Predicting cell-state-specific gene regulation from DNA across species
- 1Arc Institute, Palo Alto, CA, USA
Abstract
DNA encodes the cell's phenotype, but the same DNA sequence is interpreted differently in every cell type, developmental stage, and activity state. Existing deep-learning models solve only half of this puzzle: sequence-to-function models read DNA but treat cell state as a fixed label, while single-cell foundation models represent cell state but not sequence. We developed Bolero, a sequence-to-function model whose parameters condition on atlas-scale cell-state embeddings and additional biological variables to predict cell-state-specific chromatin accessibility and transcript abundance from DNA. Trained on 10.7 million cells from 33 datasets across six mammals, Bolero matched or exceeded specialised models and resolved regulatory differences between closely related cell states, including cortical regional and developmental gradients. Extrapolating cell-state embeddings extended Bolero's predictions to cell states that are absent from training. Cell-state conditioning facilitated tracing of the effects of non-coding variants to the cell types in which they act. Using an AP-1 activity score as a further continuous conditioning axis, Bolero predicted activity-regulated chromatin changes without requiring actual stimulation data. Across human cell types, we reconstructed the cooperative grammar of motif identity and spacing underlying predicted AP-1 responses. Bolero further predicted cell-type-specific chromatin accessibility and AP-1 responses in species excluded from training, extending comparative genomics from sequence conservation to predicted functional conservation. Across 84 mammals, we found that predicted AP-1 responses were less conserved than chromatin accessibility at cis-regulatory elements, driven by the human-specific sequence variants on AP-1 and cooperating motifs. By combining reference genomes with reference cell atlases, Bolero allowed genome sequences, either native or mutated, to be queried across cell states, perturbations, and different species for their chromatin and expression signatures.
Related
- Predicting forage yield and water use efficiency using machine learning across diverse forage species under varying irrigation and fertilizer management in the Ethiopian highlands
- ANIMA: predicting protein-protein interactions across species
- Behavioral divergence across replicated predation regimes is species-specific in two sympatric poeciliid fishes
- Contrasting Spatial Genomic Structure in Two Widespread Grassland Species Across Norway Supports Species‐Specific Seed Sourcing
- A survey of bacterial communities reveals host-specific associations across multiple Pentatomomorpha species in France
BibTeX
@misc{liu2026predicting,
title = {{Predicting cell-state-specific gene regulation from DNA across species}},
author = {H. Liu and S. Mao and T. Gallent and Y. Chen and Z. Yan and I. Vogel and L. Traunmuller and R. Zhang and J. D. Buenrostro and J. Zhou and M. E. Greenberg},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.19.760026},
url = {https://www.biorxiv.org/content/10.64898/2026.10.19.760026v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Predicting cell-state-specific gene regulation from DNA across species
AU - H. Liu
AU - S. Mao
AU - T. Gallent
AU - Y. Chen
AU - Z. Yan
AU - I. Vogel
AU - L. Traunmuller
AU - R. Zhang
AU - J. D. Buenrostro
AU - J. Zhou
AU - M. E. Greenberg
PY - 2026
DA - 2026/10/19
PB - bioRxiv
DO - 10.64898/2026.10.19.760026
UR - https://www.biorxiv.org/content/10.64898/2026.10.19.760026v1
AB - DNA encodes the cell's phenotype, but the same DNA sequence is interpreted differently in every cell type, developmental stage, and activity state. Existing deep-learning models solve only half of this puzzle: sequence-to-function models read DNA but treat cell state as a fixed label, while single-cell foundation models represent cell state but not sequence. We developed Bolero, a sequence-to-function model whose parameters condition on atlas-scale cell-state embeddings and additional biological variables to predict cell-state-specific chromatin accessibility and transcript abundance from DNA. Trained on 10.7 million cells from 33 datasets across six mammals, Bolero matched or exceeded specialised models and resolved regulatory differences between closely related cell states, including cortical regional and developmental gradients. Extrapolating cell-state embeddings extended Bolero's predictions to cell states that are absent from training. Cell-state conditioning facilitated tracing of the effects of non-coding variants to the cell types in which they act. Using an AP-1 activity score as a further continuous conditioning axis, Bolero predicted activity-regulated chromatin changes without requiring actual stimulation data. Across human cell types, we reconstructed the cooperative grammar of motif identity and spacing underlying predicted AP-1 responses. Bolero further predicted cell-type-specific chromatin accessibility and AP-1 responses in species excluded from training, extending comparative genomics from sequence conservation to predicted functional conservation. Across 84 mammals, we found that predicted AP-1 responses were less conserved than chromatin accessibility at cis-regulatory elements, driven by the human-specific sequence variants on AP-1 and cooperating motifs. By combining reference genomes with reference cell atlases, Bolero allowed genome sequences, either native or mutated, to be queried across cell states, perturbations, and different species for their chromatin and expression signatures.
ER -