Arc Institute's Bolero predicts cell-state-specific gene regulation from DNA alone

anshulkundaje · x · 2026-10-21

Arc Institute introduced Bolero, a sequence-to-function model conditioned on atlas-scale cell-state embeddings that predicts chromatin accessibility and transcript abundance from DNA alone. Trained on 10.7M cells from 33 datasets across six mammals, it matched or beat specialized models and resolved subtle regulatory differences between related cell states. Cell-state conditioning enables extrapolation to unseen states, tracing non-coding variants to the cell types where they act, and predicting activity-regulated chromatin changes via an AP-1 activity axis. Across 84 mammals, predicted AP-1 responses proved less conserved than chromatin accessibility, driven by human-specific variants.

Original post →

More from AGI Musings

AGI Musings channel →