Learning behavior during sensorimotor adaptation

To study the neural processes that support sensorimotor learning, we had subjects (N = 36) perform a visuomotor rotation (VMR) task [21] in the MRI scanner. In this classic task, subjects were required to move a cursor, representing their right finger position, to intercept a target that could appear in one of 8 locations on a circular ring, using an MRI-compatible touchpad (Fig 1. After a baseline block (40 trials), subjects performed a learning block (160 trials) in which the correspondence between finger motion on the touchpad and movement of the cursor was rotated clockwise by 45^°. Thus, subjects needed to learn to adapt their movements by applying a 45^° counter-clockwise rotation in order to hit the target. At the conclusion of each trial, subjects received visual (error) feedback consisting of both the target location, as well as the position of their cursor, on the target ring. Following the learning block, subjects performed 16 “report” trials [13], in which they were asked to indicate their aim direction using a dial, controlled by their left hand, prior to executing a target-directed movement via their right hand. These report trials were presented at the end of the task, after learning had taken place, in order to avoid drawing attention to the nature of the visuomotor perturbation and thus biasing subjects’ learning behavior [22]. Critically, these report trials provided us with a direct index of subjects’ explicit knowledge of, and strategy in counteracting, the visuomotor rotation [13].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. VMR task and subject behavior.

(A) Subjects were cued to move a cursor to a target appearing on a circular ring using an MRI compatible touchpad. During learning trials, the cursor moved at a 45^° angle relative to the hand, so that subjects had to learn to adjust their movement direction in order to contact the target successfully. After learning, subjects performed a set of “report” trials in which they used a dial to indicate their aim direction prior to executing the movement. Using these report trials, we derived estimates of subjects’ total explicit and implicit learning during the task. (B) Error curves for individual subjects. Black trace in bold denotes overall mean error across subjects. (C) Distributions of subjects’ estimated explicit (top) and implicit (middle) contributions to performance, as well as early error (bottom; defined as the mean error during the first 32 trials). (D) For each subject, we computed the mean error during the learning block in each of 5 response time bins separated by the 20%, 40%, 60%, and 80% percentiles. Consistent with previous literature suggesting the use of mental rotation strategies by explicit learners [23], subjects who reported a nonzero re-aiming direction showed monotonically decreasing error with longer response times. The data and code needed to generate this figure can be found in https://zenodo.org/records/14029185.

https://doi.org/10.1371/journal.pbio.3002934.g001

Behavioral data associated with the VMR task are shown in Fig 1. We found that subjects, on average, were able to successfully learn the VMR task by aiming their hand in a direction that counteracted the visuomotor rotation (Fig 1B). Consistent with prior work [13], we defined subjects’ explicit learning as the mean difference between subjects’ reported aim direction and the target location, and defined implicit learning as the mean difference between subjects’ reported and actual aim directions. Finally, we defined task performance as subjects’ average error during the early learning period (defined as the first 100 imaging volumes, or approximately 32 trials, of the task; this was chosen for consistency with our previous work in [24], but see S5 Fig for a robustness analysis using different window sizes). The distributions of these 3 scores are depicted in panel C. Note that we focused our analysis on this early learning period due to evidence that explicit processes are established during this early learning period. Consistent with previous literature [25], we observed a correlation between subjects’ reported explicit knowledge and mean response times (RT) during both early (r = 0.45, t = 2.4, p = 0.006) and late learning (r = 0.69, t₃₄ = 5.63, p = 2.57e − 06); typically believed to reflect the additional time taken to implement a re-aiming strategy [22,25–27]. As an additional validation of our explicit reports, we further examined the relationship between subjects’ error and their RTs. Explicit strategies in the VMR task often make use of mental rotation, and so lower errors are typically observed with longer response times [23]. For each subject, we computed the mean error during the learning block in each of 5 bins, defined using the quantiles of their response time distribution (Fig 1D). Using these bins, we computed the (Spearman) correlation between error and RT for each subject, and observed a significantly lower (negative) average correlation within explicit (compared to implicit) learners (m_exp = −0.44, m_imp = 0.11, t₃4 = −2.38, p = 0.033). That is, explicit learners were more accurate at longer response times, while the accuracy of implicit learners did not vary with response time.

Our decision to elicit subjects’ reports at the end of the learning block was based on evidence that report trials may make subjects aware of the visuomotor rotation and predispose them to use an explicit strategy during the task [22,28]. Nevertheless, group-level behavioral studies suggest that subjects’ explicit re-aiming strategies may evolve during learning the process (e.g., have a larger magnitude during early than late learning; [10,13]). As such, reports collected at the end of the task may not directly reflect the re-aiming strategy subjects used during the early learning process itself. To address this possible concern, we conducted a separate behavioral experiment outside of the MRI scanner (N = 14), in which report trials were interspersed continuously throughout the learning block (see S1 Fig). Importantly, we found that subjects’ explicit reports collected in the final trials of the learning block were highly correlated with reports collected during the early learning block (r = 0.77, t₁₂ = 4.22, p = 0.0012; S1C Fig). Further, we found that subjects’ typically developed explicit awareness of the rotation early in the learning block, and that their subsequent explicit reports were highly stable throughout the VMR task (S1D Fig). Taken together, these findings suggest that the late reports collected during our MRI testing session serve as a valid proxy for the explicit strategy developed during this early learning period.

Connectivity-based approach for studying sensorimotor adaptation

To examine the neural bases that support explicit and implicit learning, as well as the encoding of performance during the task, we examined learning-dependent changes in functional connectivity between the primary motor cortex and the rest of cortex, striatum, and cerebellum (Fig 2A). By constraining our analysis to bipartite functional interactions with primary motor cortex, we were specifically interested in assessing how various regions in the cortex, striatum, and cerebellum, modulate their connectivity with cortical areas positioned in the final motor output pathway that produce motor commands resulting in the measured behavioral adjustments. For each brain region, we extracted fMRI timecourse activity from different atlas parcellations of the cerebellum, striatum, and all of cortex (see Methods), and studied functional connectivity between these brain regions with the average activity of a subset of 4 regions in the left (contralateral) hemisphere comprising the hand motor region (belonging to the left Somatomotor A network; [29]). For each subject, we estimated covariance networks during the baseline, early learning, and late learning periods (each defined as an epoch of 100 imaging volumes), allowing us to describe changes in connectivity over the course of the task. Given that a large body of recent literature suggests that functional connectivity is dominated by static, subject-level differences that can obscure any task-related variance [30], we first centered subjects’ covariance matrices to align their mean covariance across resting-state, baseline, and learning ([24], see Methods for details and for a more thorough description). These centered covariance matrices were then used to identify patterns of functional connectivity associated with distinct learning processes.

Note that, because both implicit and explicit learning result in better performance (less error), the neural correlates of these measures may also reflect effects due to the experience of errors. For example, systems involved in the goal-directed deployment of spatial attention may be necessary for successful task performance regardless of the specific mechanisms by which a given subject learns, and so may be recruited by both implicit and explicit learners. Thus, in order to derive relatively “pure” measures of these 2 learning processes, we controlled for subjects’ performance by using the mean error during the early learning epoch as a covariate in our model. Specifically, we separately regressed each of our explicit and implicit learning measures, and the functional connectivity of each ROI with motor cortex, onto subjects’ early errors, and used each of the residuals for further analysis. Separately, in order to derive neural activity patterns linked to learning performance per se, rather than any particular learning process, we also derived patterns of functional connectivity associated with subjects’ visuomotor errors. We interpret this “Performance” axis as reflecting processes related to general error-detection, or to general task-based processes which may be common to both forms of learning.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Brain networks/regions used in our neural analyses and predictive model.

(A) We studied functional connectivity between primary motor cortex and the rest of cortex, striatum, and cerebellum during VMR learning. Cortical networks were derived using the 400 region parcellation of [29] and assigned to the 17 networks classified by [31]. Here, for the cortical regions, we have arranged the network names according to their position along the principal gradient of functional cortical organization demonstrated by [32]; with networks more proximal to primary sensorimotor regions at the lower-end of the cortical gradient (bottom), and networks more distant from sensorimotor regions at the higher-end of the gradient (top). Cerebellar regions were identified using the atlas of [33]), while striatal regions were extracted from the Harvard-Oxford atlas. Motor cortex parcels were identified as belonging to either hand or tongue areas of the left and right hemispheres based on the reported findings of [29]. (B) We used a set of penalized regression models (see Methods) to predict subjects’ explicit learning, task performance, and implicit learning from patterns of whole-brain functional connectivity with primary motor cortex. The coefficients from each of the fitted models produce a set of brain maps which we hereby refer to as the explicit, performance, and implicit neural axes. The data and code needed to generate this figure can be found in https://zenodo.org/records/14029185.

https://doi.org/10.1371/journal.pbio.3002934.g002

Using the patterns of functional connectivity across regions extracted from the early learning period, we used a group-penalized ridge regression model to predict subjects’ explicit and implicit learning, as well as their performance errors ([34]; see Fig 2, right, and see Methods). We initially focused our analysis on this early learning period as this is the point in time in which explicit learning develops ([13], see also S2 Fig). We used the region-wise coefficients from this model to create whole-brain maps reflecting covariance with motor cortex during early learning, which we hereby refer to as the Explicit, Implicit, and Performance neural axes (Fig 3). Note that we inverted the sign of subjects’ error in the derivation of the performance axis, so that greater loadings were associated with greater performance (i.e., lower error). For each axis, network-specific penalties were tuned by leave-one-out cross-validation (cross-validated ).

Separate neural axes are associated with different component processes of sensorimotor adaptation

Within the Implicit neural axis, we observed the highest regional loadings in the dorsal attention network (DAN), and specifically the DAN-B subnetwork, comprising superior parietal and premotor regions known to support the sensory-guided control of movement [14–16]. Note that, although we did not additionally observe high regional loadings in the cerebellum, a separate analysis focused solely on the cerebellum returned results consistent with its role in implicit error-based adaptation; see Discussion and S3 Fig. By contrast, within the Performance neural axis, we observed the highest regional loadings within the Control network, and specifically the Control-A subnetwork, mainly comprising the dorsolateral prefrontal cortex (DLPFC), inferior parietal lobe (IPL), and anterior cingulate cortex (see barplots in Fig 3). In the sensorimotor adaptation literature, the lateral prefrontal cortex has been shown to increase its activity during early learning [18,35,36], and thus it has been suggested to play a key role in developing explicit strategies during learning [37,38]. Critically, however, for the Explicit neural axis we actually observed the highest loadings within the default mode network (DMN); specifically, within the DMN-B subnetwork, comprising higher-order regions in the middle and inferior frontal gyrus, angular gyrus, and anterior and middle temporal cortex (see barplots in Fig 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Explicit, performance, and implicit neural axes.

(Left) Barplots show mean coefficients from our penalized ridge regression model for each subnetwork [30]. (Right) The most prominent subnetworks for each neural axis are displayed on cortical surfaces. Loadings for cerebellar and striatal regions are shown at the right; note that colormaps for each surface are scaled individually. The data and code needed to generate this figure can be found in https://zenodo.org/records/14029185.

https://doi.org/10.1371/journal.pbio.3002934.g003

Note that, due to the sparsity of the coefficients returned by our models, we found that most cortical networks loaded only on a single axis (e.g., the DAN and DMN networks loaded almost exclusively on the Implicit and Explicit neural axes, respectively; Fig 4A). However, as a noteworthy departure from this general observation, we found that the frontoparietal control network exhibited substantial cross-loadings on the different neural axes. For example, Fig 4A (bottom right) shows the implicit/explicit loadings of individual regions within the Control subnetworks. As can be seen in this plot, it was solely the Control-A and Control-C networks that showed meaningful cross-loadings across both axes, with the posterior cingulate cortex (PCC) and precuneus regions, in particular, being strongly explicit-aligned. In this way, a particular brain region can be described as being either explicit- or implicit-aligned if it has a greater loading on the Explicit or Implicit neural axis, respectively (we return to this idea in the next section).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Alignment of individual cortical regions along the 3 separate neural axes.

(A) Relationship between the loadings of individual cortical regions for the Explicit, Performance, and Implicit neural axes. Note that only Control regions showed meaningful cross-loading between axes (highlighted by ellipse). These regions are shown separately in the bottom right panel. Control C subregions (posterior cingulate and precuneus) in particular were strongly aligned to the explicit axis. (B) Loadings of individual brain regions for each neural axis, plotted according to their positioning along the Principal gradient. (C) Surface maps of the 3 neural axes. Red colors denote a positive loading and blue colors denote a negative loading. (D) Relationship between the Principal gradient and Explicit-alignment axis. Plot at bottom shows the null distribution of spatial correlations between the Principal gradient and Explicit-alignm¹e²nt axis, as derived from 1,000 iterations of the spatial autocorrelation-preserving null model [39] implemented by the Neuromaps toolbox [40]. The dashed red vertical line denotes the true correlation value (p = 0.028). The data and code needed to generate this figure can be found in https://zenodo.org/records/14029185.

https://doi.org/10.1371/journal.pbio.3002934.g004

The learning-related neural axes map onto different levels of functional cortical organization

In order to better understand the large-scale organization of the different neural axes, we took the loadings of individual regions on each axis (Fig 4, panel B) and plotted them according to their position along the principal gradient of functional brain organization derived by [31] (Fig 4, panel C, see panel D for a visualization of the principal gradient). This principal gradient, which spans from unimodal cortex (i.e., visual and somatomotor networks) to transmodal cortex (i.e., the DMN), signifies a hierarchy in cortical processing from lower- to higher-order systems [17,31,41]. As can be seen in Fig 4, panel C, regions loading on the implicit axis tend to be situated lower on the principal gradient (i.e., closer to sensory and motor regions), while the performance axis comprises regions closer to transmodal cortex. By contrast, the explicit axis prominently features regions of the default mode network, located neuroanatomically furthest from sensorimotor cortex.

Next, to more directly assess the relationship between our learning-derived axes and this principal gradient of functional cortical organization, we derived a single summary axis—the “explicit alignment” axis—by computing the difference in rank loadings for each ROI (Fig 4D). According to the construction of this summary axis, regions with positive values indicate a relatively greater loading on the explicit, compared to implicit, neural axis, and vice versa for negative values. Consistent with this, the greatest explicit alignment is observed in the DMN-B and Control-C subnetworks, whereas the least explicit (most implicit) alignment is observed in the superior parietal and premotor areas of the DAN-B subnetwork. Notably, we found this explicit alignment axis was significantly correlated with the aforementioned principal gradient of functional brain organization (p = 0.028; spin permutation test: [42], Fig 4D). This suggests that explicit learning is supported by brain areas (i.e., transmodal regions) positioned at the very apex of the cortical processing hierarchy, whereas implicit adaptation is supported by brain areas (superior parietal and premotor regions) located directly adjacent to sensorimotor cortex.

Expression of the explicit neural axis predicts learning in a separate reward-based motor task

Implicit adaptation in the VMR task is thought to primarily reflect the updating, through sensory prediction errors, of an internal forward model that predicts the sensory consequences of a motor command [5,7]). In other forms of motor learning, where subjects are denied this sensory error feedback, this same kind of updating is impossible, and so learning in such cases must rely on alternative neural systems and pathways [43,44]. By contrast, explicit learning has been linked to various executive functions [18,38,45,46], and the DMN—which features prominently in our Explicit neural axis—is suggested to specialize in cognitive computations that are independent of specific perceptual inputs [41]. We thus reasoned that the Explicit neural axis might generalize across motor learning tasks that provide different forms of sensory feedback, whereas the implicit neural axis would not.

To test this idea, in a different fMRI session we had our same participants perform a separate motor learning task in which they were required to alter their hand movements purely through reward-based feedback. In this task, subjects used their right finger on the touchpad to trace—without visual feedback of their cursor—a rightward-curved path displayed on the screen (from a start position to target line). Following a baseline block (70 trials), in which subjects did not receive any feedback about their performance, they performed a learning block (200 trials). Here, they were instructed that they would now receive reward score feedback, between 0 and 100, based on how accurately they traced the visible path, and that they should attempt to maximize their score across trials. Critically—and unbeknownst to the subjects—the reward score they received was actually based on how accurately they traced the (hidden) mirror-image path (reflected across the vertical axis; i.e., they would receive an imperfect score if they traced the visible path but would receive a 100 score if they perfectly traced the mirror image path; Fig 5A). Importantly, as the cursor was invisible, subjects could not use error-based learning mechanisms to improve their performance (as they could in the sensorimotor adaptation task). Rather, they could use only the scalar reward feedback, presented at the end of each trial, to refine their movements over time.

Learning in such tasks—i.e., where behavior must be guided solely by scalar feedback—is known to be highly explicit [19,20,47–49]. Consistent with this view, we found that the vast majority of subjects displayed clearly defined learning periods during our task, book-ended by periods of relatively flat performance (see example subjects in Fig 5). To identify these learning periods in a data-driven fashion, on each trial we computed a movement score quantifying the overall leftward (towards the invisible target) or rightwardness (towards the trained path) of the trajectory, and fit a sigmoid function to these scores across trials. These resulting fits allowed us to define separate pre-learning, learning, and post-learning periods (see Fig 5C; see Methods and see fits to individual subjects in S4 Fig). We then asked whether the expression of the Explicit neural axis—specifically during the learning period—was related to subjects’ rate of learning. To test this, we performed a two-part analytical procedure: (1) For each subject, we computed the spatial correlation between each of our learning-related neural axes extracted from the VMR task (i.e., the Implicit, Performance, and Explicit axes) with the observed patterns of functional connectivity during each of the separate periods (pre-learning, learning, and post-learning) associated with the reward-based learning task. [Note that these spatial correlations provide us with a direct index of the degree to which each neural axis from the sensorimotor adaptation task is “expressed” during the different periods of reward-based learning, and we thus call these correlations the explicit (resp. performance, implicit) axis scores.] (2) We examined whether these newly derived axis scores were correlated with subjects’ learning rates. The results of this analysis are shown in Fig 6B.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Reward-based motor task and subject performance.

(A) Subjects learned to trace an invisible target trajectory with only score feedback. After learning to trace a rightward curved path (baseline trials), score feedback was introduced indicating (unbeknownst to the subject) their success in tracing the mirror image path (learning trials). (B) Spline-smoothed scores for individual subjects. Note the considerable heterogeneity in subjects’ learning, with some subjects beginning to learn immediately after score onset (as was the case in the VMR task) versus some subjects showing no improvement until much later. (C) Behavioral data from an example subject. (Top) Score across trials. (Bottom) For each trial, we quantified the overall leftward or rightwardness of the response by taking the average x-position of the movement trajectory, with positive values indicating that the response was more towards the invisible target. We described this value as the lateral movement score. Most subjects showed clearly defined learning periods, book-ended by periods of stable performance. We identified these periods by fitting a sigmoid function to each subject’s movement scores in order to identify these periods. The 0.025, 0.05, and 0.95 quantiles of the resulting curves were used to define the end of the pre-learning, beginning of the learning, and beginning of the post-learning periods, respectively. In addition, the parameters of the fitted sigmoid were used to characterize subjects’ performance. (D) Distributions of fitted model parameters across subjects. L denotes the lower asymptote, and L + U denotes the difference between the upper and lower asymptotes, respectively; m denotes the midpoint of the learning period; and k denotes the learning rate. (E) Relationship between subjects’ reaction time (RT, left) and movement time (MT, right) as a function of learning period. Data shows the group mean +/- 1 standard error of the mean. Note that subjects’ RTs and MTs tended to decrease during the post-learning period, consistent with the responses becoming more automatic once subjects have converged upon a stable solution to the task. The data and code needed to generate this figure can be found in https://zenodo.org/records/14029185.

https://doi.org/10.1371/journal.pbio.3002934.g005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Neural results for the separate reward-based motor learning task.

(A) We examined whole-brain functional connectivity with motor cortex during each period of the reward-based motor task (pre-learning, learning, and post-learning) and asked whether learning-related increases in the expression of the Explicit neural axis was associated with better learning (i.e., k parameters from Fig 5). For each period of the reward-based motor task, we computed functional connectivity with motor cortex and then examined the correlation between these patterns of functional connectivity, and the 3 neural axes derived from the VMR task (shown in Fig 4). These correlations served as measures of the expression of the Explicit, Performance, and Implicit axes during reward-based learning. (B) Learning rates, derived from subjects’ sigmoid fits, were significantly correlated with the increase in Explicit axis expression during learning, but not with the Implicit or Performance axis expressions. As a control, we repeated our analysis, including the derivation of the VMR axes, using functional connectivity with the ipsilateral (to the response hand) tongue and hand areas of the somatomotor network, as well as with the contralateral tongue area. In each case, the correlation with learning rate was abolished. The data and code needed to generate this figure can be found in https://zenodo.org/records/14029185.

https://doi.org/10.1371/journal.pbio.3002934.g006

Consistent with our predictions, we observed a significant positive correlation between the expression of the Explicit neural axis (derived from the VMR task) during the learning period of the reward-based task, with subjects’ learning rates. That is, the larger the increase in the expression of the explicit component during the learning phase, the more rapidly subjects learned the hidden path (r = 0.34, t₃₄ = 2.04, p = 0.049). Critically, as an important control analysis, we found that this across-task prediction was unique to the Explicit neural axis, as we observed no such relationship for either of the Implicit (r = −0.12, t₃₄ = −0.71, p = 0.48) or Performance (r = −0.05, t₃₄ = −0.29, p = 0.77) neural axes during the same learning period. Moreover, as a further control, we also observed no correlation between the expression of the Explicit axis and performance in the reward learning task when we separately examined both the pre-learning (r = −0.06, t₃₄ = −0.34, p = 0.74) and post-learning periods of the task (r = −0.18, t₃₄ = −1.01, p = 0.32). Together, these results strongly suggest that the brain systems comprising our Explicit axis are recruited specifically during the learning period itself—when subjects are actively adjusting their behavior—rather than in the preceding and subsequent periods of relatively stable performance, respectively.

One alternative explanation of the above results is that, rather than the Explicit neural axis uniquely reflecting a modulation of motor regions involved in response generation, they instead reflect a more global, whole-brain change in functional connectivity. To test this alternative explanation of the results, we performed an additional set of control analyses in which we repeated the construction of the 3 axes (Explicit, Implicit, and Performance) using patterns of functional connectivity with the contra- and ipsilateral (to the response hand) tongue area of somatomotor cortex, as well as the ipsilateral hand area during the VMR task. Critically, in none of these cases did we observe any significant correlation between axis expression and performance in the reward-based learning task (Fig 6, panel B). [Note that, although the correlation between Explicit axis expression and learning rate was not significant in the right hand area, neither was the difference in correlation between the 2 hemispheres significant.]

Another alternative explanation of our main results is that Explicit neural axis, instead of reflecting processes related to the learning of an explicit strategy, relates purely to the implementation of the learned strategy (e.g., mental rotation of the response). If this were true, then we would expect to also observe the same pattern of across-task prediction effects when deriving an Explicit neural axis using instead the late learning epoch of the VMR task (the last 4 blocks of 8 trials). Indeed, during this late learning epoch subjects’ performance has plateaued and their RTs are decreased (Fig 1D), consistent with an increased automaticity in the motor responses. To test this idea, we again repeated the construction of the 3 neural axes using the patterns of functional connectivity during the late learning epoch of the VMR task, and again found no correlation between any of these axes’ expression and performance in the reward-based learning task (S6A Fig). Note also that the explicit axis derived from subjects’ late learning periods was markedly different from its early learning equivalent (S6B and S6C Fig). During late learning, the explicit axis prominently featured regions of the dorsal attention and visual networks—in particular, premotor and posterior parietal regions, as well as central and peripheral visual regions. Although we do not directly assay the specific strategy used by explicit subjects, mental rotation strategies are commonly used by explicit learners in the VMR task [23], and regions of the premotor, posterior parietal, and visual cortex are frequently implicated in mental rotation [50]. This interpretation would be consistent with the failure to observe a correlation between the expression of the late explicit axis and performance in the reward-based learning task, as mental rotation is unlikely to support learning in the reward-based task. Thus, we interpret this late explicit axis as reflecting the implementation of a strategy specific to the VMR task (e.g., mental rotation), while the early explicit axis likely involves components related to the learning of a strategy itself.

Collectively, these control analyses indicate that it is the expression of the Explicit neural axis, and not the Implicit or Performance neural axes, that predicts learning in a separate motor task that does not afford learning via sensory prediction errors. Moreover, these additional analyses demonstrate an important selectivity to this effect—specifically, we find that the across-task prediction is exclusive to (1) subjects’ individual learning phase (and not their pre- and post-learning phases); and (2) connectivity with the contralateral hand area of motor cortex, indicating that these patterns of connectivity are both neuroanatomically and behaviorally relevant.

Participants

Forty-six right-handed subjects (27 females, aged 18 to 28 years, mean age: 20.3 years) participated in 3 separate testing sessions, each spaced approximately 1 week apart: the first, an MRI-training session, was followed by 2 subsequent MRI experimental sessions. Of these 46 subjects, 10 participants were excluded from the final analysis [1 subject was excluded due of excessive head motion in the MRI scanner (motion greater than 2 mm or 2° rotation within a single scan); 1 subject was excluded due interruption of the scan during the learning phase of the reward-based motor task; 5 subjects were excluded due to poor behavioral performance in one or both tasks (4 of these participants were excluded because >25% of trials were not completed within the maximum trial duration; and one because >20% of trials had missing data due to insufficient pressure of the fingertip on the MRI-compatible tablet); 3 subjects were excluded due to a failure to properly perform the reward-based task (these subjects did not trace the visible rightward path during the baseline phase, but rather continued to trace a straight line for the entire duration of the task, suggesting that they did not attend to the task)]. Right-handedness was assessed with the Edinburgh handedness questionnaire [87]. Participants’ informed consent was obtained before beginning the experimental protocol. The Queen’s University Research Ethics Board approved the study and it was conducted in coherence to the principles outlines in the Canadian Tri-Council Policy Statement on Ethical Conduct for Research Involving Humans and the principles of the Deceleration of Helsinki (1964).

Procedure

In the first session, participants took part in an MRI-training session inside a mock (0 T) scanner that was made to look and sound like a real MRI scanner. We undertook this training session to (1) introduce participants to features of the VMR and reward-based motor tasks that would be subsequently performed in the MRI scanner; (2) ensure that subjects obtained baseline performance levels on those 2 tasks; and (3) ensure that participants could remain still for 1.5 h. To equate baseline performance levels across participants, subjects performed 80 trials per task. To train participants to remain still in the scanner, we monitored subjects’ head movement via a Polhemus sensor attached to each subject’s forehead (Polhemus, Colchester, Vermont). This allowed a real-time read-out of subject head displacement in the 3 axes of translation and rotation (6 dimensions total). Whenever subjects’ head translation and/or rotation reached 0.5 mm or 0.5° rotation, subjects received an unpleasant auditory tone, delivered through a speaker system located near the head. All of the subjects used in the study learned to constrain their head movement through this training regimen. Following this first session, subjects then subsequently participated in reward-based VMR motor tasks, respectively (see below for details).

Apparatus

During testing in the mock (0 T) scanner, subjects performed hand movements that were directed towards a target by applying fingertip pressure on a digitizing touchscreen tablet (Wacom Intuos Pro M tablet). During the actual MRI testing sessions, subjects used an MRI-compatible digitizing tablet (Hybridmojo LLC, California, United States of America). In both the mock and real MRI scanner, the target and cursor stimuli were rear-projected with an LCD projector (NEC LT265 DLP projector, 1,024 × 768 resolution, 60 Hz refresh rate) onto a screen mounted behind the participant. The stimuli on the screen were viewed through a mirror fixated on the MRI coil directly above the participants’ eyes, thus preventing the participant from being able to see their hand.

Sensorimotor adaptation task

To study sensorimotor adaptation, we used the well-characterized VMR paradigm ([21]; Fig 1). During the VMR task, participants performed baseline trials in which they used their right index finger to perform center-out target-directed movements. After these baseline trials, we applied a 45^° clockwise (CW) rotation to the viewed cursor, allowing investigation of VMR learning. Following this, we assessed participants’ re-aiming strategy associated with their learning.

Each trial started with the participant moving the cursor (3 mm radius cyan circle) into the start position (4 mm radius white circle) in the center of the screen by sliding the index finger on the tablet. To guide the cursor to the start position, a ring centered around the start position indicated the distance between the cursor and the start position. The cursor became visible when its center was within 8 mm of the center of the start position. After the cursor was held within the start position for 0.5 s, a target (5 mm radius red circle) was shown on top of a gray ring with a radius of 60 mm (i.e., the target distance) centered around the start position. The target was presented at one of 8 locations, separated by 45^° (0, 45, 90, 135, 180, 225, 270, and 315^°), in pseudorandomized bins of 8 trials. Participants were instructed to hit the target with the cursor by making a fast finger movement on the tablet. They were instructed to “slice” the cursor through the target to minimize online corrections during the reach. If the movement was initiated (i.e., the cursor had moved fully out of the start circle) before the target appeared, the trial was aborted and a feedback text message “Too early” appeared centrally on the screen. In trials with correct timing, the cursor was visible during the movement to the ring and then became stationary for 1 s when it reached the ring, providing the participant with visual feedback of their endpoint reach error. If any part of the stationary cursor overlapped with any part of the target, the target colored green to indicate a hit. Each trial was terminated after 4.5 s, independent of whether the cursor had reached the target. After a delay of 1.5 s, allowing time to save the data, the next trial started with the presentation of the start position.

During the participant training session in the mock MRI scanner (i.e., 2-weeks prior to the VMR MRI testing session), participants performed a practice block of 40 trials with veridical feedback (i.e., no rotation was applied to the cursor). This training session exposed participants to several key features of the task (e.g., use of the touchscreen tablet, trial timing, presence, and removal of cursor feedback) and allowed us to establish adequate performance levels.

At the beginning of the MRI testing session, but prior to the first scan being collected, participants re-acquainted themselves with the VMR task by performing 80 practice trials with veridical cursor feedback. Next, we collected an anatomical scan, followed by the baseline fMRI experimental run, wherein subjects performed 64 trials of the VMR task with veridical cursor feedback using their right hand. Next, they performed the learning scan, wherein subjects performed 160 trials in which feedback of the cursor during the reach was rotated clockwise by 45^°. Following this fMRI experimental run, participants were asked to report their strategic aiming direction (over 16 trials). In these trials, a line between the start and target position appeared on the screen at the start of each trial. Participants were asked to use a separate MRI joystick (Current Designs, Inc.) positioned at their left hip to rotate the line to the direction that they would aim their finger movement in for the cursor to hit the target, and click the button on the joystick box when satisfied. Following the button click, the trial proceeded as a normal reach trial. Specifically, the following text was displayed on the screen to participants at the onset of the reporting trials: “Use the joystick to rotate the line to the direction that you will aim your finger in for the cursor to hit the target. If you think your finger should aim straight at the target, then leave the line at the target and click the button on the joystick. If you think your finger should aim somewhere else for the cursor to hit the target, rotate the line to your aiming direction, and then click the button. Next, hit the target with your right finger.”

These 16 “report” trials were followed by 16 normal rotation trials. As several subjects expressed confusion about the nature of the report trials, often failing to adjust the line at all during the first few trials, we discarded the first 8 report trials for all subjects, and calculated the mean aim direction using the final 8 trials (note, however, that we observed very similar results whether estimates of the explicit and implicit components were based on these 8 versus all 16 report trials, see S2 Fig). Note that we had subjects perform these report trials following learning (as in [49]) given our prior behavioral work showing that the inter-mixing of report trials during learning can lead to more participants adopting an explicit, re-aiming strategy [22,49], thereby distorting participants’ learning curves. Participants were not informed about nature or presence of the visuomotor rotation before or during the experiment, and received no “coaching” about how to specifically perform the reporting trials (we adopted this “hands-off” approach given our concern that any experimenter instructions or invention might bias subjects’ learning and reporting behavior).

Note that all subjects performed the VMR task after having performed the reward-based motor task (i.e., testing order was not counterbalanced across the RL and EL tasks). We made this decision in light of our knowledge that subjects may alter their behavior after performing explicit report trials in the VMR task [22] (i.e., often developing explicit awareness of the perturbation and using that to guide their performance). We were concerned that, had subjects performed the VMR task first, they may have anticipated some experimental manipulation (or deception), and thus would not have approached the reward-based motor task in a naive fashion. From this perspective, our approach of having defined the explicit and implicit neural axes on the (second) VMR task and then using those neural axes to predict performance on the (first) reward-based motor task, strengthens our interpretations of the effects.

Reward-based learning task

In the reward-based motor task (Fig 5, left), participants learned to produce, through reward-based feedback, finger movement trajectories with a specific (unseen) shape. Specifically, subjects were instructed to repeatedly trace, without visual cursor feedback of their actual finger paths, a subtly curved path displayed on the screen (the visible path). During learning trials, participants were told that, following each trial, they would receive a score based on how accurately they traced the visible path (and were instructed to maximize points across trials). However, unbeknownst to them, they actually received points based on how well they traced the mirror-image path (the target path). Critically, because participants received no visual feedback about their actual finger trajectories or the “rewarded” shape, they could not use error-based information to guide learning. Our task was inspired by the motor learning tasks developed by [48,82].

Each trial started with the participant moving the cursor (3 mm radius cyan circle) into the start position (4 mm radius white circle) at the bottom of the screen by sliding their index finger on the tablet. The cursor was only visible when it was within 30 mm of the start position. After the cursor was held within the start position for 0.5 s, the cursor disappeared and a curved path and a target distance marker appeared on the screen. The target distance marker was a horizontal red line (30 × 1 mm) that appeared 60 mm above the start position. The visible path connected the start position and target distance marker and had the shape of a half sine wave with an amplitude of 0.15 times the target distance. Participants were instructed to trace the curved visible path. When the cursor reached the target marker distance, the marker changed color from red to green to indicate that the trial was completed. Importantly, participants did not receive feedback about the position of their cursor during the trial.

In the baseline block, participants did not receive feedback about their performance. In the learning block, participants were rewarded 0 to 100 points after reaching the target distance marker, and participants were instructed to do their best to maximize this score across trials (the points were displayed as text centrally on the screen). They were told that to increase the score, they had to “trace the line more accurately.” Each trial was terminated after 4.5 s, independent of whether the cursor had reached the target. After a delay of 1.5 s, allowing time to save the data, the next trial started with the presentation of the start position.

To calculate the score on each trial in the learning block, the x position of the cursor was interpolated at each cm displacement from the start position in the y direction (i.e., at exactly 10, 20, 30, 40, 50, and 60 mm). For each of the 6 y positions, the absolute distance between the interpolated x position of the cursor and the x position of the rewarded path (mirror image of visible path) was calculated. The sum of these errors was scaled by dividing it by the sum of errors obtained for a half cycle sine-shaped path with an amplitude of 0.5 times the target distance, and then multiplied by 100 to obtain a score ranging between 0 and 100. The scaling worked out so that a perfectly traced visible path would result in an imperfect score of 40 points. This scaling was chosen on the basis of extensive pilot testing in order to achieve variation in subject performance, and to ensure that subjects still received informative score feedback when tracing in the vicinity of the visible trajectory.

During the participant training session in the mock MRI scanner (i.e., 1-week prior to the MRI testing session), participants only performed a practice block in which they traced a straight line with (40 trials) and then without (40 trials) visual feedback of the position of the cursor during the reach. As with the VMR task, this training session exposed participants to several key features of the task (e.g., use of the touchscreen tablet, trial timing, used of cursor feedback to correct for errors) and allowed us to establish adequate baseline performance levels. Importantly, however, this training session did not allow for any reward-based learning to take place.

At the beginning of the MRI testing session, but prior to the first scan being collected, participants re-acquainted themselves with the reward-based motor task by first performing a practice block in which they traced a straight line with (40 trials) and then without (40 trials) visual feedback of the position of the cursor during the reach. Next, we collected an anatomical scan, following by a DTI scan, followed by a resting-state fMRI scan. During the latter resting-state scan, participants were instructed to rest with their eyes open while fixating on a central cross location presented on the screen. Following this, participants performed the reward-based motor task, which consisted of 2 separate experimental runs without visual feedback of the cursor: (1) a baseline block of 70 trials in which they attempted to trace the curved visible path and no score was provided; and (2) a separate learning block of 200 trials in which participants were instructed to maximize their score shown at the end of each trial.

Behavioral data analysis

MRI acquisition

Participants were scanned using a 3-Tesla Siemens TIM MAGNETOM Trio MRI scanner located at the Centre for Neuroscience Studies, Queen’s University (Kingston, Ontario, Canada). Functional MRI volumes were acquired using a 32-channel head coil and a T2*-weighted single-shot gradient-echo echo-planar imaging (EPI) acquisition sequence (time to repetition (TR) = 2,000 ms, slice thickness = 4 mm, in-plane resolution = 3 mm × 3 mm, time to echo (TE) = 30 ms, field of view = 240 mm × 240 mm, matrix size = 80 × 80, flip angle = 90°, and acceleration factor (integrated parallel acquisition technologies, iPAT) = 2 with generalized auto-calibrating partially parallel acquisitions (GRAPPA) reconstruction. Each volume comprised 34 contiguous (no gap) oblique slices acquired at a 30° caudal tilt with respect to the plane of the anterior and posterior commissure (AC-PC), providing whole-brain coverage of the cerebrum and cerebellum. Each of the task-related scans included an additional 8 imaging volumes at both the beginning and end of the scan. On average, each of the MRI testing sessions lasted 2 h.

At the beginning of the reward-based motor task MRI testing session, a T1-weighted ADNI MPRAGE anatomical was also collected (TR = 1,760 ms, TE = 2.98 ms, field of view = 192 mm × 240 mm × 256 mm, matrix size = 192 × 240 × 256, flip angle = 9°, 1 mm isotropic voxels). This was followed by a series of Diffusion-Weighted scans, wherein we acquired 2 sets of whole-brain diffusion-weighted volumes (30 directions, b = 1,000 s mm-2, 65 slices, voxel size = 2 × 2 × 2 mm3, TR = 9.3 s, TE = 94 ms) plus 2 volumes without diffusion-weighting (b = 0 s mm-2). Next, we collected a resting-state scan, wherein 300 imaging volumes were acquired. For the baseline and learning scans during the motor task, 222 and 612 imaging volumes were acquired, respectively.

At the beginning of the VMR MRI testing session, we gathered high-resolution whole-brain T1-weighted (T1w) and T2-weighted (T2w) anatomical images (in-plane resolution 0.7 × 0.7 mm2; 320 × 320 matrix; slice thickness: 0.7 mm; 256 AC-PC transverse slices; anterior-to-posterior encoding; 2 × acceleration factor; T1w TR 2,400 ms; TE 2.13 ms; flip angle 8°; echo spacing 6.5 ms; T2w TR 3,200 ms; TE 567 ms; variable flip angle; echo spacing 3.74 ms). These protocols were selected on the basis of protocol optimizations designed by [89]. Following this, for the baseline and learning functional scans, 204 and 492 imaging volumes were acquired, respectively.

fMRI preprocessing

Results included in this manuscript come from preprocessing performed using fMRIPrep 1.4.1 ([90,91]; RRID:SCR_016216), which is based on Nipype 1.2.0 ([92,93]; RRID:SCR_002502).

Construction of the VMR learning-related neural axes and prediction of reward-based motor performance.

Here, we describe our approach for the main analysis—i.e., construction of the (performance corrected) explicit neural axis derived from the contralateral hand area, during the early learning epoch.

For each tangent vector in the early learning epoch, we computed the mean connectivity between the contralateral hand area (within the somatomotor network) and the remaining ROIs. The resulting vectors (containing an entry for each non-motor ROI) were used as the features for our predictive model.

Call the resulting matrix of predictors X = [x₁, x₂,…, x_n], where x_i is a vector containing the mean connectivity of the i’th ROI with the hand area of motor cortex. We then computed the explicit learning score for each subject, defined as the circular mean difference between the target location and the subject’s reported aim direction. These scores formed the response variable y. Our corrected scores were then derived by orthogonalizing with respect to subjects’ VMR performance scores. That is, we regressed y onto the vector of VMR mean early error and used the residuals for our analysis, and the same was done for each predictor (that is, each column in X). Both the predictors and outcome were then standardized prior to analysis.

We then fit a ridge regression model with separate network-level penalty terms using the model described by [34], and implemented in the R package GRridge [116]. That is, each of the 17 networks comprising the predictors—15 non-motor cortical networks, plus cerebellum and basal ganglia—possessed their own shrinkage parameters. These parameters were separately tuned using an internal leave-one-out cross-validation loop, so that networks whose regions contributed relatively little to overall predictive performance were shrunken more harshly relative to predictively useful networks. The precise arguments were as follows:

The resulting parameter vector β contains, for each ROI, its estimated conditional contribution to subjects’ explicit report.

We then derived, for each subject, the mean functional connectivity of each region with the motor cortex during the learning period of the reward-based motor task (derived from the centered tangent vectors, as described above). The Pearson correlation between this vector and the vector β was the subject’s Explicit neural axis score.

Note that all other models (such as the derivation of the implicit or performance neural axis), or neural axes using separate areas of motor cortex (e.g., tongue areas) or different time windows (e.g., baseline or pre-learning) were constructed identically, mutatis mutandis (e.g., using implicit scores rather than explicit reports).

Behavioral control study

Fourteen right-handed paid volunteers (8 females; mean age: 22.3 years) participated in a behavioral control study conducted outside the MRI scanner. This study adhered to the same research ethics approvals and consent procedures as the MRI study. In this behavioral testing, participants performed a VMR task similar in structure and number of trials to the MRI testing session, except for changes in the experimental setup and the inclusion of intermittent reporting trials during the learning phase.