Cell-state-conditioned DNA sequence model that predicts regulatory activity and variant effects from 1 Mb of sequence, even in unseen cell states.
No providers recorded yet. Browse all providers
A variant that raises expression of ETS2 only in gut macrophages looks inert when a model reports one averaged signal for the whole colon. Most disease-associated variants sit in non-coding DNA and act in particular cell types or states, yet earlier sequence-to-function models predicted fixed sets of bulk tracks in parallel, and distinguished closely related cell states poorly. AlphaGenome 2 (AG2), from Google DeepMind, links single-cell atlases to variant interpretation.
The key change from AlphaGenome is conditioning. AG2 takes a 1-Mb human or mouse DNA sequence together with a condition: a tissue, cell line, cell type or cell state. It predicts regulatory readouts for that one condition. The condition comes either from a learned dictionary of ontology-defined reference conditions or from a cell encoder that turns a scRNA-seq profile into an embedding. The embedding modulates sequence representations from the earliest layers onward, so the same DNA can be read differently in each cell type. Conditioning on expression extends AG2 to newly profiled cohorts and unseen cell states without fine-tuning.
AG2 was announced on 11 November 2026 and described in a bioRxiv preprint posted 17 November 2026 that has not yet been peer reviewed.
AG2 keeps a U-Net-style backbone similar to AlphaGenome 1. A convolutional encoder, a central transformer tower and a decoder integrate proximal and distal context across the 1-Mb window, and 14 modality heads produce the outputs. Training ran in four stages. Eight randomly initialized models were pre-trained across reference conditions. Refining their condition embeddings and modulation layers yielded eight teachers. A single student was distilled from the teacher ensemble on randomly mutated sequences. Finally, the frozen student was post-trained at single-cell resolution by mapping the expression autoencoder's embeddings into its condition space. Beyond AlphaGenome 1's bulk corpus (ENCODE, GTEx, 4D Nucleome, FANTOM5), AG2 adds 3,977 scRNA-seq and scATAC-seq pseudobulk tracks from 36 human and mouse studies and cell-type-resolved CpG methylation across 185 human conditions. Post-training sampled from 107 million human scRNA-seq profiles, including CZ CELLxGENE.
The main comparison baseline is AG1-sc, AlphaGenome 1 fine-tuned on the same single-cell reference conditions. On TenK10K immune sc-eQTLs, AG2 raised recall at 90% precision from 16% to 30% for causality and from 23% to 43% for sign (AUPRC 0.76 vs 0.72 for AG1-sc and 0.69 for Scooby). On GTEx eQTLs, recall at 90% precision rose from 17% to 26% over AlphaGenome 1 (AUPRC 0.74 vs 0.72, and 0.69 for Borzoi). On the held-out ENCODE-rE2G CRISPRi benchmark, gradient-based enhancer–gene scores reached an AUPRC of 0.78, against 0.73 for AlphaGenome 1, 0.61 for a Borzoi ensemble and 0.57 for ENCODE-rE2G. AG2's advantage was largest at 100–2,500 kb from the TSS. On a Parkinson's disease cohort, expression conditioning outperformed CREsted and cohort-fine-tuned Scooby. The paper does not report a parameter count.
AG2 is aimed at statistical geneticists and single-cell biologists who need to connect a GWAS locus to a causal variant, target gene, direction of effect and cell state. The paper's examples include recovering the ETS2 enhancer variant in gut macrophages and a severe-COVID-19 variant that acts mainly in vaccine-activated T cells. The model is available for non-commercial research through the AlphaGenome API and Python client, with Colab tutorials for expression conditioning. Commercial access is in private preview on Google Cloud. Source code and weights have not been released; the preprint states they will be provided on final publication.
AG2 shows that conditioning on expression can replace per-dataset fine-tuning of sequence models. Its limitations are stated plainly by the authors. At high precision it recovers only a minority of fine-mapped eQTLs and causal GWAS variants. It underestimates effects outside accessible chromatin, at very distal elements and through post-transcriptional mechanisms. Transient, stimulated and disease states are underrepresented, and fine-grained states such as neuronal subtypes remain hard to separate. Because condition embeddings are learned from steady-state observational data, its ability to predict responses to knockouts or state transitions is untested. It is not validated for clinical or diagnostic use.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.