Fig. 6: DCGAN-reconstructed functional connectivity maps capture individual-specific information.

a For each participant in the test data set, we generated functional connectivity (FC) maps from all vertices within each of the compromised regions. Here, we show FC maps calculated from two randomly selected seeds on the cortical surface, one in the motor region of the lateral parietal cortex and one in the lateral temporal cortex. We generated FC maps from intact BOLD frames (first row), reconstructed BOLD frames (second row), group-averaged functional connectivity data (third row), and data from each participant’s most similar individual (MSI) from the training data set. To identify each test participant’s MSI, we correlated that participant’s FC vectors inside the compromised regions with the corresponding vectors in each of the training data set individuals. The MSI was the individual displaying the highest average correlation. b If reconstructed BOLD frames capture specific details about an individual’s functional brain organization, we would expect the reconstructed FC maps to be more similar to the intact FC maps than to any other FC map. Box-and-whisker plots are shown, depicting the reconstructive accuracy for reconstructed FC maps relative to the intact, MSI, and training group FC maps. The center line indicates the median, box limits indicate upper and lower quartiles, whiskers indicate 1.5 times the interquartile range, and plus signs indicate outliers. As expected, we found that the correlation between participants’ reconstructed and intact FC maps, averaged over all vertices within the five cortical masks (r = 0.66 ± 0.03), was significantly higher than the correlation between reconstructed and MSI FC maps (r = 0.55 ± 0.02; t(19) = 20.70, bootstrapped p = 0.001) and also significantly higher than the correlation between reconstructed and training group-averaged FC maps (r = 0.44 ± 0.02; t(19) = 34.53, bootstrapped p = 0.001). The diagonal lines between the first two boxes join data points from the same participants. All statistical tests were two-sided and corrected for multiple comparisons. Source data are provided as a Source Data file. *p ≤ 0.001.