Fig. 4: Spatial input sensitivity analysis of Krakencoder predictions.

a, Brain network masking and analysis workflow. We estimate the sensitivity of the Krakencoder’s prediction accuracy to regions in a given Yeo functional network by replacing all connectome edge values not in that network with the population mean, and feeding these masked connectomes into the connectome and demographic predictions. The accuracy after masking reveals the amount of information that the particular network’s regional connections have in the mapping and the demographic and behavioral predictions. Upper right panel shows the overall number of edges in each network. bal.acc, balanced accuracy; Cbm, cerebellum; Con, control; DA, dorsal attention; Def, default; Lim, limbic; SM, somatomotor; Sub, subcortex; S/VA, salience/ventral attention; Vis, visual. b, Masked connectome predictions. Connectome prediction identifiability (avgrank, top) and reconstruction accuracy (avgcorr_demean, bottom) using only the connections to and from regions within each brain network shows how much information the model is utilizing from each network when predicting whole-brain connectomes. The Non-masked column on the right shows performance using original connectomes without masking. Values that are lower indicate that the network’s connections do not contribute as much to the accuracy of the Krakencoder’s mapping. c, Similarly, predicting subject sex (top), age (middle), and total cognition (bottom) from masked input data shows the relative amount of information that the connections to and from regions in each network contain about those demographic and cognitive features.