CoLa-VAE: Cell-Cell Communication-aware Variational Autoencoder with Dynamic Graph Laplacian Constraints

Intercellular communication is a fundamental driver of cellular identity and tissue homeostasis, yet current single-cell representation learning frameworks predominantly model cell state as a function of intrinsic gene expression, neglecting the extrinsic signaling context. Conversely, dedicated communication inference tools are often constrained by the sparsity and noise of raw transcriptomic data. Here, we present CoLa-VAE, a deep generative framework that explicitly integrates cell-cell communication (CCC) constraints into latent variable learning. By employing a dynamic graph Laplacian regularization derived from pairwise ligand-receptor interactions, CoLa-VAE disentangles communication-driven topology from intrinsic transcriptional heterogeneity within a variational autoencoder architecture. We demonstrate that CoLa-VAE serves as a robust, method-agnostic framework compatible with diverse signaling definitions, consistently outperforming state-of-the-art baselines in structural clustering metrics and denoising fidelity across heterogeneous sequencing platforms.