Task structure and overview of fMRI analysis.

(A) Subject setup in the MRI scanner. (B) Trial structure of the reward-based motor learning task. On each trial, subjects were required to trace a curved (Visible) path from a start location to a target line (in red), without visual feedback of their finger location. Following a baseline block of trials, subjects were instructed that they would receive score feedback, presented at the end of the trial, based on their accuracy in tracing the visible path. However, unbeknownst to subjects, the score they received was actually based on how accurately they traced the mirror-image path (Reward Path), which was invisible to participants. (C) Example subject data from learning trials in the task. Coloured traces show individual trials over time (each trace is separated by ten trials to give a sense of the trajectory changes over time; 20 trials shown in total). (D) Average participant performance throughout the learning task. Three equal-length task epochs for subsequent neural analyses are indicated by the gray shaded boxes. (E) Neural analysis approach. For each participant and each task epoch (Baseline, Early and Late learning), we estimated functional connectivity matrices using region-wise timeseries extracted from the Schaefer 1000 cortical parcellation and the Harvard-Oxford subcortical parcellation. We estimated functional connectivity manifolds for each task epoch using PCA with centered and thresholded connectivity matrices (see Methods, as well as Fig. 2). All manifolds were aligned to a common template manifold created from a group-average Baseline connectivity matrix (right) using Proscrustes alignment. This allowed us to assess learning-related changes in manifold structure from this Baseline architecture.

Riemmanian centering removes subject-level clustering.

UMAP visualization of the similarity of connectivity matrices, both before centering (A) and after (B) centering. In these plots, each point represents a single functional connectivity matrix, color-coded either to subject identity (left panels) or task epoch (right panels), with its location in the multidimensional space based on the similarity between matrices. Note that the uncentered connectivity matrices in A show a high-degree of subject-level clustering, thus obscuring any differences in task structure. By contrast, the Riemmanian manifold centering approach (in B) abolishes this subject-level clustering. To help illustrate this point, the dashed circles in both A and B indicate the functional connectivity matrices belonging to the same single subject (Subject 1; S1).

Baseline manifold structure and eccentricity.

(A) Region loadings for the top three PCs. (B) Percent variance explained for the first 10 PCs. (C) The Baseline (template) manifold in low-dimensional space, with regions colored according to functional network assignment 58,68. (D) Scatter plots showing the embedding of each region along the top three PCs. Probability density histograms at top and at right show the distribution of each functional network along each PC. Vis: Visual. SomMot: Somatomotor. DorsAttn: Dorsal attention. SalVentAttn: Salience/Ventral attention. Cont: Control. (E) Illustration of how eccentricity is calculated. Region eccentricity along the manifold is computed as the Euclidean distance (dashed line) from manifold centroid (white circle). The eccentricity of three example brain regions is highlighted (bordered coloured circles). (F) Regional eccentricity during Baseline. Each brain region’s eccentricity is color-coded in the low-dimensional manifold space (left) and on the cortical and subcortical surfaces (right). White circle with black bordering denotes the center of the manifold (manifold centroid).

Changes in manifold structure during reward-based motor learning.

(A & B) Pairwise contrasts of eccentricity between task epochs. Positive (red) and negative (blue) values show significant increases and decreases in eccentricity (i.e., expansion and contraction along the manifold), respectively, following FDR correction for region-wise paired t-tests (at q<0.05). The spider plot, at center, summarizes these patterns of changes in connectivity at the network-level (according to the Yeo networks, 68). Note that the black circle in the spider plot denotes t=0 (i.e., no change in eccentricity between the epochs being compared). Radial axis values indicate t-values for the associated contrast (see color legend). (C & D) Temporal trajectories of statistically significant regions from A and B, shown in the low-dimensional manifold space. Traces show the displacement of each region for the relevant contrast and filled coloured circles indicate each region’s final position along the manifold for a given contrast (see insets for legends). Each region is coloured according to its functional network assignment (middle). Nonsignificant regions are denoted by the gray point cloud. White circle with black bordering denotes the center of the manifold (manifold centroid).

Main patterns of connectivity changes that underlie manifold expansions and contractions.

(A-C) Connectivity changes for each seed region. Selected seed regions are shown in yellow and are also indicated by arrows. Positive (red) and negative (blue) values show increases and decreases in connectivity, respectively, from Baseline to Early learning (leftmost panel) and Early to Late learning (second from leftmost panel). Second from the rightmost panel shows the eccentricity of each region for each participant, with the black dashed line showing the group mean across task epochs. Rightmost panel contains spider plots, which summarize these patterns of changes in connectivity at the network-level (according to the Yeo 17-networks parcellation 68). Note that the black circle in the spider plot denotes t=0 (i.e., zero change in eccentricity between the epochs being compared). Radial axis values indicate t-values for associated contrast (see color legend).

Relationship between learning performance and regional changes in eccentricity.

(A) Individual subject learning curves for the learning task. Solid black line denotes the mean across all subjects whereas light gray lines denote individual participants. The green trace denotes an example fast learner and the red trace denotes an example slow learner (see text). (B) Derivation of subject learning scores. We performed functional principal component analysis on subjects’ learning curves in order to identify the dominant patterns of variability during learning. The top component, which encodes overall learning, explained the majority of the observed variance (∼75%). The green and red bands denote the effect of positive and negative component scores, respectively, relative to mean performance. Thus, subjects who learned more quickly than average have a higher loading (in green) on this ‘Learning score’ component than subjects who learned more slowly (in red) than average. (C & D) Whole-brain correlation map between subject Learning score and the change in regional eccentricity from Baseline to Early learning (C) and Early to Late learning (D). Black bordering denotes regions that are significant at p<0.05. (E & F) Results of the spin-test permutation procedure, assessing whether the topography of correlations in C and D are specific to individual functional brain networks. Single points indicate the real correlation value for each of the 17 Yeo et al. networks 68, whereas the boxplots represent the parameters of a null distribution of correlations derived from 1000 iterations of a spatial autocorrelation-preserving null model 76,77. In the boxplots, the ends of the boxes represent the first (25%) and third (75%) quartiles, the center line represents the median, and the whiskers represent the min-max range of the null distribution. All correlations were corrected for multiple comparisons (q<0.05). The dashed horizontal blue line indicates a correlation value of zero and the gray shading encompasses correlation values that do not significantly differ from zero (p>0.05). [Note that in the spin-test procedure, due to the sign of the correlations, it is possible for some networks to be significantly different from the null distribution, and yet not significantly different from zero. Thus, to be considered significant in our analyses, a brain network must satisfy both constraints; i.e., show a correlation that is significantly different from zero and from the spatial null distribution]. Right, scatterplots show the relationships between subject Learning score and the change in eccentricity from Baseline to Early learning (top) and Early to Late learning (bottom) for the DAN-A network (depicted in yellow on the cortical surface at top), the only brain network to satisfy the two constraints of our statistical testing procedure. Black line denotes the best-fit regression line, with shading indicating +1/− standard error of the mean. Dots indicate single participants. (G) Connectivity changes for the DAN-A network (highlighted in yellow) across epochs. Positive (red) and negative (blue) values show increases and decreases in connectivity, respectively, from Baseline to Early learning (left panel) and Early to Late learning (right panel). Spider plot, at right, summarizes the patterns of changes in connectivity at the network-level. Note that the black circle in the spider plot denotes t=0 (i.e., no change in eccentricity between the epochs being compared). Radial axis values indicate t-values for associated contrast (see color legend). VisCent: Visual Central. VisPer: Visual Peripheral. SomMotA: Somatomotor A. SomMotB: Somatomotor B. TempPar: Temporal Parietal. DorsAttnA: Dorsal Attention A. DorsAttnB: Dorsal Attention B. SalVentAttnA: Salience/Ventral Attention A. SalVentAttnB: Salience/Ventral Attention B. ContA: Control A. ContB: Control B. ContC: Control C.