Learning induces coordinated neuronal plasticity of metabolic demands and functional brain networks

Abstract

The neurobiological basis of learning is reflected in adaptations of brain structure, network organization and energy metabolism. However, it is still unknown how different neuroplastic mechanisms act together and if cognitive advancements relate to general or task-specific changes. Therefore, we tested how hierarchical network interactions contribute to improvements in the performance of a visuo-spatial processing task by employing simultaneous PET/MR neuroimaging before and after a 4-week learning period. We combined functional PET and metabolic connectivity mapping (MCM) to infer directional interactions across brain regions. Learning altered the top-down regulation of the salience network onto the occipital cortex, with increases in MCM at resting-state and decreases during task execution. Accordingly, a higher divergence between resting-state and task-specific effects was associated with better cognitive performance, indicating that these adaptations are complementary and both required for successful visuo-spatial skill learning. Simulations further showed that changes at resting-state were dependent on glucose metabolism, whereas those during task performance were driven by functional connectivity between salience and visual networks. Referring to previous work, we suggest that learning establishes a metabolically expensive skill engram at rest, whose retrieval serves for efficient task execution by minimizing prediction errors between neuronal representations of brain regions on different hierarchical levels.

Fig. 1. Design and analysis.

a After the initial screening, participants were randomly assigned to the training or the control group. All subjects underwent two simultaneous PET/MRI examinations for acquisition of structural, functional and metabolic data at resting-state and while performing a challenging visuo-spatial processing task (the video game Tetris®, Supplementary Fig. S1). In the 4-week period between the two PET/MRI scans, the training group regularly practiced the task using an online platform, whereas the control group did not. After the second PET/MRI scan, no further training was carried out and the training group completed a final task session on a laptop. Additional testing of different cognitive domains was performed at both PET/MRI examinations. b To obtain a robust estimate of task-specific increases in energy demands, the imaging parameters of glucose metabolism (blue), cerebral blood flow (green) and BOLD-derived activation (red) were combined in a conjunction analysis (intersection, orange). Joint active areas served as target regions for the subsequent network analysis (Supplementary Fig. S2). c We extended metabolic connectivity mapping (MCM) to the whole-brain level to assess learning-induced adaptations in directional connectivity towards regions with high task-specific energy demands. The BOLD signal of each brain voxel (exemplarily shown as yellow/orange squares) yields a certain functional connectivity pattern in the target region (here the occipital cortex). Computing the spatial correlation between patterns of functional connectivity (yellow/orange) and glucose metabolism (blue/green) results in an MCM value for each brain voxel that reflects the directional connectivity to the target. d Finally, simulations were carried out to disentangle the individual contribution of glucose metabolism and functional connectivity to MCM learning effects. Voxels in the target region were gradually removed based on values of connectivity or metabolism (here 50% black voxels in left and right columns, respectively), followed by recalculation of MCM values and the corresponding learning effects.

Fig. 2. Behavioral data for the video game Tetris® measured as score per minute.

Changes in task performance differed between the two PET/MRI measurements (M1 and M2), groups and task conditions (group*time*condition interaction, p < 10⁻⁵). a For the easy task condition, the training group (n = 21) showed a 2.7-fold increase in performance, which was significantly higher compared to the control group (n = 20). b For the hard task, changes in performance followed a similar pattern but effects were more pronounced, with the training group showing a 3.1-fold improvement in performance. Also, task performance for the hard task condition further increased even without training until the final visit (FV). The time between measurements/visits was 4 weeks. Initial performance at measurement 1 was not significantly different between the groups for both task conditions (p > 0.5). c Monitoring the task performance during the training period highlights the continuous improvement. The learning curve further matched with the performance of the two PET/MRI measurements as indicated by the dots (average values of M1 and M2 in b). Solid and dotted lines represent mean and standard deviation, respectively. Data were cut after 21 days as less than 1/3 of the subjects trained longer than this period. For a and b, post-hoc comparisons indicate significant differences for the group*time interactions (^#p < 0.05, ^####p < 10⁻⁹), for the differences between the two measurements (**p < 0.01, ***p < 0.001, ****p < 10⁻¹⁰) and for the difference between measurement 2 and the final visit (^♦p < 0.05). All p-values were corrected for multiple comparisons with the Bonferroni-Holm procedure. Boxplots indicate median values (center line), upper and lower quartiles (box limits) and 1.5× interquartile range (whiskers). Data for the plots are provided in Supplementary Data 1–3.

Fig. 3. Learning-induced changes in metabolic connectivity mapping (MCM) with the occipital cortex as target region.

Four weeks of training the video game Tetris® resulted in specific adaptations of connectivity from the right insula (a) and the dorsal anterior cingulate cortex (dACC, b) to the occipital cortex (group*time*condition interaction, p < 0.05 FWE-corrected cluster level). Post-hoc comparisons showed that at rest MCM increased for both connections in the training group (n = 21) as compared to the control group (n = 20). In contrast, MCM decreased during the hard task condition in the training group. There were no significant changes in the control group between the two measurements (M1, M2). Furthermore, MCM values between training and control groups at measurement 1 were not significantly different. Boxplots show the MCM z-scores of the clusters indicated by the crosshair. Post-hoc comparisons indicate significant differences for the group*time interaction (^#p < 0.05, ^##p < 0.01) and for the differences between the two measurements (*p < 0.05, **p < 0.01), corrected for multiple comparisons with the Bonferroni-Holm procedure. Boxplots indicate median values (center line), upper and lower quartiles (box limits) and 1.5× interquartile range (whiskers). Data for the plots are provided in Supplementary Data 4–7.

Fig. 4. Associations between MCM adaptations and cognitive performance.

Based on the training-induced effects in the salience network (Fig. 3), one would expect that subjects with high MCM values at rest and low values during task execution (i.e., a high divergence between rest and task) show the best cognitive performance after learning. Thus, the difference of MCM values between rest and the hard task condition at the second PET/MRI scan was correlated with task performance. Positive associations of dACC MCM values were observed with the Tetris® score (high score = high performance) of the second PET/MRI measurement (a, rho = 0.46, p < 0.05) and that obtained during the 4-week training period (b, rho = 0.56, p < 0.01, normalized area under curve). Further, dACC MCM values were negatively associated with the mental rotation performance (duration/number of correct answers with low value = high performance, c, rho = −0.56, p < 0.01). All values were rank transformed to account for one outlier, thus correlation values represent Spearman’s rho (n = 21). Data for the plots are provided in Supplementary Data 8–10.

Fig. 5. Simulated perturbations of learning-induced changes in MCM.

We aimed to identify whether learning-specific MCM effects in Fig. 3 were driven by glucose metabolism (CMRGlu) or functional connectivity (FC). Voxels of the occipital cortex (i.e., the MCM target region) were progressively removed based on increasing values of CMRGlu (solid lines) or FC (dashed lines) and training effects were recalculated (F-value of group*time interaction). At resting-state simulated removal of voxels based on CMRGlu abolished training-induced MCM effects for both connections towards the occipital cortex, which was, however, not the case for FC (top row). The inverse pattern was observed for the hard task condition, where training-specific decreases in MCM were nullified when removing voxels based on FC, but not CMRGlu (bottom row). ^#/solid black lines: p < 0.05 when removing voxels based on CMRGlu. */dashed black lines: p < 0.05 when removing voxels based on FC values. 0% of voxels removed represents the results shown in Fig. 3 (i.e., when using the entire target region). Of note, randomly removing up to 90% of voxels in the occipital cortex did not affect the learning-induced changes in MCM at all (right top panel, all p < 0.05), highlighting the specificity of CMRGlu and FC to drive MCM changes and indicating that effects are not dependent on the size of the target region. The colors for the random removal match those of the other panels. Data for the plots are provided in supplementary data 11–13.

Fig. 6. Schematic illustration of training effects and potential neurobiological mechanisms.

Cognitive skill learning results in complementary metabolic adaptations at rest (top row) and functional network reorganization during task execution (bottom row). Glass brains depict directional connectivity from the higher-order salience network (grayscale circles) to the lower-order occipital region (colored circles) as assessed with metabolic connectivity mapping (MCM, number of lines). a Training naïve subjects exhibit low directional connectivity at resting-state (solid lines) between unorganized state units (blue-green circles), since the skill trace is not yet established (random circle arrangement, crossing lines). b Due to the lack of training, higher-order representations in the SN are inaccurate (blurred grayscale Tetris®) in comparison to lower-order visual sensory information (colored Tetris®), resulting in a high prediction error (large thunderbolt) encoded by error units (red–yellow circles). The representational inaccuracy requires substantial dynamic optimization between brain regions of different hierarchies (numerous dashed lines). c With repeated task performance during the learning period functional network reorganization approaches an optimal solution. Presumably, this is realized by a high frequency of synaptic tagging, where optimal task representations are gradually encoded in the salience network by synaptic capture and subsequent anchoring of glutamatergic AMPA receptors. d After the learning period, state unit directional connectivity increases, which equals the consolidated skill engram (parallel lines between organized circles). The metabolic emphasis of this process suggests the energy-intensive formation of clustered and potentiated synapses (line thickness). e The established skill engram can then be retrieved for task execution. This results in a decreased prediction error (small thunderbolt) as representations between higher- and lower-order brain regions became more accurate (sharpened grayscale Tetris®). Thus, only minor cognitive control is required (few dashed lines) to apply an efficient task strategy. In sum, these observations indicate that effects of skill learning at resting-state and during task execution are two sides of the same coin, where different neurobiological mechanisms complement each other to improve task performance. The glass brain was kindly provided by Dr. Gill Brown (https://neuroscience-graphicdesign.com/) under CC BY-NC 4.0.