Omnipresence of the sensorimotor-association axis topography in the human connectome

Abstract

Low-dimensional representations are increasingly used to study meaningful organizational principles within the human brain. Most notably, the sensorimotor-association axis consistently explains the most variance in the human connectome as its so-called principal gradient, suggesting that it represents a fundamental organizational principle. While recent work indicates these low dimensional representations are relatively robust, they are limited by modeling only certain aspects of the functional connectivity structure. To date, the majority of studies have restricted these approaches to the strongest connections in the brain, treating weaker or negative connections as noise despite evidence of meaningful structure among them. The present work examines connectivity gradients of the human connectome across a full range of connectivity strengths and explores the implications for outcomes of individual differences, identifying potential dependencies on thresholds and opportunities to improve prediction tasks. Interestingly, the sensorimotor-association axis emerged as the principal gradient of the human connectome across the entire range of connectivity levels. Moreover, the principal gradient of connections at intermediate strengths encoded individual differences, better followed individual-specific anatomical features, and was also more predictive of intelligence. Taken together, our results add to evidence of the sensorimotor-association axis as a fundamental principle of the brain's functional organization, since it is evident even in the connectivity structure of more lenient connectivity thresholds. These more loosely coupled connections further appear to contain valuable and potentially important information that could be used to improve our understanding of individual differences, diagnosis, and the prediction of treatment outcomes.

Keywords: Connectivity gradients; Individual differences; Resting-state fMRI; Somatosensory-association axis; Weak connections.

Fig. 1.

Overview of the connectivity gradient workflow. In contrast to keeping only the strongest connections, the functional connectivity structure is vertex-wise (row-wise) thresholded based on a defined percentile bin with lower and upper boundary. This results in functional connectivity (FC) profiles with specific subsets of connections that are systematically defined based on their functional connectivity strengths. A similarity matrix is established by the vertex-wise cosine distance of this thresholded and binarized connectivity matrix. Principal component analysis (PCA) is then applied to this similarity matrix to establish a percentile bin-specific embedding and connectivity gradients.

Fig. 2.

Similar principal gradients are observed for the top-ranked connections (functional affinity) and the opposing bottom-ranked connections (functional enmity). The thresholded functional connectivity (FC) profiles demonstrate a clear converse functional topography, while their intrinsic similarity structures exhibit a relatively common organizational pattern. The principal gradients in both cortex and subcortex exhibit a high spatial similarity (r >= 0.84).

Fig. 3.

The sensorimotor-association axis is omnipresent in the functional organization thresholded at different connectivity levels. (A) The thresholded connectivity profiles illustrate a gradual transition between an antagonistic spatial organization, but their intrinsic similarity structures resemble a shared pattern of coherence across the different connectivity levels. (B) Only the principal gradient is omnipresent across the different connectivity thresholds, explaining always the most variance as the first component. The topographical patterns of the somatomotor-visual and task-negative to task-positive gradients vary across the connectivity strengths. The somatomotor-visual gradient emerges only with increasing connectivity strengths, emphasizing its characterization of stronger short-range connections. The task-negative to task-positive gradient describes a similar spatial pattern for the more top- and bottom-ranked connections, but diverges for the intermediate connectivity structure.

Fig. 4.

Clustering of the gradient signatures across the connectivity levels reveals a distinct spatial pattern that reflects cortical hierarchy. (A) The sensorimotor and association cluster, characterizing the opposing ends of the connectivity spectrum, are separated by an intermediate cluster that indicates regional variation across the connectivity thresholds. (A similar spatial cluster pattern is observed for the cortex and in subcortical structures. At intermediate connectivity levels, the gradient coefficients of the association and the variable cluster are more similar. (B) The cluster configuration is repeatedly found across 100 individuals in the HCP sample, with the highest variability in the variable cluster. The variable cluster showed the largest overlap in attention and limbic networks, that are bordering the antagonistic association and sensorimotor clusters. (C) On the individual level, the clustering follows individual microstructural features such as cortical myelination (T1w/T2w ratio) and cortical thickness. (D) At intermediate connectivity levels, the gradient scores of brain regions associated with attention and higher-order networks are more similar, suggesting a stronger integration. (E) Particularly the principal gradients of intermediate connectivity bins are associated with individual microstructure as quantified by cortical myelination (T1w/T2w ratio) and cortical thickness.

Fig. 5.

The principal gradients of intermediate connections capture individual differences, and (A) demonstrate an increased ability to identify individuals. (B) Intraclass correlation between the principal gradients of different connectivity bins emphasizes the distinct topographical patterns of intermediate connections, and their reliability across sessions. C-E) Leveraging the principal gradients of intermediate connections as features can improve the prediction of Sex, Age, and full-scale intelligence quotient (FSIQ).