Google DeepMind releases AlphaGenome 2 for cell-specific variant analysis
Google DeepMind has released AlphaGenome 2, a conditional DNA sequence-to-function model designed to predict regulatory activity and genetic variant effects for particular cell states. It processes a 1 Mb (1,048,576 bp) human or mouse DNA window and one condition per call. The condition is a 256-dimensional embedding that can be derived from raw single-cell RNA sequencing counts, allowing the system to represent previously unseen cell states without fine-tuning.
The model uses a convolutional encoder, an 11-block transformer and a decoder, with conditioning applied through FiLM layers. Its outputs include RNA coverage, CAGE, splicing, chromatin accessibility, histone marks, transcription-factor binding, contact maps and, newly, CpG methylation. Training incorporated bulk data and 3,977 single-cell RNA and ATAC pseudobulk tracks from 36 studies, followed by refinement, distillation and post-training with an expression encoder trained on more than 100 million human scRNA-seq profiles.
In DeepMind’s evaluation, AlphaGenome 2 raised causal-variant recall at 90% precision on TenK10K sc-eQTLs from 16% for fine-tuned AlphaGenome 1 to 30%, while direction-of-effect recall increased from 23% to 43%. It also achieved 0.78 enhancer-gene AUPRC on held-out ENCODE-rE2G CRISPRi data. The model is currently available through a hosted AlphaGenome API for non-commercial research; weights are planned after final publication.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.