Advancing regulatory variant effect prediction with AlphaGenome
Natasha S. Latysheva, Natasha S. Latysheva, Jun Cheng, Guido Novati, Kyle R. Taylor, Tom Ward et al. · 2026 · 242 citationsRead the paper
Abstract Deep learning models that predict functional genomic measurements from DNA sequences are powerful tools for deciphering the genetic regulatory code. Existing methods involve a trade-off between input sequence length and prediction resolution, thereby limiting their modality scope and performance 1–5 . We present AlphaGenome, a unified DNA sequence model, which takes as input 1 Mb of DNA sequence and predicts thousands of functional genomic tracks up to single-base-pair resolution across diverse modalities. The modalities include gene expression, transcription initiation, chromatin accessibility, histone modifications, transcription factor binding, chromatin contact maps, splice site usage and splice junction coordinates and strength. Trained on human and mouse genomes, AlphaGenome matches or exceeds the strongest available external models in 25 of 26 evaluations of variant effect prediction. The ability of AlphaGenome to simultaneously score variant effects across all modalities accurately recapitulates the mechanisms of clinically relevant variants near the TAL1 oncogene 6 . To facilitate broader use, we provide tools for making genome track and variant effect predictions from sequence.
What the full text shows
Working prototype- Matched or outperformed external SOTA models in variant effect prediction across most benchmarks. — 25 of 26 evaluations (Using the distilled student model)
- Improved prediction of cell-type-specific gene-level expression log-fold change compared to Borzoi. — +14.7% relative improvement (Out-of-fold evaluations)
- Improved sign prediction for expression QTLs (eQTLs) compared to Borzoi. — +25.5% (GTEx eQTL benchmarks)
- Improved sign prediction for accessibility QTLs compared to ChromBPNet. — +8.0% (Averaged across five datasets)
- Increased recovery of GTEx eQTLs at a 90% sign prediction accuracy threshold compared to Borzoi. — 41% vs 19% (GTEx eQTLs)
- Outperformed Orca on contact map Pearson r and cell-type-specific differences. — +6.3% Pearson r; +42.3% cell-type-specific differences (Out-of-fold evaluations)
- Setup
- Trained on human (5,930 tracks) and mouse (1,128 tracks) genomes using 1 Mb input sequences; evaluated on NVIDIA H100 GPU.
- How
- A U-Net-inspired backbone architecture using convolutional layers for local patterns and transformer blocks for long-range dependencies. The model was trained via a two-stage process: 4-fold cross-validation pretraining followed by distillation into a single student model.
- Limitations
- Difficulty capturing influence of distal regulatory elements beyond 100 kb. · Challenges in precisely predicting tissue-specific nuances and condition-specific variant effects. · Training and evaluation heavily focused on protein-coding genes, with limited coverage of non-coding genes like microRNAs. · Limited species coverage (human and mouse only).
Summarized by Gemma 4 from the open-access full text · 8 Oct 2026. Read it