A covariance-aware variational generative model preserving gene correlation structure in single-cell transcriptomic data

Measuring gene activity in individual cells reveals the diversity within tumors but is costly, technically noisy, and often captures too few rare cells. Computer-generated cells could fill these gaps. Here we show, across two independent groups of pancreatic cancer patients, that a deep generative model produces realistic synthetic cells. Explicitly modeling correlations within the model’s compressed description of each cell does not improve preservation of gene-gene relationships over a simpler matched model, which appears to capture this structure implicitly. Nevertheless, the model’s sampling strategy yields cells that faithfully reproduce gene activity patterns, cell-type markers, and cellular organization and compare favorably with established methods. The model trains 100 to 1,000 times faster than a diffusion-based alternative. Synthetic cells blend with real data and support standard analyses, including reconstruction of a precancerous transition. These findings support using synthetic cells to enrich rare populations, benchmark methods, and simulate transitional states in tumors.