Fig. 2: Benchmarking of DECIPHER and state-of-the-art methods. | Nature Communications

Fig. 2: Benchmarking of DECIPHER and state-of-the-art methods.

From: DECIPHER for learning disentangled cellular embeddings in large-scale heterogeneous spatial omics data

Fig. 2

a Spatial-oriented benchmark results, in which NMI and ARI were reported against the original spatial region annotations. From left to right are results on the 10x mimic dataset, the MERFISH brain dataset and the Xenium breast tumor dataset. The UMAP visualization of spatial regions are shown in Supplementary Fig. 3. b Omics-oriented benchmark results, in which NMI and ARI are reported against the original cell type annotations. From left to right are the results on the 10x mimic dataset, the MERFISH brain dataset and the Xenium breast tumor dataset. STAGATE failed on MERFISH brain dataset because of GPU memory overflow (capping at 80 GB). GraphST failed on MERFISH brain and Xenium breast dataset because of memory overflow (capping at 256 GB). BASS failed on MERFISH brain and Xenium breast dataset because of overtime (capping at 72 h). Hamrony and STAIDA are not applicable to MERFISH brain and Xenium breast datasets which do not contain batch effects. The UMAP visualization cell types are shown in Supplementary Fig. 4. c Benchmark results for batch correction on 10x simulation dataset. batch ASW, batch GC, batch iLSI, and batch kBET were calculated against spatial region annotations (first row) and cell type annotations (second row), respectively. The UMAP visualization of batches were shown in Supplementary Fig. 6. n = 8 repeats with different random seeds. The error bars indicate the means ± s.d. Raw data was provided as Supplementary Data 6. These methods can be mainly divided into three groups: spatial modeling (DECIPHER, Banksy, STAGATE, SLAT, scNiche, BASS), single cell modeling/integration (scVI, Scanpy, Harmony), and spatial batch correction (STADIA, GraphST). Source data are provided as a Source Data file.

Back to article page