, which efficiently contrasts mutated and unmutated DNA to assess impact across modalities.
It also features splice-junction modeling, a first-of-its-kind approach to predicting RNA splicing disruptions tied to diseases like cystic fibrosis and spinal muscular atrophy.
In synthetic biology, AlphaGenome could help design regulatory sequences that activate genes selectively, for instance, in nerve cells but not muscle cells.
The model could also prove useful in studying rare variants with large biological effects, such as those responsible for Mendelian disorders.
In a test case, AlphaGenome accurately predicted how a leukemia-linked mutation introduces a MYB DNA binding motif that activates the TAL1 gene, mirroring known mechanisms in T-cell acute lymphoblastic leukemia and showcasing its power to connect non-coding variants to disease genes.
While AlphaGenome marks a major advance, it's not designed or validated for personal genome interpretation or clinical use. It also faces challenges in modeling very distant regulatory interactions — especially those over 100,000 DNA letters away — and in fully capturing cell- and tissue-specific patterns.
Still, researchers say it lays a strong foundation for future expansion, with potential to be adapted for additional species, modalities, and lab-specific datasets.
AlphaGenome is now available in preview for non-commercial use via the AlphaGenome API. Google is inviting researchers worldwide to explore use cases, ask questions, and share feedback. The AI-powered tool's predictions are intended strictly for research purposes.
'We hope AlphaGenome will help deepen our understanding of the complex cellular processes encoded in DNA,' Google said, 'and drive new discoveries in genomics and healthcare.'
