Transfer of motor learning is associated with patterns of activity in the default mode network

Abstract

An often-desired feature of motor learning is that it generalizes to untrained scenarios. Yet, how this is supported by brain activity remains poorly understood. Here we show, using human functional MRI and a sensorimotor adaptation task involving the transfer of learning from the trained to untrained hand, that the transfer phase of adaptation re-instantiates a highly similar large-scale pattern of brain activity to that observed during initial adaptation. Notably, we find that these neural changes, rather than occurring at the level of sensorimotor regions, predominantly occur across distributed areas of higher-order transmodal cortex, specifically in regions of the default mode network (DMN). Moreover, we show that these learning-related neural changes relate to the structural properties of transmodal cortex (its myelin content and neurotransmitter receptor density), and that intersubject differences in DMN activity relate to both adaptation- and transfer-phase task performance. Together, these findings suggest that the transfer of learning across the hands is supported by the re-expression of the same activity patterns in the DMN as those that support initial learning. Collectively, these results offer a unique characterization of the whole-brain macroscale changes associated with sensorimotor learning and generalization, and establish a key role for higher-order brain areas, such as the DMN, in the transfer of learning to untrained scenarios.

Fig 1. Task structure and overview of fMRI analysis.

(A) MRI setup (top) and task structure (bottom). (B) Average participant learning curves (median across 8-trial bins). Orange and green traces denote periods during the task in which the left hand (LH) and right hand (RH) were used, respectively. Banding around each trace denotes ±1 standard error of the mean (SEM). Inset bar graphs at right compare the initial error (first 16 trials) for RH learning and LH transfer trials. ** denotes p < 0.01 (C) Neural analysis approach. Right, for each participant, we extracted time series data from the Schaefer 400 cortical parcellation, Tian 32-region subcortical parcellation, and Nettekoven 32-region cerebellar parcellation for six equal-length task epochs (indicated by the coloured boxes) and then computed functional connectivity (FC) matrices for each epoch. We then estimated connectivity manifolds for each task epoch using PCA with centered and thresholded connectivity matrices (see Materials and methods). To visualize how the dominant pattern of FC changed across task phases before alignment to a common space, the surface brain maps depict the first Principal Component (PC1) loadings calculated from a representative subject’s covariance matrices for each specific task epoch (e.g., LH Baseline, RH Baseline, etc.). The color bar indicates the PC1 loading value for each brain region, reflecting the strength and direction of that region’s contribution to the epoch’s primary connectivity pattern. These maps serve an illustrative purpose and will differ slightly from the first PC of the common template Baseline manifold shown in Fig 2, which provides the reference space for all subsequent alignment and analyses. Left, construction of this template Baseline manifold. All manifolds were aligned using Procrustes transform to a common template manifold created from a group-averaged FC matrix based on the mean across the LH and RH Baseline epochs. This allowed us to assess learning-related changes in manifold structure from this Baseline architecture. (D and E) Subject-level clustering is abolished through a Riemannian centering approach. UMAP visualization of the similarity of FC matrices, both before centering (D) and after (E) centering. In these plots, each point represents a single FC matrix, color-coded either to subject identity (left panels) or task epoch (right panels), with its location in the two-dimensional space based on the similarity between matrices. Note that the uncentered connectivity matrices in D show a high degree of subject-level clustering, thus obscuring any differences in task structure. By contrast, the Riemannian manifold centering approach (in E) abolishes this subject-level clustering. The data and code needed to generate this figure can be found in https://zenodo.org/records/15648991.

Fig 2. Main connectivity axes extracted from Baseline trials.

(A) Region loadings across cortex (top), subcortex (middle) and cerebellum (bottom) for the top three PCs. (B) Variance explained for the first 10 PCs. Black trace shows the cumulative variance explained across PCs, whereas the blue bars denote the variance explained for each individual PC. (C) The Baseline (template) manifold in low-dimensional space, with regions colored according to the Yeo and colleagues, 17-network assignment [58,70]. (D) Illustration of how eccentricity is computed. A single region’s eccentricity along the manifold is calculated as the Euclidean distance (dashed line) from manifold centroid (black square). The eccentricity of three example brain regions is highlighted. (E) Regional eccentricity during Baseline trials. Each brain region’s eccentricity is color-coded in the low-dimensional manifold space (left) and on the cortical, subcortical, and cerebellar surfaces (right). Black square denotes the center of the manifold (manifold centroid). L = left; R = right.

Fig 3. Multiplexing of task-based information in global manifold architecture.

(A) UMAP visualization of the similarity in neural states across task epochs (based on the data in B). In this plot, each point represents the across-subject mean whole-brain eccentricity for different task epochs (see legend at bottom), with nearby points representing more similar patterns of whole-brain eccentricity. Note that the two UMAP dimensions appear to capture different types of information about the task structure (see text). (B) Average across-subject representational similarity matrix (RSM) for the patterns of whole-brain eccentricity associated with the six task epochs (the value in each cell represents the Pearson spatial correlation). Note that the RSM is symmetric about the diagonal (black squares, which represent self-correlations). (C) Idealized model RSMs representing different hypotheses about task-related structure. (D) An idealized model RSM representing a control hypothesis based on the passage of time. In these model RSMs (C and D), values of 1 indicate predicted similarity between epochs according to the model, while 0 indicates predicted dissimilarity. These models are compared to the empirical data RSM shown in B. (E) Model comparisons. Each bar indicates the across-subject mean of cosine similarity between the RSM of the whole-brain eccentricity data (in B) and each of the models (in C and D; higher values mean higher similarity). Inference was performed via bootstrap resampling of subjects (1,000 bootstrap samples). Error bars indicate the standard error of the mean. The significance of each model for a one-sided comparison against zero is marked by white dew drops on the horizontal axis, and against the lower-bound estimate of the noise ceiling (gray icicles; FDR-corrected for 4 models). Models’ Pairwise differences are summarized by stars. ** denotes q < 0.05. The data and code needed to generate this figure can be found in https://zenodo.org/records/15648991.

Fig 4. Effector-related changes in manifold eccentricity.

(A) Brain areas showing a significant main-effect of Hand, based on region-wise two-way repeated measures ANOVAs using a false discovery rate (FDR) correction for multiple comparisons (q < 0.05). (B) Aggregation of significant brain regions in (A) according to functional network assignment (Yeo 17-networks parcellation [70]). (C) Temporal trajectories of significant regions from A in low-dimensional manifold space. Colored circles indicate each region’s initial position during the LH Baseline epoch, and the accompanying trace shows the unfolding displacement of that region across the subsequent five phases of the task (RH Baseline, RH Learning Early, RH Learning Late, LH Transfer Early, and LH Transfer Late). Nonsignificant regions are shown in gray point cloud. (D) Meta-analyses of the main effects depicted in A based on the NiMARE correlation decoder tool [75] with the Neurosynth database [76]. Word clouds show the top and bottom 15 processes that are associated with the brain map in A. The text size and color within each word cloud denote their correlation value (bigger words have higher correlation values) and their polarity (words in red and blue are positively and negatively associated, respectively, with the main effect brain maps). (E) Patterns of regional changes in manifold eccentricity underlying the main effects in A. Left, Pairwise contrast denoting the significant main effect of Hand. Note the general contralateral organization of eccentricity changes. Right, Scatter plots show the eccentricity for each significant region (averaged across participants) and separated according to hemisphere (left vs. right). The line plot overlays show the group mean (across ROIs) over task epochs. Note that data points are color-coded, as in Fig 1, according to the hand used during each epoch (orange = left hand; green = right hand). Left = left; Right = right. The data and code needed to generate this figure can be found in https://zenodo.org/records/15648991.

Fig 5. Learning/transfer-related changes in manifold eccentricity.

(A) Brain areas showing a significant main-effect of Task epoch, based on region-wise two-way repeated measures ANOVAs using a false discovery rate (FDR) correction for multiple comparisons (q < 0.05). (B) Aggregation of significant brain regions in (A) according to functional network assignment (Yeo 17-networks parcellation [70]). (C) Temporal trajectories of significant regions from (A) in low-dimensional manifold space (plotted the same as in Fig 4). (D) Meta-analyses of the main effects depicted in A based on the NiMARE correlation decoder tool [75] with the Neurosynth database [76]. See Fig 4 caption for details. (E) Patterns of regional changes in manifold eccentricity underlying the main effects in A. Pairwise contrasts (left) and eccentricity plots (right) denoting the significant main effect of Task epoch. The eccentricity plots highlight the noticeable increases in eccentricity during each of the RH Learning Early and LH Transfer Early epochs. Left = left; Right = right. The data and code needed to generate this figure can be found in https://zenodo.org/records/15648991.

Fig 6. Re-expression of early learning manifold structure across epochs.

(A) Mean eccentricity values for the RH Learning Early epoch for brain regions exhibiting a significant main effect of Task Epoch (i.e., areas from Fig 5A). (B) Pattern similarity (Pearson r) between RH Learning Early and the three subsequent task epochs—RH Learning Late, LH Transfer Early, and LH Transfer Late—computed across the regions shown in A. The black line shows the across-subject mean, and individual points represent single subjects. (C) Mean eccentricity values for the RH Learning Early epoch for brain regions exhibiting a significant main effect of Hand (i.e., areas from Fig 4A). (D) Pattern similarity between RH Learning Early and the same three comparison epochs as in (A), but computed across the Hand-significant regions shown in C. * denotes p < 0.05 for one-tailed paired-samples t tests. The data and code needed to generate this figure can be found in https://zenodo.org/records/15648991.

Fig 7. Effector-related and learning/transfer-related main-effects relate to underlying brain structure.

Spatial correlations between our two main-effect F-stat maps (shown at top) with (A) cortical myelin concentration (T1w/T2w ratio) and (B) the first principal component of receptor density (B, from [52]). Spatial null model testing was performed using the Neuromaps toolbox [98,97]. L = left; R = right.

Fig 8. Connectivity changes that underlie the main effects of hand and task epoch.

(A and B) Connectivity changes for each M1 seed region. Seed regions are denoted in yellow and indicated by arrows. Positive (red) and negative (blue) values show increases and decreases in connectivity for the contrast of Contralateral > Ipsilateral hand. Scatter plots at right show the eccentricity for each participant, and the line plot overlays show the group mean across task epochs. Note that data points are color-coded, as in Fig 1, according to the hand used during each epoch (orange = left hand; green = right hand). (C and D) Connectivity changes for each MPFC seed region (denoted in yellow). Positive (red) and negative (blue) values show increases and decreases in connectivity, respectively, for Baseline to Early learning/transfer (leftmost panel) and for Early to Late learning/transfer (adjacent right panel). Eccentricity scatter plots (middle) are the same as in A and B. Rightmost panel contains spider (polar) plots, which summarize these patterns of changes in connectivity at the network-level (according to the Yeo 17-networks parcellation [70]). Note that the black circle in the spider plot denotes t = 0 (i.e., zero change in eccentricity between the epochs being compared). The data and code needed to generate this figure can be found in https://zenodo.org/records/15648991.

Fig 9. Individual differences in motor learning and generalization.

(A) Individual subject learning curves. The Solid black line denotes the mean across all subjects, binned by trial block, whereas the light gray traces denote individual participants. The green and red traces denote the learning curves of a good (low error) and poor (high error) learner, respectively. (B) Distribution of median angular error during the RH Learning Early and LH Transfer Early epochs (corresponding to the timepoints covered by the faint blue boxes in A), as well as Transfer Rate, constructed by subtracting the RH learning Late angular error (red box) from the LH Transfer Early epoch error. Single data points denote the median error of individuals. (C) Correlation between subjects’ angular error during the RH Learning Early and LH Transfer Early epochs. (D) Correlation between subjects’ Transfer Rate and angular error during the LH Transfer Early epoch. (E) Correlation between subjects' explicit reports (collected at the end of the Learning block) and angular error during the RH Learning Early epoch (F) Same as E, but for the angular error during the LH Transfer Early epoch. Single data points denote individuals, and the black line denotes the best-fit regression line, with shading indicating ±1 standard error of the mean (SEM). The data and code needed to generate this figure can be found in https://zenodo.org/records/15648991.

Fig 10. Relationship between learning performance and learning-related changes in eccentricity.

(A) Whole-brain correlation map between subjects’ RH early error and the change in regional eccentricity from RH Baseline to RH Learning Early. Black bordering denotes regions that are significant at p < 0.05. (B) Results of the spin-test permutation procedure, assessing whether the spatial topography of correlations in A are specific to individual functional brain networks in each hemisphere. The density graph denotes the null distribution for statistically significant brain networks only, as derived from 1,000 iterations of a spatial autocorrelation-preserving null model [98,97] (see Fig G in S1 Text for the results from other non-significant networks). The dashed vertical line denotes the true correlation value. All correlations were corrected for multiple comparisons (q < 0.05). Scatterplots above panel B show the correlation between the change in eccentricity for a representative brain region from each significant network (region denoted in yellow) with subjects’ median angular error during the RH Learning Early epoch. (C) Underlying pattern of functional network connectivity, and its relationship to learning performance, for two of the significant networks in B. Positive (red) and negative (blue) values denote where an increase in inter-network connectivity was associated with either higher or lower angular errors, respectively (i.e., blue values denote where increased connectivity led to lower errors, or better performance). Spider plots, at bottom, summarize these patterns of correlation changes at the network-level. Note that the black circle in the spider plot denotes r = 0 (i.e., zero correlation between the change in functional connectivity and performance). (D) Same as A but for the correlation between subjects’ LH early error and the change in regional eccentricity from LH Baseline to LH Transfer Early. (E) Same as B, but for the correlation map in D. (F) Underlying pattern of functional network connectivity, and its relationship to transfer performance, for two of the significant networks in E (see Fig G in S1 Text for the results from other non-significant networks). Data is presented the same as in C, but for angular errors during the LH Transfer Early epoch. Again, here, blue values denote where increased connectivity was associated with lower errors.