CoLa-VAE

Cell-cell communication-aware variational autoencoder for representation learning and expression denoising in single-cell transcriptomics.

CoLa-VAE is a cell-cell communication-aware Variational Autoencoder (VAE) that jointly learns low-dimensional latent representations and denoised expression profiles from sparse single-cell RNA sequencing (scRNA-seq) and spatial transcriptomic data.

Single-cell RNA sequencing provides unprecedented resolution into cellular heterogeneity, but severe dropout noise and sparsity hinder downstream biological discovery. CoLa-VAE incorporates ligand-receptor communication topology through dynamic graph Laplacian regularization, establishing a principled communication-guided framework for AI in the life sciences.

Motivation & Background

  • Dropout & Sparsity in scRNA-seq: Single-cell sequencing technologies frequently suffer from technical noise where expressed transcripts fail to be amplified or detected.
  • Vicious Cycle in Communication Analysis: Ligand-receptor based cell-cell communication inference relies on accurate expression levels, but dropout creates substantial false negatives. Standard denoising methods ignore intercellular communication structures during reconstruction.
  • Core Innovation: Rather than treating denoising as an independent precursor to biological analysis, CoLa-VAE dynamically couples representation learning, denoising, and communication network structure into a unified generative model.

Key Ideas

  • Dynamic Graph Laplacian Regularization: Leverages ligand-receptor interactions across cells to build an intercellular communication topology, regularizing both latent representations and expression reconstruction.
  • Iterative Communication Refinement: Uses denoised expression matrices to iteratively update communication estimates, which in turn refine the geometric organization of the latent space.
  • Improved Downstream Biological Discovery:
    • More sensitive and robust identification of differential cell-cell communication programs.
    • Mitigation of technical batch effects without distorting genuine biological variation.
    • Enhanced cell-type deconvolution in spatial transcriptomics when spatially constrained interaction priors are integrated.

Citation

If you use CoLa-VAE in your research, please cite:

@article{Chen2026.03.28.715052.CCC,
  author    = {Chen, Yeqing and Qi, Cong and Fang, Hanzhang and Luan, Feiyang and Zhang, Zhirong and Arya, Shivvrat and Wei, Zhi},
  title     = {CoLa-VAE: A Cell-Cell Communication-Aware Variational Autoencoder for Representation Learning and Expression Denoising},
  journal   = {bioRxiv},
  year      = {2026},
  doi       = {10.64898/2026.03.28.715052},
  publisher = {Cold Spring Harbor Laboratory}
}