Mitigating autocorrelation during spatially resolved transcriptomics data analysis

Several computational methods have recently been developed for characterizing molecular tissue regions in spatially resolved transcriptomics (SRT) data. However, each method fundamentally relies on spatially smoothing transcriptomic features across neighboring cells. Here, we demonstrate that smoothing increases autocorrelation between neighboring cells, causing latent space to encode physical adjacency rather than spatial transcriptomic patterns. We find that randomly sub-sampling neighbors before smoothing mitigates autocorrelation, improving the performance of existing methods and further enabling a simpler, more efficient approach that we call spatial integration (SPIN). SPIN leverages the conventional single-cell toolkit, yielding spatial analogies to each tool: clustering identifies molecular tissue regions; differentially expressed gene analysis calculates region marker genes; trajectory inference reveals continuous, molecularly defined ana tomical axes; and integration allows joint analysis across multiple SRT datasets, regardless of tissue morphology, spatial resolution, or experimental technology. We apply SPIN to SRT datasets from mouse and marmoset brains to calculate shared and species-specific region marker genes as well as a molecularly defined neocortical depth axis along which several genes and cell types differ across species.