What changes compared with AlphaGenome 1
Google DeepMind presented AlphaGenome 2, a DNA sequence model that predicts variant effects in specific cell types, according to a report by MarkTechPost. Unlike its predecessor, which according to the source averaged signals across bulk tissues, the new model is conditional: it takes as input a 1 Mb window of human or murine DNA together with a reference cell state, and produces predictions for a single condition per call.
The architecture is U-Net-like, with a convolutional encoder, an 11-block transformer, and a decoder. The condition is a 256-dimensional embedding that modulates activations via FiLM layers. According to the source, conditioning only the decoder or the output layers yields minimal improvements, while most of the gain comes from conditioning the encoder and transformer. The outputs cover RNA coverage, CAGE, splicing, chromatin accessibility, histone modifications, transcription factor binding, and contact maps; CpG methylation is new compared with the previous model.
The benchmark numbers, all internal to DeepMind
The source specifies that the benchmark rows come from the AlphaGenome 2 paper and constitute an evaluation by DeepMind itself. On GTEx tissues, the median Pearson correlation for expression changes across tissues rises from 0.61 to 0.71; the cell-type variance explained within tissues grows from 23% to 34% for single-cell RNA. The main result concerns the TenK10K single-cell eQTLs across 28 immune cell types: at 90% precision the model recovers 30% of causal variants, versus 16% for fine-tuned AlphaGenome 1, while recall on effect direction rises from 23% to 43%.
On CRISPRi ENCODE-rE2G data unseen in training, the enhancer-gene AUPRC reaches 0.78, versus 0.73 for AlphaGenome 1 and 0.61 for a Borzoi ensemble; the gap is wider for enhancers 100 to 2,500 kb away, 0.62 versus 0.48. In assay imputation, the model beats the Avocado baseline on 166 of 181 unseen tracks and, on an unseen Parkinson's cohort, reaches 41% recall versus 17% for Scooby tuned on the cohort. The source also flags the worst limitation: outside accessible chromatin, the AUROC for sign prediction stops at 0.63.
Access, training, and what is missing
Training took place in four stages: pre-training on bulk data and on 3,977 pseudobulk single-cell RNA and ATAC tracks from 36 studies; refinement of the condition on the embeddings and FiLM layers only; distillation of 8 teacher models into a single student; post-training that maps a single-cell expression encoder into the condition space of the frozen model. The encoder was trained on more than 100 million human scRNA-seq profiles, largely from CELLxGENE, and the documents allow up to 256 cells per call. The source indicates the use of 2,048 TPU v4 chips.
