How AlphaGenome Predicts Regulatory DNA Variants with 1‑bp Precision

Overview

AlphaGenome is an artificial‑intelligence model that accepts DNA sequences up to 1 megabase (≈entire yeast genome) and predicts, at single‑base resolution, how individual variants or mutations affect a broad spectrum of regulatory biological processes.

Key Capabilities

Input: ultra‑long 1 Mb sequences.

Output: 5,930 human and 1,128 mouse functional tracks at single‑base resolution.

Coverage: eleven regulatory modalities, including gene expression, splicing, chromatin accessibility, histone modifications, and 3D contact maps.

Performance: achieves state‑of‑the‑art results on 25 of 26 variant‑effect benchmarks.

Model Architecture

Figure 1a: U‑Net‑Transformer hybrid backbone; the 1 Mb sequence is split into eight parallel segments processed across eight TPUv3 devices.

Key components:

Encoder : four‑stage down‑sampling from 1 bp to 128 bp.

Transformer Tower : models enhancer‑promoter long‑range interactions at 128 bp resolution.

Decoder : up‑samples back to 1 bp with skip‑connections to preserve fine‑grained detail.

2D Pairwise Branch : generates an additional 2 kb‑resolution chromatin contact map.

Training Strategy

Pre‑training : four‑fold cross‑validation with training/validation splits defined by genomic intervals to prevent information leakage.

Distillation : an ensemble of four teacher models is distilled into a single student model; input sequences are randomly mutated and reverse‑complemented to improve robustness.

Benchmark Performance

Splicing variant (ClinVar deep intron) : +3 % auPRC over the best baseline.

Expression QTL (GTEx eQTL direction prediction) : +25.5 % auROC.

Chromatin accessibility (caQTL causal inference) : +8 % average precision.

3D contact map (Micro‑C) : +42 % cell‑specific correlation.

Real‑World Case Study ①: TAL1 Oncogenic Enhancer Mutation

Post‑mutation H3K27ac and H3K4me1 signals increase.

TAL1 expression rises downstream of the mutation (≈7.5 kb).

In‑silico mutagenesis reveals newly created MYB motifs, matching experimental observations.

Real‑World Case Study ②: Splicing Mutation “One Variant, Three Effects”

The same variant can affect:

Splice‑site strength.

Competitive usage of the site.

Specific splice‑junction counts.

AlphaGenome outputs three separate scores; a composite score yields:

GTEx rare splicing abnormal samples: auPRC 0.66.

MFASS experimental validation: performance surpasses SpliceAI and DeltaSplice.

Ablation Experiments

Resolution : training at 1 bp resolution significantly outperforms 32 bp or 128 bp resolution for splicing and accessibility tasks.

Sequence length : training on 1 Mb versus 32 kb improves eQTL sign‑prediction accuracy by +12 %.

Distillation : distilling from 64 teachers yields a student model whose performance matches a 4‑model ensemble while inference is ~4× faster.

Multimodal training : joint training on all functional tracks outperforms single‑modality models, with the largest gains on eQTL tasks.

References

https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/

Advancing regulatory variant effect prediction with AlphaGenome

https://www.nature.com/articles/s41586-025-10014-0