S3RL: Enhancing Spatial Single-Cell Transcriptomics With Separable Representation Learning

Spatial transcriptomics enables in situ mapping of gene expression, offering insights into tissue organization and cell-cell interactions. However, its utility is limited by data sparsity and technical noise for decoding complex tissue microenvironments. Here, we introduce S3RL, a separable representation learning framework designed to enhance the fidelity of raw spatial transcriptomic data. By effectively denoising sparse measurements and amplifying biologically relevant signals, S3RL enables the recovery of fine-grained spatial expression patterns and regulatory relationships that are otherwise lost. Applied across diverse human, mouse and plant tissues, S3RL not only achieved improved accuracy in spatial domain identification and multi-slice alignment (up to 170% ARI improvement), but also uncovered previously unrecognized ligand-receptor signaling and spatial gene expression gradients that are critical for understanding immune-tumor crosstalk and plant developmental trajectories. These results establish S3RL as a powerful tool for extracting latent biological programs from noisy spatial transcriptomic datasets, paving the way for deeper exploration of tissue biology and disease mechanisms.