AlphaGenome 2 Predicts How DNA Sequence Variations Impact Cell Types and Disease

AlphaGenome 2, the genomic AI model developed by Google DeepMind, predicts how single-letter changes in human DNA alter gene expression across specific cell types. Building on earlier genomic sequence models like Enformer and AlphaFold 3, AlphaGenome 2 maps non-coding genetic variants directly to downstream functional consequences in tissues ranging from brain neurons to immune cells. The update allows researchers to screen millions of non-coding mutations computationally before conducting in vitro experiments.

What's new

AlphaGenome 2 expands prediction scope beyond protein-coding regions to target the 98% of human DNA that regulates gene activity. The model processes long-range genomic contexts up to hundreds of kilobases, mapping how distant enhancer regions interact with gene promoters in distinct cell lineages.

Google DeepMind trained AlphaGenome 2 on unified multi-omic datasets, enabling simultaneous predictions for chromatin accessibility, histone modifications, and RNA transcription levels. The model evaluates genetic variant impacts across thousands of distinct human cell types and tissue contexts, delivering higher accuracy on rare mutation scoring than previous architectures.

How it fits your workflow

For computational biologists and geneticists, AlphaGenome 2 acts as a preliminary screening filter for genome-wide association studies (GWAS). Instead of testing every candidate single nucleotide polymorphism (SNP) in physical assays, researchers can query AlphaGenome 2 to rank variants by predicted impact on specific disease-relevant cells, such as pancreatic beta cells in diabetes studies or microglial cells in neurodegenerative research.

Compared to broad protein structure predictors like AlphaFold 3 or Meta's ESMFold, AlphaGenome 2 specifically targets transcriptional regulation and non-coding sequence function. It replaces multi-step pipelines that combined separate tools for splice prediction, enhancer mapping, and transcription factor binding, offering a single architecture for variant interpretation.

What it costs / how to try it

Google DeepMind provides research access to AlphaGenome 2 through its research site and code repositories. Access details and model weights for non-commercial academic research are available directly via Google DeepMind's release documentation.