Liang, Q.; Lyu, Q. R.

Most single-cell generative models rely on highly variable genes (HVGs) or low-dimensional latent representations, limiting their capacity to capture the complexity of full-gene features. We present scJET, a patch-based Transformer denoising framework that operates in full-gene space. scJET preserves global manifold structure, local neighborhood statistics, and gene-level expression programs. By combining scalable patch tokenization with full-gene denoising, scJET provides an efficient framework for transcriptome-wide single-cell matrix generation.