Title: Scalable Single-Cell Gene Expression Generation with Latent Diffusion Models

Abstract:

Accurately modeling single-cell gene expression is essential for deciphering cellular mechanisms, yet producing realistic expression profiles remains a significant hurdle. This challenge is largely driven by the discrete, count-based nature of the data and the intricate latent dependencies that exist between genes. Current generative approaches frequently suffer from limitations, such as imposing artificial gene orderings or depending on shallow neural network structures. To address these issues, we present scLDM, a scalable latent diffusion model designed specifically for single-cell gene expression data that honors the data’s fundamental exchangeability property. Our approach utilizes a Variational Autoencoder (VAE) with fixed-size latent variables, powered by a unified Multi-head Cross-Attention Block (MCAB). This architecture fulfills two critical functions: it performs permutation-invariant pooling within the encoder and permutation-equivariant unpooling in the decoder. Furthermore, we improve this framework by substituting the traditional Gaussian prior with a latent diffusion model that employs Diffusion Transformers and linear interpolants. This modification facilitates high-quality generation through multi-conditional classifier-free guidance. Experimental results demonstrate that our method outperforms existing techniques across various tasks, including the analysis of both observational and perturbational single-cell data, as well as downstream applications such as cell-level classification.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC