Fig. 1: Benchmarking Rastermap on simulated data with multiplexed neural activity.

a–d, These panels illustrate how Rastermap works. a, First, Rastermap divides neurons into 50–200 clusters based on their activity (left). The cross-correlations between different clusters are computed at several time lags (right). b, The cluster correlations at different positive time lags are shown for a subset of clusters, and the entry-wise maximum of these matrices over a time window from 0 to T_max defines an ‘asymmetric similarity matrix’. c, The asymmetric similarity matrix is sorted to match the ‘matching matrix’, which is a sum of a global similarity matrix and a local similarity matrix. d, The cluster features are upsampled using a locally linear interpolation method, and then each neuron is assigned to an upsampled cluster center. e, The simulated neurons were sorted by Rastermap or t-SNE and then averaged in bins of 30 neurons—the averages of these neurons are called ‘superneurons’. f, The sorted asymmetric similarity matrix for the simulation. g, The activity of the superneurons aligned to different stimulus events. h, The sorting of neurons from various algorithms plotted against the ground truth sorting. i, For each module of the simulation and each algorithm in h, the percentage of correctly ordered triplets is shown (n = 10 simulations; error bars represent s.e.m.). j, The percentage of contamination in a module with neurons from other modules (n = 10 simulations; error bars represent s.e.m.). Corr, correlation; stim, stimulus.