AlphaGenome on Nature Cover: Predicts Variant Effects Across All Modalities in <1 s

In June 2025, DeepMind announced AlphaGenome, a genome‑scale deep‑learning model that can predict the impact of single‑base DNA variants on a wide range of regulatory processes. The model accepts up to 1 Mb of DNA sequence (human or mouse) together with species information and outputs predictions for 5,930 human or 1,128 mouse genomic tracks covering 11 modalities, including RNA‑seq, CAGE, PRO‑cap, detailed splicing patterns, chromatin state (DNase, ATAC‑seq, histone marks, TF binding) and contact maps.

The architecture follows a class‑U‑Net design. One‑dimensional embeddings (1 bp and 128 bp resolution) encode linear genomic sequence for track prediction, while two‑dimensional embeddings (2048 bp resolution) capture spatial interactions for contact‑map prediction. Convolutional layers model local motifs, and Transformer modules capture long‑range dependencies such as enhancer‑promoter contacts. Training runs on eight interconnected TPUv3 devices and processes the full 1 Mb sequence at single‑base resolution.

Training proceeds in two stages. First, a pre‑training phase uses existing experimental data to train two families of models: (1) fold‑specific models trained with four‑fold cross‑validation on 3/4 of the reference genome and evaluated on the held‑out 1/4, and (2) all‑folds models trained on the entire reference genome to serve as teachers. Second, a distillation phase trains a shared‑architecture student model that, given randomly augmented inputs, learns to reproduce the aggregated outputs of the teacher ensemble. This design enables the student to predict all modalities and cell‑type‑specific variant effects in a single inference call.

On an NVIDIA H100 GPU, the student model scores each variant in under one second, dramatically speeding up large‑scale variant‑effect prediction compared with traditional multi‑model ensembles.

Benchmarking on 24 genomic‑track tasks shows AlphaGenome outperforms existing methods on 22 of them. Notably, it surpasses the multimodal model Borzoi by +17.4 % on cell‑type‑specific gene‑expression (log‑fold‑change) prediction, exceeds Orca by +6.3 % Pearson correlation on contact‑map prediction (and +42.3 % on cell‑type‑specific differences), beats ProCapNet by +15 % Pearson on transcription‑start‑site trajectories, and improves chromatin‑accessibility prediction over ChromBPNet by +8 % (ATAC‑seq) and +19 % (DNase‑seq). AlphaGenome is the only model that jointly predicts all evaluated modalities.

Experts have praised the model as a milestone. Jun Cheng (co‑first author) highlighted its ability to predict splicing junctions directly from sequence for variant‑effect analysis. Caleb Lareau (MSKCC) called it “the first model with long context, single‑base precision, and top‑tier performance across a broad set of genomic tasks.” Pushmeet Kohli (DeepMind VP) noted that AlphaGenome offers a comprehensive view of the non‑coding genome, deepening disease‑biology understanding. Marc Mansour (UCL) emphasized its utility for large‑scale non‑coding variant discovery in cancer.

AlphaGenome is released for non‑commercial research use, with the expectation that the academic community will build upon it for disease‑mechanism studies, synthetic‑biology design of cell‑type‑specific regulatory elements, and fundamental genomics investigations.