Google DeepMind has released AlphaGenome 2, a DNA model that predicts variant effects in specific cell types. AlphaGenome 2 is a conditional sequence-to-function model. It reads 1 Mb of DNA plus a target cell state. Its predecessor averaged across bulk tissues. The new model resolves closely related cell types and adapts to unseen single-cell datasets without fine-tuning. It is available through the AlphaGenome API for non-commercial research.
TL;DR
- Size: Input is a 1 Mb (1,048,576 bp) DNA window, human or mouse.
- Runs on: Hosted API only for now. Training used 2048 TPU v4 chips; weights come “upon final publication”.
- Performance: Beats fine-tuned AlphaGenome 1, Borzoi and Scooby on single-cell eQTL and enhancer-gene benchmarks.
- Best: 43% recall at 90% sign accuracy on TenK10K sc-eQTLs, versus 23% for AlphaGenome 1 fine-tuned.
- Worst: Outside accessible chromatin, sign prediction AUROC reaches only 0.63.
- Bottom line: Best: cell-state conditioning from raw scRNA-seq. Worst: no local weights, non-commercial only.
What is AlphaGenome 2?
AlphaGenome 2 is a U-Net style model with a convolutional encoder, an 11-block transformer and a decoder. AlphaGenome 1 predicted a fixed catalog of thousands of tracks in parallel. AlphaGenome 2 instead predicts outputs for 1 condition per call.
The condition is a 256-dimensional embedding. It modulates activations through FiLM layers in the encoder, transformer and decoder. Conditioning only the decoder or output layers gave minimal gains. Conditioning the encoder and transformer drove most of the improvement.
Outputs cover RNA coverage, CAGE, splicing, chromatin accessibility, histone marks, TF binding and contact maps. CpG methylation is new.
How was AlphaGenome 2 trained?
Training runs in 4 stages:
- Pre-training across bulk data plus 3,977 single-cell RNA and ATAC pseudobulk tracks from 36 studies.
- Condition refinement, which tunes only the embeddings and FiLM layers.
- Distillation of 8 teacher models into 1 student, using mutated input sequences.
- Post-training that maps a single-cell expression encoder into the frozen model’s condition space.
That encoder was trained on over 100 million human scRNA-seq profiles, largely from CELLxGENE. Researchers can pass raw counts and get a condition embedding. The docs allow up to 256 cells per call.
How does AlphaGenome 2 perform on benchmarks?
On GTEx tissues, median Pearson r for cross-tissue expression changes rose from 0.61 to 0.71. Within tissues, explained cell-type variance rose from 23% to 34% for scRNA-seq.
The main result is on TenK10K single-cell eQTLs across 28 immune cell types. At 90% precision, AlphaGenome 2 recovered 30% of causal variants. AlphaGenome 1 fine-tuned on single-cell data recovered 16%. Direction-of-effect recall rose from 23% to 43%.
On held-out ENCODE-rE2G CRISPRi data, enhancer-gene AUPRC reached 0.78. AlphaGenome 1 scored 0.73, a Borzoi ensemble 0.61. The gap was widest for enhancers 100 to 2,500 kb away: 0.62 versus 0.48.
For assay imputation, AlphaGenome 2 beat the Avocado baseline on 166 of 181 held-out tracks. On an unseen Parkinson’s cohort, it reached 41% recall versus 17% for cohort-tuned Scooby.