UC Berkeley

Mapping the functional connectome underlying recorded time-series of brain activity can reveal meaningful pathways of large-scale neural computation. However, it remains commonplace to estimate functional connectomes using pairwise correlative methods which are prone to spurious correlations and vulnerable to structured noise. Multivariate statistical estimates of functional connectivity offer new possibilities for generating parsimonious networks. We generated functional connectomes from resting state EEG (electroencephalography) and fMRI (functional magnetic resonance imaging) collected from infants and children under the age of 3. We generated directed multivariate functional connectomes using the ensemble statistical framework Union of Intersections (UoI) to perform regularized multivariate linear regression. We also generated pairwise connections using the Pearson correlation coefficient between each pair of time series in the recording. We found that multivariate estimates of functional connectivity from EEG are sparse and small-world, while pairwise connectomes are lattice-like and spatially correlated. We found a significant difference in small-worldness $\omega$ ($p<<0.0001$) and proximity dependence of coupling strength ($p<< 0.0001$) between pairwise and multivariate EEG functional connectomes. Functional connectomes generated from fMRI were spatially distributed for both pairwise and multivariate estimates, but we observed a significant difference in $\omega$ ($p<<0.001$), with multivariate connectomes clustered tightly around 0, indicating small-worldness. We observed lateralized structure in the multivariate fMRI connectomes that remained stable across age groups. In particular, coupling between the language network posterior superior temporal gyrus (pSTG) and salience network supramarginal gyrus (SMG) showed strong positive ipsilateral connections and strong negative contralateral connections, reinforcing the known lateralization of the language network. The same grid-like structure was observed in sensorimotor lateral fields coupling with the dorsal attention network intraparietal sulcus (IPS). This lateralized structure was not found in pairwise estimates.

We used fMRI functional connectomes to predict cognitive development scores assessed with the Mullen Scales of Early Learning (MSEL). We partitioned subjects into three groups using unsupervised clustering based on their raw MSEL scores. Multiple feature sets were extracted from the functional connectomes, by principal components analysis (PCA) and sparse PCA, and by selecting task-relevant functional network pairs. Random forest classifiers were used to predict the Mullen groups from neural feature vectors. We found that sparse PCs predicted raw scores significantly better than chance for multivariate connectomes, while both PCs and sparse PCs scored significantly better than chance for pairwise connectomes. For multivariate connectomes, functional connectivity between language and salience (SN) scored better than chance. For pairwise connectomes, coupling between the default mode network (DMN) and the frontoparietal network (FPN), and between SN and DMN scored better than chance.

Our understanding of distributed representations of complex sounds in auditory cortex remains incomplete. This is in part due a lack of experimental data for neural responses to complex naturalistic stimuli. Here, we describe the development of a large, diverse dataset of natural sounds. The sources and semantic content of the sounds are described, and acoustic features are calculated for a hand-selected set of 99 high-quality sounds and used to predict semantic labels using a support vector machine. We briefly discuss the intended use case of the sound database as a stimulus set for the characterization of neural representations of natural sound statistics, and selection for features that discriminate semantic categories.