A single-cell multimodal foundation model enables cross modality analysis between transcriptome and chromatin
- 1University of Michigan
Abstract
Most single-cell studies measure either gene expression or chromatin accessibility, and connecting the two usually requires fitting a model to each study. Here we present MUTUAL (MUltiome Transfer in a Unified Aligned Latent space), which keeps the single-cell foundation models scFoundation and EpiAgent frozen, combines their embeddings with linear summaries of each modality and learns a shared RNA--ATAC space from 3.0 million paired cells of 400 public multiome datasets. Applied without refitting to ten datasets not used for training, MUTUAL matched cells' RNA and ATAC profiles more accurately than methods fitted on other donors of the same dataset, annotated ATAC-only cells more accurately and transferred transcription-factor motif activity to RNA-only cells beyond what cell-type assignment provides. In a maturation series of stem-cell-derived islets, motif activity predicted from RNA largely reproduced the measured stage-wise changes when other libraries of the series served as the reference. MUTUAL makes public multiome data directly usable by single-modality studies.
Related
- Beyond the Transcriptome: Chromatin-Informed Prediction of Cell-State-Dependent Perturbation Responses
- CUBE: Multimodal Representation Learning Reveals Biological Structure Across Histomorphology, Spatial Protein Phenotypes, and Transcriptome-Associated Signals
- Reversible chromatin remodeling enables Prosopis cineraria survival under recurrent heat extremes.
- SMCHD1's DNA binding activity enables its stable retention on chromatin
- DAPE-PD: Diffusion adaptation of genomic foundation model enables tailored-generation of cis-regulatory DNA
BibTeX
@misc{bao2026singlecell,
title = {{A single-cell multimodal foundation model enables cross modality analysis between transcriptome and chromatin}},
author = {X. Bao and J. Zhu and X. Song and X. Zhang and H. Zeng and H. T. H. Vu and H. Zhang and Z. Zhang and M. Yao and M. Liang and K. Liu and S. C. Parker and S. Chen and J. Liu},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.28.762297},
url = {https://www.biorxiv.org/content/10.64898/2026.10.28.762297v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - A single-cell multimodal foundation model enables cross modality analysis between transcriptome and chromatin
AU - X. Bao
AU - J. Zhu
AU - X. Song
AU - X. Zhang
AU - H. Zeng
AU - H. T. H. Vu
AU - H. Zhang
AU - Z. Zhang
AU - M. Yao
AU - M. Liang
AU - K. Liu
AU - S. C. Parker
AU - S. Chen
AU - J. Liu
PY - 2026
DA - 2026/10/28
PB - bioRxiv
DO - 10.64898/2026.10.28.762297
UR - https://www.biorxiv.org/content/10.64898/2026.10.28.762297v1
AB - Most single-cell studies measure either gene expression or chromatin accessibility, and connecting the two usually requires fitting a model to each study. Here we present MUTUAL (MUltiome Transfer in a Unified Aligned Latent space), which keeps the single-cell foundation models scFoundation and EpiAgent frozen, combines their embeddings with linear summaries of each modality and learns a shared RNA--ATAC space from 3.0 million paired cells of 400 public multiome datasets. Applied without refitting to ten datasets not used for training, MUTUAL matched cells' RNA and ATAC profiles more accurately than methods fitted on other donors of the same dataset, annotated ATAC-only cells more accurately and transferred transcription-factor motif activity to RNA-only cells beyond what cell-type assignment provides. In a maturation series of stem-cell-derived islets, motif activity predicted from RNA largely reproduced the measured stage-wise changes when other libraries of the series served as the reference. MUTUAL makes public multiome data directly usable by single-modality studies.
ER -