AlphaGenome 2 Improves Single-Cell eQTL Reproducibility by Approximately 2x

Google DeepMind released AlphaGenome 2 on November 11, 2026, which predicts gene regulation and variant effects by taking 1Mb DNA sequences and specific cell conditions as input
In the TenK10K immune cell single-cell eQTL evaluation, the causal variant reproducibility at 90% precision improved from 16% in AlphaGenome 1 to 30%, an approximately 2x increase
Transitioned from a parallel output structure to a conditional prediction model, applying 256-dimensional embeddings and FiLM techniques to more precisely capture subtle differences between cell types
Provides a free API v2.0.0 for non-commercial research and a pre-computed atlas of over 6.8 million GWAS variants; it is unvalidated for clinical diagnostic use

Technical Changes

Adopted a conditional model structure that takes 1Mb DNA sequences and one biological condition as input, performing predictions only for that specific condition
Applied FiLM technique to inject condition vectors throughout the network and changed batch normalization to RMS/layer normalization
Implemented in JAX and Haiku, optimizing performance through a 4-stage training procedure from pre-training to distillation and integration with a single-cell autoencoder

Pearson correlation for inter-tissue expression differences in GTEx improved from 0.61 to 0.71, and the proportion of variance explained for inter-cell type expression differences improved from 23% to 34%
In cross-modality imputation for predicting unmeasured experimental data, it outperformed existing methods by 1.7 to 2.9 times in AUPRC
API access is restricted to non-commercial organizations only, and model source code and weights are scheduled to be released after the formal publication of the paper