Advancing regulatory variant effect prediction with AlphaGenome

N/A (pan-genomic prediction model)

AlphaGenome: unified sequence-to-function model, 2 Mb DNA sequence input → simultaneous predictions across 50+ functional modalities (ATAC-seq, histone ChIP-seq, RNA-seq, splicing, methylation); multi-resolution transformer architecture; 10× more precise than Enformer on pathogenic regulatory variants

AlphaGenome is a deep learning model (DeepMind) unifying functional effect prediction of regulatory variants across 50+ genomic modalities from a 2 Mb DNA sequence. The multi-resolution transformer architecture enables nucleotide-level resolution across the entire locus. AlphaGenome outperforms Enformer and Basenji by a factor of 10 on known pathogenic regulatory variants, and simultaneously predicts effects on chromatin (ATAC, histone ChIP), expression (RNA-seq), splicing (SpliceAI-like) and methylation.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

AlphaGenome represents the absolute state-of-the-art for non-coding variants, a major weakness of classical exome and WGS panels. Its integration into secondary annotation pipelines would allow scoring of variants in regulatory regions, enhancers, and deep intronic zones. Available soon via the AlphaFold Server API according to the authors.

Nature (top journal) +3; SOTA model surpassing Enformer/Basenji +2; 50+ unified functional modalities +2; applicable to non-coding variants in clinical setting +2