Positive Connectivity Predicts the Dynamic Intrinsic Topology of the Human Brain Network

Abstract

Functional connectivity MRI (fcMRI) has become instrumental in facilitating research of human brain network organization in terms of coincident interactions between positive and negative synchronizations of large-scale neuronal systems. Although there is a common agreement concerning the interpretation of positive couplings between brain areas, a major debate has been made in disentangling the nature of negative connectivity patterns in terms of its emergence in several methodological approaches and its significance/meaning in specific neuropsychiatric diseases. It is still not clear what information the functional negative correlations or connectivity provides or how they relate to the positive connectivity. Through implementing stepwise functional connectivity (SFC) analysis and studying the causality of functional topological patterns, this study aims to shed light on the relationship between positive and negative connectivity in the human brain functional connectome. We found that the strength of negative correlations between voxel-pairs relates to their positive connectivity path-length. More importantly, our study describes how the spatio-temporal patterns of positive connectivity explain the evolving changes of negative connectivity over time, but not the other way around. This finding suggests that positive and negative connectivity do not display equivalent forces but shows that the positive connectivity has a dominant role in the overall human brain functional connectome. This phenomenon provides novel insights about the nature of positive and negative correlations in fcMRI and will potentially help new developments for neuroimaging biomarkers.

Keywords: functional connectivity MRI; negative connectivity; positive connectivity; stepwise functional connectivity; topological causality.

Figure 1

Diagram of the neuroimaging-processing pipeline in one subject. (A) Functional magnetic resonance images (fMRI) were transformed into functional connectivity processed BOLD signals (top-left). Using an overlapping sliding window approach (top-left; 15 time-points, 1 time-point lag), functional connectivity processed BOLD signals of voxels were used to create network association matrices based on Pearson R (Fisher transformed) correlations (top-right). (B) Network association matrices (M) were the input for two distinctive analytical strategies: (1) stepwise functional connectivity analysis (SFC) to compute optimal distances of the network (bottom-left) and (2) Granger causality analysis of Euclidian distance similarities over time (bottom-right). SFC analysis diagram represents the SFC matrices of positive correlations with the corresponding step distances from 2 to 7 in each network configuration (T1, T2 … T106) (bottom-left). Before the Granger causality analysis, each sliding window association matrix was separated into positive and negative correlations. (C) Then, each node-based network of positive or negative connections (rows or columns in the positive or negative association matrices; also known as star network) were used to compute the Euclidean distance E, where lower values represent a similar topology between networks and higher different topology between networks (displayed in similar or different colors). Euclidean distances between consecutive pairs of network configurations were computed within positive (bottom-right, left matrix in C) or negative connectivity (bottom-right, right matrix in C), resulting in two matrices containing Euclidean distances of 1,902 voxels by 105 network transition pairs. Finally, a Granger test is applied to the Euclidean distance of positive and negative connectivity time transitions (red lines in Euclidean distance matrices; bottom-right) to determinate if positive-over-negative or negative-over-positive connectivity changes dominate network landscape. In (D) transitions from positive to negative connectivity are displayed for illustration purposes. Assuming a theoretical BOLD time series (red wave), it is possible to observe the arise of negative correlations due to specific lags and graph connectivity distances.

Figure 2

Relationship between stepwise functional connectivity (SFC) optimal distance and average negative correlations in the discovery dataset. X-axis represents the SFC optimal distance for all positive connections. X-axis represents the average negative correlation value. Error bars indicate the standard deviation across all subjects. Of note, line graph starts in two steps -as negative R-values are incomputable if a positive functional connectivity is already established.

Figure 3

Cortical Maps of Granger causality between positive and negative network configurations (via Euclidean distance similarities) over time in the discovery and replication datasets. Cortical maps show the Granger causality findings from two testing hypotheses (positive-explaining-negative and negative-explaining-positive connectivity) in three statistical conditions (I) no statistical threshold, (II) p-value < 0.05, and (III) p-value corrected by an FDR approach with q level at 0.05. Color scale shows the subtraction of F-statistics with a normalized intensity of positive and negative values using a 0–98% transformation.