Dissociating individual connectome traits using low-rank learning.

Intrinsic functional connectivity (FC) exhibits high variability across individuals, which may account for the diversity of cognitive and behavioural ability. This variability in connectivity could be attributed to individual-specific trait and inter-session state differences (intra-subject differences), as well as a small amount of noise. However, it is still a challenge to perform accurate identification of connectivity traits from FC. Here, we introduced a novel low-rank learning model to solve this problem with a new constraint item that could reduce intra-subject differences. The model could dissociate FC into a substrate (substrate) that delineates functional characteristics common across the population and connectivity traits that are expected to account for individual behavioural differences. Subsequently, we performed a sparse dictionary learning algorithm on the extracted connectivity traits and obtained a dictionary matrix, named connectivity dictionary. We could then predict cognitive behaviours, including fluid intelligence, oral reading recognition, grip strength and anger-aggression, more accurately using the connectivity dictionary than the original FC. The results reflect that we captured individual connectivity traits that more effectively represent cognitive behaviour. Moreover, we found that the functional substrate is significantly correlated with large-scale anatomical brain architecture, and individual differences in connectivity traits are constrained by the connectivity substrate. Our findings may advance our understanding of the relationships among anatomy, function, and behaviour.