# Cross-modal contrastive learning decodes developmental regulatory features through chromatin potential analysis

## TL;DR

This paper introduces Attune, a new method that combines gene expression and chromatin data to better understand gene regulation during development.

## Contribution

Attune introduces a cross-modal contrastive learning framework for integrating and analyzing paired gene expression and chromatin accessibility data.

## Key findings

- Attune outperforms existing methods in omics integration and gene expression prediction.
- The method identifies regulatory interactions and differential signals in cell subtypes.
- Attune reveals chromatin potential of Gli3 and transmitted states in cortical neuron differentiation.

## Abstract

Emerging large-scale multimodal single-cell data jointly measure chromatin accessibility and transcription in the same cell, thus reconciling matched data paves an integrated route for comprehensive regulatory analysis.

Here, we introduce Attune, a cross-modal contrastive learning framework to align paired gene expression and accessibility information. Systematic benchmarking shows Attune’s superior performance for omics integration and gene expression prediction. We further introduce transformer-based cross-modal attention over fine-tuned gene and peak embeddings to infer regulatory interaction and discover significant differential signals of cell subtypes. Applied to a hair follicle maturation dataset, Attune reveals chromatin potential for the bifunctional transcription factor Gli3 at the gene level. In addition, the paired representations determine transmitted states across neonatal and mature cell types of cortical neuron differentiation at the cell level. Taken together, Attune offers an approach for regulatory inference across omics layers and enables more advanced omics analyses.

Attune offers a versatile framework for integrating gene expression and chromatin accessibility, enabling the inference of regulatory mechanisms and the prediction of gene expression from cross-modal data.