Extended Data Fig. 7: Comparison of the Krakencoder’s SC to FC mapping to state-of-the-art ML-based SC to FC mapping.

Krakencoder’s performance in mapping from SC to FC, compared against a deep neural network (‘deepnet’)²⁴ and a deep graph neural network (‘graphnet’)²⁵. Top-1 accuracy (top1acc), average rank percentile (avgrank), average correlation of predicted and measured FC after (avgcorr_demean) and before (avgcorr) de-meaning are presented in the top row of panels, one value for each FC flavor as output (columns) and various SCs as input (rows). Network metrics are provided in the bottom row of figure panels. There are three SC inputs for the Krakencoder model: volume-normalized deterministic SC for that parcellation only [kraken:SCdt], volume-normalized probabilistic SC for that parcellation only [kraken:SCpr] and the fusion of all 6 SC inputs in the latent space [kraken:fusionSC]. There are two SC input flavors (deterministic [SCdt] and probabilistic [SCpr] volume-normalized SC) for the deepnet and graphnet models. The same metrics between pairs of measured FC from varied familial relatedness categories is shown as a comparison. Deepnet and graphnet models were trained and evaluated using the same data and subject splits as the Krakencoder model. Colors indicate the correlation’s p-value. Bold entries denote significant correlation (p_perm < 10⁻³, two-sided permutation test, 1000 permutations, FDR-corrected across 72 comparisons per metric).