Understanding the Link Between Functional Profiles and Intelligence Through Dimensionality Reduction and Graph Analysis

Abstract

There is a growing interest in neuroscience for how individual-specific structural and functional features of the cortex relate to cognitive traits. This work builds on previous research which, by using classical high-dimensional approaches, has proven that the interindividual variability of functional connectivity (FC) profiles reflects differences in fluid intelligence. To provide an additional perspective into this relationship, the present study uses a recent framework for investigating cortical organization: functional gradients. This approach places local connectivity profiles within a common low-dimensional space whose axes are functionally interpretable dimensions. Specifically, this study uses a data-driven approach to model the association between FC variability and interindividual differences in intelligence. For one of these loci, in the right ventral-lateral prefrontal cortex (vlPFC), we describe an association between fluid intelligence and the relative functional distance of this area from sensory and high-cognition systems. Furthermore, the topological properties of this region indicate that, with decreasing functional affinity with high-cognition systems, vlPFC functional connections are more evenly distributed across all networks. Participating in multiple functional networks may reflect a better ability to coordinate sensory and high-order cognitive systems.

Keywords: functional connectivity; functional gradients; intelligence; interindividual differences; topology.

FIGURE 1

(A–C) The first three gradients of functional connectivity (group median) displayed on the inflated cortical surface. These three dimensions each recapitulate a different functional axis, as follows: Sensory‐DMN axis (A), FPN‐DMN (B), and visual‐somatomotor (C). (D–E) Scatterplot of the median location of all cortical vertices along the first and second gradient (D), and first and third gradient (E). Vertices are colored based on their affiliation to the seven resting‐state functional networks as per legend (bottom right). (F) Average scree plot of the variance explained by the first 1000 components of the individual embeddings before (pink) and after (purple) weighting by the fraction of variance explained by the group embedding of individual components.

FIGURE 2

Visual summary of the analyses. The analyses are composed of two streamlines. First (gray box), we concatenated individual resting‐state time series and used GCCA to project them into a common latent space. Within this space we measured vertex‐wise dispersion maps that were then thresholded to identify clusters of maximum variability. Then (outside the gray box), we parcellated the original time series using an atlas to which the variability clusters had been added. We built a correlation matrix of regional time series and thresholded it to the top 10% of connections to obtain a graph of functional connectivity. Graph topology was then analyzed to characterize the role of different regions within it. BOLD, blood‐oxygen‐level‐dependent; FC, functional connectivity; GCCA, generalized canonical correlation.

FIGURE 3

(A) Map of cross‐subject vertex dispersion in gradient space displayed on the inflated cortical surface. (B) Clusters of high interindividual variability obtained by thresholding the surface map of interquartile range at the 95th percentile. (C) Topological profiles of the variability clusters visualized as radar plots. The plots are colored based on the cluster they correspond to in panel B. BC: betweenness centrality; CC: clustering coefficient; dlPFC: dorsolateral prefrontal cortex; GE: global efficiency; LE: local efficiency; OP: occipital pole; PC: participation coefficient; S: strength; TPJ: temporo‐parietal junction; vlPFC: ventrolateral prefrontal cortex.

FIGURE 4

Fluid intelligence (A), card sorting (B), and flanker inhibition (C) scores plotted against the principal gradient of the right ventrolateral prefrontal cluster (R‐vlPFC) after correction for age, gender, education, and handedness. Each marker represents the principal gradient and cognitive scores of R‐vlPFC in an individual and its color represents the cluster's participation coefficient. G1: gradient 1; resid.: residuals; R‐vlPFC: right ventrolateral prefrontal cortex.