Fig. 1: Overview of the steps to turn functional MRI data into a network.

Starting from preprocessed and denoised data, the following steps are involved. (i) Use of data with vs without global signal regression (GSR), in addition to other denoising protocol (aCompCor for NYU-short, NYU-long and Cambridge datasets; FIX-ICA for HCP); (ii) Definition of nodes (based on anatomical features, local and global functional characteristics, or multimodal features; or Independent Components Analysis); (iii) Choice of number of nodes (approximately 100, 200, or 400); (iv) Definition of connectivity measure (from Pearson correlation or mutual information); (v) Choice of edges to retain (8 filtering schemes considered, based on a priori choices of network density, or minimum edge weight, or data-driven strategies to optimise the balance between network efficiency and wiring cost), (vi) Use of binary or weighted edges. In total, we consider 2 × 4 × 3 × 2 × 8 × 2 = 768 unique pipelines. For each pipeline, the resulting functional networks are compared for the same subject across different time-spans (minutes, weeks, or months) using the Portrait Divergence. A network portrait for a binary network is a matrix B whose rows each correspond to a histogram obtained by thresholding the matrix of shortest paths between the networks’s constituent nodes, at each path length l between 0 and the network’s diameter L, such that entry B_l,k encodes the number of nodes that have k nodes at distance l. For weighted networks, the histogram is obtained by binning (see Methods). Illustration of parcellations adapted from refs. ³⁴ and ⁴⁴; illustration of Portrait Divergence adapted from ref. ⁵³.