Benchmarking Computational Integration Methods for Spatial Transcriptomics Data

The increasing popularity of spatial transcriptomics has allowed researchers to analyze transcriptome data in its tissue samples spatial context. Various methods have been developed for detecting SV (spatially variable) genes, with distinct spatial expression patterns. However, the accuracy of using these SV genes in clustering has not been thoroughly studied. On the other hand, in single cell resolution sequencing data without spatial context, clustering analysis is usually done on highly variable (HV) genes. Here we investigate if integrating SV genes and HV genes from spatial transcriptomics data can improve clustering performance beyond using SV genes alone. We examined three methods that detect SV genes, including Giotto, spatialDE, and SPARK, and evaluated six methods that integrate different features measured from the same samples including MOFA+, scVI, Seurat v4, CIMLR, SNF, and the straightforward concatenation approach. We applied these methods on 19 real datasets from three different spatial transcriptomics technologies (merFISH, SeqFISH+, and Visium) as well as 20 simulated datasets of varying spatial expression conditions. Our evaluations show that MOFA+ and simple concatenation have good performances in general, despite the variations among datasets and spatial transcriptomics platforms. This work shows that integrating highly variable and spatially variable genes in the spatial transcriptomics data can improve clustering beyond using spatially variable genes only. It also provides practical guides on the choices of computational methods to accomplish this goal.