Fig. 4

Benchmarking spatial clustering methods using synthetic data from SRTsim. The synthetic data are either generated based on the DLPFC reference data with seven original layers (A–D) or the STARmap data with four layers (E–G). A The reference DLPFC data contains seven layers, including six DLPFC layers and white matter (WM). Sample 151673 in the DLPFC data is used to serve as the reference. B, C The performances of six spatial clustering methods (x-axis) are evaluated based on the adjusted Rand index (y-axis) using synthetic data generated from SRTsim via either the tissue-based simulations (B) or the domain-specific simulations (C) based on the DLPFC reference. The six spatial clustering methods include BayesSpace, SNN, stLearn with Louvain, stLearn with Kmeans, spaGCN, and HMRF. D The reference STARmap data contains four cortical layers. E, F The performances of six spatial clustering methods (x-axis) are evaluated based on the adjusted Rand index (y-axis) using synthetic data generated from SRTsim via either the tissue-based simulations (E) or the domain-specific simulations (F) based on the STARmap reference. For B, C and E, F, ARI results are summarized across 100 simulation replicates. For the HMRF, we included the best result obtained by examining a range of values for the spatial parameter beta. For the SpaGCN, there is no refined version for the STARmap-based simulation, as no histology image is available for the data generated by this platform. G The performances of six spatial clustering methods in terms of ARI (x-axis) are evaluated under varying-sequencing-depth (scenario I), varying-location-number (scenario II), and fixing-average-sequencing-depth (scenario III)