How AlphaGenome Decodes 98% of the Genome’s Dark Matter

AlphaGenome Overview

AlphaGenome is a DeepMind model that processes up to 1 M DNA base pairs in a single forward pass and predicts the functional impact of genetic variants.

Key Technical Contributions

Full‑stack prediction: Simultaneously outputs gene‑expression, splicing, chromatin accessibility, and protein‑DNA binding profiles for non‑coding regions.

Extended sequence length: Built on the Enformer architecture, the receptive field is enlarged to capture regulatory interactions over hundreds of kilobases while retaining base‑pair‑resolution outputs.

Performance

Variant impact: At ~90 % accuracy AlphaGenome identifies 41 % of known functional variants, compared with 19 % for the previous best model.

Splicing disruption: Ranks first on six of seven splicing benchmark datasets.

Chromatin state: Outperforms specialized tools in predicting changes to DNA packaging.

Cancer‑mutation case study: Accurately predicts activation of the TAL1 oncogene in T‑cell leukemia, matching experimental conclusions.

Model Architecture

AlphaGenome extends the transformer‑based Enformer model. It accepts a 1 M‑nt input window, uses dilated convolutions and attention to aggregate information across >500 kb, and produces per‑base predictions for multiple modalities.

Training Data and Evaluation

Training combines large‑scale functional genomics datasets (e.g., RNA‑seq, ATAC‑seq, ChIP‑seq) covering diverse cell types. Evaluation uses benchmark suites for variant effect prediction, splicing, and chromatin accessibility.

Open‑Source Release

Code, pretrained weights, and documentation are publicly available at https://github.com/google-deepmind/alphagenome_research. The repository includes scripts for data preprocessing, model training, and inference on custom genomic regions.

Usage Example

# Clone the repository
git clone https://github.com/google-deepmind/alphagenome_research.git
cd alphagenome_research

# Install dependencies
pip install -r requirements.txt

# Run inference on a FASTA region (1 M bp)
python predict.py --fasta my_region.fasta --output predictions.tsv

Implications

By providing a unified sequence‑to‑function model for the 98 % non‑coding genome, AlphaGenome enables systematic assessment of rare‑variant effects and supports downstream applications such as disease‑gene prioritization and therapeutic target discovery.