Google DeepMind has released AlphaGenome 2, a DNA mannequin that predicts variant results in particular cell sorts. AlphaGenome 2 is a conditional sequence-to-function mannequin. It reads 1 Mb of DNA plus a goal cell state. Its predecessor averaged throughout bulk tissues. The brand new mannequin resolves carefully associated cell sorts and adapts to unseen single-cell datasets with out fine-tuning. It’s out there by the AlphaGenome API for non-commercial analysis.
TL;DR
- Measurement: Enter is a 1 Mb (1,048,576 bp) DNA window, human or mouse.
- Runs on: Hosted API just for now. Coaching used 2048 TPU v4 chips; weights come “upon final publication”.
- Efficiency: Beats fine-tuned AlphaGenome 1, Borzoi and Scooby on single-cell eQTL and enhancer-gene benchmarks.
- Greatest: 43% recall at 90% signal accuracy on TenK10K sc-eQTLs, versus 23% for AlphaGenome 1 fine-tuned.
- Worst: Outdoors accessible chromatin, signal prediction AUROC reaches solely 0.63.
- Backside line: Greatest: cell-state conditioning from uncooked scRNA-seq. Worst: no native weights, non-commercial solely.
What’s AlphaGenome 2?
AlphaGenome 2 is a U-Web model mannequin with a convolutional encoder, an 11-block transformer and a decoder. AlphaGenome 1 predicted a hard and fast catalog of 1000’s of tracks in parallel. AlphaGenome 2 as a substitute predicts outputs for 1 situation per name.
The situation is a 256-dimensional embedding. It modulates activations by FiLM layers within the encoder, transformer and decoder. Conditioning solely the decoder or output layers gave minimal good points. Conditioning the encoder and transformer drove a lot of the enchancment.
Outputs cowl RNA protection, CAGE, splicing, chromatin accessibility, histone marks, TF binding and phone maps. CpG methylation is new.
How was AlphaGenome 2 skilled?
Coaching runs in 4 phases:
- Pre-training throughout bulk information plus 3,977 single-cell RNA and ATAC pseudobulk tracks from 36 research.
- Situation refinement, which tunes solely the embeddings and FiLM layers.
- Distillation of 8 instructor fashions into 1 pupil, utilizing mutated enter sequences.
- Submit-training that maps a single-cell expression encoder into the frozen mannequin’s situation house.
That encoder was skilled on over 100 million human scRNA-seq profiles, largely from CELLxGENE. Researchers can move uncooked counts and get a situation embedding. The docs permit as much as 256 cells per name.
How does AlphaGenome 2 carry out on benchmarks?
On GTEx tissues, median Pearson r for cross-tissue expression adjustments rose from 0.61 to 0.71. Inside tissues, defined cell-type variance rose from 23% to 34% for scRNA-seq.
The primary result’s on TenK10K single-cell eQTLs throughout 28 immune cell sorts. At 90% precision, AlphaGenome 2 recovered 30% of causal variants. AlphaGenome 1 fine-tuned on single-cell information recovered 16%. Course-of-effect recall rose from 23% to 43%.
On held-out ENCODE-rE2G CRISPRi information, enhancer-gene AUPRC reached 0.78. AlphaGenome 1 scored 0.73, a Borzoi ensemble 0.61. The hole 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.
window.addEventListener(‘message’,operate(e){if(e.information&&e.information.ag2h){var f=doc.getElementById(‘ag2-frame’);if(f)f.model.top=e.information.ag2h+’px’;}});
How does it examine with different DNA sequence fashions?
| Function | AlphaGenome 2 | AlphaGenome 1 | Borzoi | Scooby |
|---|---|---|---|---|
| Enter DNA | 1 Mb | 1 Mb | 524 kb | 524 kb |
| Situation dealing with | Reference or scRNA-seq embedding | Fastened multitask tracks | Fastened multitask tracks | Cell-embedding decoder, per dataset |
| Single-cell outputs | scRNA, scATAC | No | No | scRNA, scATAC |
| Parameters | Not disclosed | Not disclosed | 235M | Not disclosed |
| {Hardware} | Hosted API | H100 beneficial | Not disclosed | NVIDIA GPU (examined on A40) |
| Weights / license | API now, weights later, non-commercial | Kaggle and HF, non-commercial phrases | Apache-2.0 | MIT |
| TenK10K sc-eQTL causality AUPRC* | 0.76 | 0.72 (fine-tuned) | Not evaluated | 0.69 |
| TenK10K sc-eQTL signal AUC* | 0.85 | 0.79 (fine-tuned) | Not evaluated | 0.77 |
| rE2G held-out CRISPRi AUPRC* | 0.78 | 0.73 | 0.61 (4-model ensemble) | Not evaluated |
*Benchmark rows come from the AlphaGenome 2 paper and are DeepMind’s personal analysis.
What can builders entry immediately?
AlphaGenome 2 and Colab tutorials are dwell by the API. An API key is free for non-commercial use. Industrial analysis is in non-public preview. The alphagenome Python package deal exposes the mannequin below alphagenome.fashions.v2.
DeepMind additionally launched precomputed atlases. One covers enhancer-gene hyperlinks throughout 640 main cell situations. One other scores about 7 million widespread and GWAS variants throughout 510 situations. Each are on the downloads page.
Key Takeaways
- AlphaGenome 2 situations a 1 Mb DNA mannequin on 1 cell kind per name.
- Causal sc-eQTL recall at 90% precision rose from 16% to 30%.
- Uncooked scRNA-seq counts can outline new cell states with out retraining.
- Precomputed scores cowl about 7 million variants throughout 510 situations.
- Entry is API-only and non-commercial for now.
Try the paper, the DeepMind blog, the API docs and the GitHub repo. Additionally, be at liberty to comply with us on Twitter and don’t overlook to affix our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Have to companion with us for selling your GitHub Repo OR Hugging Face Web page OR Product Launch OR Webinar and so on.? Connect with us
The publish Google DeepMind Releases AlphaGenome 2 for Cell-Type-Specific Variant Effects appeared first on MarkTechPost.

