Sex-based Network-Specific Differences in Connectomes: A Krakencoder-Based Analysis

A simulation study of 702 healthy adults found that removing the Default Mode Network caused the largest disruptions in cross-modal brain connectome predictions, while sex-specific information in the resulting perturbation signatures remained subtle, according to a preprint posted to arXiv on June 15, 2026 [1]. The work used the Krakencoder, a deep-learning framework, to simulate how deficiencies in one brain-connectivity modality — structural or functional — propagate to the other [1][2]. Researchers drew structural and functional connectomes from the Human Connectome Project and systematically removed each of the seven Yeo-7 functional networks, preserving the remaining six, to create seven perturbation scenarios [1][3]. A connectome is a comprehensive map of neural connections, often described as the brain’s “wiring diagram,” and can be captured at millimeter scale using MRI [10]. Three complementary metrics — KL divergence on eigenvalue spectra, Frobenius norm, and Wasserstein distance — were used to quantify the resulting deviations in cross-modal predictions [2][4]. Across all metrics and both prediction directions, the Default Mode Network produced the largest perturbations, while the Somatomotor network yielded the smallest [1][3]. The study computed a difference matrix for each network removal, capturing the deviation induced by the absence of a specific network [3]. Researchers also tested whether sex-specific information persisted in the perturbed connectomes. When features were drawn from connectomes predicted under network-removal conditions, the best sex-classification accuracy reached 66.09% [1][4]. In contrast, connectomes predicted from intact inputs achieved substantially higher accuracy, reaching 84.76% [1][4]. “The consistent improvement of combined features over either modality alone suggests complementary sex-discriminative information across directions,” the authors note in the preprint [4]. These findings arrive amid broader efforts to characterize sex differences in brain networks. A separate analysis of resting-state fMRI data from 1,948 young adults reported that biological sex accounts for roughly 10% of the differences in nodal centrality consensus rankings, with females showing higher rankings in regions with stronger intra-system connections and males dominating in areas with stronger inter-system connections [5]. Earlier work on structural connectomes identified individual “implicator edges” whose weights alone could infer sex with more than 67% accuracy, with male-implicating edges concentrated in anterior brain regions and female-implicating edges in posterior regions [6]. The Krakencoder-based study validates network-cut-off as a tool for probing structure-function relationships and confirms that full predicted connectomes retain considerably more sex-discriminative information than perturbation-derived signatures alone [3][4].