Species-Shared and -Specific Brain Functional Connectomes Revealed by Shared-Unique Variational Autoencoder

A comparative study of large-scale species-shared and -specific functional connectomes, that are respectively inherited or diverged from their common ancestor, is important to the understanding of emergence and evolution of brain functions, cognitions and behaviors. However, recent works largely relied on the scheme that one species is used as the “reference” to which another is aligned and contrasted, whereas a more reasonable “reference” could be related to their common ancestor and unknown. To this end, we proposed a novel method termed shared-unique variational autoencoder (SU-VAE), and applied it to macaque and human MRI datasets to disentangle species-specific variation of functional connectomes from species-shared one. The reconstructed shared and specific connectomes gain supports from reports. The proposed method was further validated by the results that human-specific latent features, in contrast to shared features, better capture the variation of behavior variables unique to human, such as language comprehension. Our studies outperform other methods developed on VAE and linear regression in the aforementioned validation study. Finally, a graph analysis on the identified shared and specific connectomes reveals that, human/macaque -specific connecomtes positively/negatively contribute to the enhancement of network efficiency.