PAST enables scalable training and prediction on large data while preserving global spatial patterns. (A) Quantitative performance evaluation of spatial domain characterization on the Stereo-seq MOBS1 data set via supervised cross-validation and unsupervised spatial clustering with a specified number of clusters (Ncluster) and with default resolution (Dlouvain). The cross-validation performance was evaluated by the average score of Cohen's kappa value (κ) and mean F1 score (mF1), respectively, in the fivefold experiments. The spatial clustering performance was evaluated by adjusted rand index (ARI) and adjusted mutual information (AMI). (B) The manual annotation of MOBS1. (C) Visualization of the results of spatial clustering with a specified number of clusters on MOBS1. (D) Sampling strategy of STAGATE for minibatch training. (E) Ripple walk sampling strategy of PAST for scalable subgraph-based training and prediction. (F) Average time cost of different methods on MOBS1. (G) Peak memory usage of different methods on MOBS1.
