Training at asymptote stabilizes motor memories by reducing intracortical excitation

Abstract

Learning similar motor skills in close succession is limited by interference, a phenomenon that takes place early after acquisition when motor memories are unstable. Interference can be bidirectional, as the first memory can be disrupted by the second (retrograde interference), or the second memory can be disrupted by the first (anterograde interference). The heightened plastic state of primary motor cortex after learning is thought to underlie interference, as unstable motor memories compete for neural resources. While time-dependent consolidation processes reduce interference, the passage of time (~6 h) required for memory stabilization limits our capacity to learn multiple motor skills at once. Here, we demonstrate in humans that prolonged training at asymptote of an initial motor skill reduces both retrograde and anterograde interference when a second motor skill is acquired in close succession. Neurophysiological assessments via transcranial magnetic stimulation reflect this online stabilization process. Specifically, excitatory neurotransmission in primary motor cortex increased after short training and decreased after prolonged training at performance asymptote. Of note, this reduction in intracortical excitation after prolonged training was proportional to better skill retention the following day. Importantly, these neurophysiological effects were not observed after motor practice without learning or after a temporal delay. Together, these findings indicate that prolonged training at asymptote improves the capacity to learn multiple motor skills in close succession, and that downregulation of excitatory neurotransmission in primary motor cortex may be a marker of online motor memory stabilization.

1. Introduction

The ability to acquire and retain motor skills is critical to everyday life. However, learning similar skills in close succession is limited by interference, a phenomenon that takes place early after skill acquisition (Brashers-Krug et al., 1996; Shadmehr & Brashers-Krug, 1997; Walker et al., 2003; Krakauer et al., 2005; Ghilardi et al., 2009; Cantarero, Tang, et al., 2013; Lerner et al., 2020). Retrograde interference refers to disruption of the first memory by the second, whereas anterograde interference refers to disruption of the second memory by the first. Although the susceptibility of motor memories to bidirectional interference decreases through time-dependent consolidation (Brashers-Krug et al., 1996; Shadmehr & Brashers-Krug, 1997; Walker et al., 2003; Krakauer et al., 2005; Cantarero, Tang, et al., 2013; Lerner et al., 2020), the passage of time (~6 h) required for memory stabilization limits our capacity to learn multiple motor skills in series.

Previous studies have indicated that prolonged training at performance asymptote of an initial motor memory can reduce retrograde interference (Krakauer et al., 2005; Ghilardi et al., 2009). It has been proposed that consolidation mechanisms may be initiated online during prolonged training (Krakauer et al., 2005), offering a temporally efficient method for stabilizing motor memories. A recent study showed that a long inter-acquisition interval (~6 h) is required to prevent anterograde intereference of a second motor memory by a preceding memory (Lerner et al., 2020). Whether stabilizing an initial motor memory through asymptote training can also reduce anterograde interference, critical to learning multiple motor skills, and the neural mechanisms underlying this online stabilization effect are unclear.

Evidence from animal and human studies indicates that motor skill learning elicits long-term potentiation (LTP)-like plasticity within primary motor cortex (M1) (Rioult-Pedotti et al., 1998, 2000; Ziemann et al., 2004; Rosenkranz et al., 2007). LTP is characterized by increased efficacy of excitatory synapses and a reduction in inhibitory tone (Feldman, 2009). It has been proposed that the heightened plastic state of M1 after learning may underlie the susceptibility of newly acquired motor skills to retrograde interference, as well as impair the learning of a subsequent motor skills (i.e., anterograde interference), due to competition for neural resources between unstable memories (Cantarero et al., 2013a, 2013b). In visual perceptual learning, prolonged training reduces the weighting of excitatory processing in primary visual cortex and protects perceptual memories from retrograde interference (Shibata et al., 2017). Due to the prominent role M1 plays in the consolidation of motor skills (Muellbacher et al., 2002; Censor et al., 2010; Kantak et al., 2010; Cantarero, Lloyd, & Celnik, 2013), training duration-dependent changes in excitatory and inhibitory neurotransmission within M1 may underlie the stability of newly acquired motor memories.

Here we sought to determine whether training an initial motor skill (skill A) at performance asymptote, would reduce both retrograde and anterograde interference when a subsequent motor skill (skill B) is acquired in close succession. We also sought to assess whether training duration-dependent modulation of inhibitory and excitatory neurotransmission within M1 underlies motor skill stability. We hypothesized that prolonged training of skill A at performance asymptote, before training on skill B, would reduce both retrograde interference of skill A and anterograde interference of skill B. Furthermore, this online stabilization effect would be associated with reduced excitatory or increased inhibitory neurotransmission in M1.

2. Materials and methods

2.1. Participants

Sixty healthy young adults (21 males; age range 18–35 years, mean age ± SD: 24 ± 4 years) participated in the study. We screened participants prior to enrollment for contraindications to transcranial magnetic stimulation (TMS) using a questionnaire. We assessed handedness (52 right, 8 left) using the Edinburgh Handedness Inventory (Oldfield, 1971). The Johns Hopkins School of Medicine Institutional Review Board, in accordance to the declaration of Helsinki, approved the study and participants provided written informed consent.

2.2. Experimental design

For the main experiment, participants completed two sessions, an acquisition session and a retention session, on consecutive days. In the acquisition session, each participant learned an established sequential visuomotor isometric pinch task (Reis et al., 2009; Cantarero, Tang, et al., 2013; Spampinato & Celnik, 2017). There were two configurations of the task, skill A and skill B (Fig. 1A and B). We randomly assigned participants to one of four groups (Fig. 1C). Sample sizes were determined based on estimated effect sizes from previous motor learning studies demonstrating reduced retrograde interference with prolonged training (Krakauer et al., 2005; Ghilardi et al., 2009). With desired power of .8 and alpha of .05, a sample size of 8 per group was required. Thus, we conservatively tested n = 10 per group. Groups ${\text{A}}_{\text{SHORT}}$ and ${\text{A}}_{\text{LONG}}$ trained on skill A for five and ten blocks (30 trials/block) respectively. Groups ${\text{A}}_{\text{SHORT}}+\text{B}$ and ${\text{A}}_{\text{LONG}}+\text{B}$ trained on skill A for five and ten blocks respectively, then immediately trained on skill B for five blocks. We acquired neurophysiological data before and after training on skill A using TMS. Specifically, intracortical facilitation (ICF) and short-interval intracortical inhibition (SICI) were used to index excitatory and inhibitory neurotransmission within M1 respectively (Ziemann et al., 2015). In the retention session, each participant completed one block (30 trials) of skill A and skill B. To control for the mere effects of practice on neurophysiological variables, another group (n = 10) trained on a random version of the task, in which learning does not accumulate (Cantarero, Tang, et al., 2013; Spampinato & Celnik, 2017), for ten blocks. To control for the additional time spent training on skill A by ${\text{A}}_{\text{LONG}}+\text{B}$ compared with ${\text{A}}_{\text{SHORT}}+\text{B}$, another group $({\text{A}}_{\text{SHORT}}+{\text{B}}_{\text{DELAY}};\text{n}=10)$ trained on skill A for five blocks, then trained on skill B for five blocks after a 30 min delay.

Fig. 1 –

Experimental design. (A) Depiction of the sequential visuomotor isometric pinch task for skill A and skill B. The black square on the left represents the cursor within the home position. Participants learned to control the movement of the cursor from the home position to each of the targets in the order home-1-home-2-home-3-home-4-home-5. (B) Logarithmic (solid) and exponential (dash) mapping between pinch force and cursor position applied for skill A and skill B respectively. (C) Training was performed on day one. The ${\text{A}}_{\text{SHORT}}$ and ${\text{A}}_{\text{LONG}}$ groups (both n = 10) trained on skill A for five and ten blocks (30 trials/block) respectively. The ${\text{A}}_{\text{SHORT}}+\text{B}$ and ${\text{A}}_{\text{LONG}}+\text{B}$ groups (both n = 10) trained on skill A for five and ten blocks respectively, then trained on skill B for five blocks. Neurophysiological data were obtained before and immediately after training on skill A using transcranial magnetic stimulation (TMS). Retention of skill A and skill B was assessed 24 h later on day two with one block of 30 trials. To control for the mere effects of practice on neurophysiological variables, the random group (n = 10) trained on a random version of the pinch task, in which learning does not accumulate, for ten blocks. To control for the additional time spent training on skill A by the ${\text{A}}_{\text{LONG}}+\text{B}$ group compared with ${\text{A}}_{\text{SHORT}}+\text{B}$, the ${\text{A}}_{\text{Short}}+{\text{B}}_{\text{DELAY}}$ group (n = 10) trained on skill A for five blocks, then trained on skill B for five blocks after a 30 min delay.

2.3. Sequential visuomotor isometric pinch task

As previously described (Cantarero, Tang, et al., 2013; Reis et al., 2009; Spampinato & Celnik, 2017), participants sat in front of a computer screen and held a LMD300 force transducer (Futek, Irvine, CA) between the thumb and index finger of their dominant hand. Pinching the transducer generated horizontal displacement of an on-screen cursor. The goal was to move the cursor between a home position and five targets in a specific sequence (home-1-home-2-home-3-home-4-home-5). The target positions were different for skill A and skill B (Fig. 1A), however the sequence remained the same. Participants were made explicitly aware of the target sequence before beginning training and the target numbers were displayed at all times. Learnable logarithmic and exponential transformations were also applied to the relationship between applied force and cursor movement for skill A and B respectively (Fig. 1B). For the random version of the task, the target positions and sequence was the same as skill A. However, for each trial, a different transformation was applied to the relationship between applied force and cursor movement.

We instructed participants to complete the trials at a self-selected pace, with the aim of completing each trial as quickly and as accurately as possible. A trial was considered correct only if the cursor endpoint was within each respective target and in the correct order. Therefore, accuracy for a trial was binary (i.e., error or no error), regardless of the number of errors committed. Whilst both target inaccuracy and misidentification of the sequence order could contribute to trial error, we did not separate these error sources as it is difficult to disambiguate under- or over-shooting a respective target from a sequencing error. Participants completed 30 trials per block, with rest periods of 1 min provided between blocks to prevent fatigue.

2.4. Surface electromyography

We recorded surface electromyography (EMG) from the dominant first dorsal interosseous (FDI) using 25 mm square Ag–AgCl recording electrodes (Vermed, Buffalo, NY), arranged in a belly-tendon montage, with a ground electrode positioned on the styloid process of the ulna. EMG signals were amplified (1000 ×) and band-pass filtered (10–1000 Hz) using an AMT-8 amplifier (Bortec Biomedical Ltd, Alberta, Canada), sampled at 5 kHz using a Micro1401-4 interface (CED, Cambridge, England) and recorded with Signal software (Version 4.02; CED, Cambridge, England).

2.5. Transcranial magnetic stimulation

We applied single- and paired-pulse TMS to dominant M1. A 70 mm-diameter figure-of-eight coil, connected to two monophasic Magstim 200² magnetic stimulators via a Bistim² module (Magstim, Whitland, Wales), was held tangentially to the scalp (~45° to the mid-sagittal line) to induce posterioranterior current in the brain. We identified the optimal site to elicit a consistent motor evoked potential (MEP) in the dominant FDI, which was marked over M1 and kept constant throughout using Brainsight neuronavigation (Rogue Research, Montreal, Canada).

We determined active motor threshold (AMT) as the stimulus intensity required to elicit a MEP of at least 200 μV in amplitude, with FDI pre-activated (~10% of the participant’s perceived maximum voluntary contraction), in five out of ten trials. We assessed ICF and SICI by delivering subthreshold conditioning stimuli, set to 70, 80 and 90% AMT, 10, 1 and 3 ms prior to a suprathreshold test stimulus set to elicit an ~1 mV MEP at rest (Kujirai et al., 1993). We measured SICI at both at both 1 and 3 ms as these intervals are proposed to be mechanistically distinct (Stagg et al., 2011). Furthermore, we selected conditioning intensities that were below the level at which SICI at 3 ms is contaminated by short-interval intracortical facilitation (Peurala et al., 2008). If necessary, we adjusted the intensity of the test stimulus after training to maintain an ~1 mV nonconditioned MEP. Group mean AMT, test stimulus intensity and nonconditioned MEP amplitude can be found in Table 1. We delivered twelve trials for each condition (120 trials total) in a randomized order pre and post training. For the random group, ICF and SICI were assessed before and after five (post1) and ten (post2) blocks of training.

Table 1 –

Transcranial magnetic stimulation parameters.

	${\text{A}}_{\text{SHORT}}(\text{n}=20)$	${\text{A}}_{\text{LONG}}(\text{n}=20)$	Random (n = 10)	${\text{A}}_{\text{SHORT}}+{\text{B}}_{\text{DELAY}}(\text{n}=8)$
AMT (% MSO)	45 (1)	43 (2)	46 (2)	47 (3)
S1mV pre (% MSO)	64 (2)	61 (3)	60 (3)	61 (3)
NC MEP pre (mV)	1.01 (.07)	1.00 (.07)	1.01 (.12)	1.11 (.11)
S1mV post (% MSO)	64 (2)	57 (4)	61 (3)	61 (4)
NC MEP post (mV)	1.01 (.07)	1.11 (.07)	1.07 (.12)	1.00 (.11)
S1mV post2 (% MSO)	–	–	61 (4)	–
NC MEP post2 (mV)	–	–	1.04 (.15)	–

Note: Data are shown as group mean (SEM). AMT–active motor threshold, MSO–maximum stimulator output, NC–nonconditioned, MEP–motor evoked potential.

2.6. Data analysis

We quantified skill for each individual participant using the following function (Reis et al., 2009):

$$\mathit{\text{Skill}}=\frac{1-\mathit{\text{error}}\mathit{\text{rate}}}{\mathit{\text{error}}\mathit{\text{rate}}\left(\text{ln}(\mathit{\text{trial}}\mathit{\text{duration}}{)}^b\right)}$$

where error rate is the proportion of trials (out of 30) with at least one error, trial duration is the total time from the onset of cursor movement until target 5 was reached and $b$ is the dimension free parameter. We calculated average trial duration and error rate for each block of 30 trials and held the value of $b$ constant at 5.424 (Cantarero, Tang, et al., 2013; Reis et al., 2009; Spampinato & Celnik, 2017). To homogenize variance, the logarithm of the skill parameter was used as the skill measure (Reis et al., 2009). We quantified online learning as the skill measure difference between the last and first block of training on day one. We quantified retention as the skill measure difference between the block on day two and the last block on day one. Online learning and retention were also quantified for error rate and trial duration using the values at the same time points.

We also quantified the average number of targets hit and the cumulative distance error on each trial for each participant. The cumulative distance error was calculated by measuring the shortest distance from the cursor to the outer boundary of the corresponding target and summing them (Hardwick et al., 2017). Attempts falling within the boundary of the target thus had error = 0.

For neurophysiology data, we discarded trials contaminated by pre-stimulus EMG activity (root mean squared EMG >10 μV; 100 ms before stimulation) or with MEP amplitudes >2SD outside the mean. All participants had at least 10 valid trials per condition. We quantified ICF and SICI by expressing the conditioned MEP amplitude for each conditioning intensity as a ratio of the nonconditioned MEP amplitude. At baseline, the conditioning intensity which produced the greatest inhibition/facilitation for ICF, ${\text{SICI}}_1$ and ${\text{SICI}}_3$ was determined as the maximum for each participant. For the main experiment, we grouped participants according to whether they trained on skill A for five $({\text{A}}_{\text{SHORT}};\text{n}=20)$ or ten $({\text{A}}_{\text{LONG}};\text{n}=20)$ blocks. Due to technical issues, we were unable to collect post training data from two participants in the ${\text{A}}_{\text{SHORT}}+{\text{B}}_{\text{DELAY}}$ control group, therefore we excluded their neurophysiology data.

2.7. Statistical analysis

We performed all statistical analyses using SPSS software (Version 25; IBM, Armonk, NY). We assessed normality using Shapiro–Wilk’s test and homogeneity of variance using Levene’s test. Non-normal data were log transformed for statistical analysis.

We analyzed baseline performance, online learning and retention of skill, error rate and trial duration for skill A using separate one-way ANOVAs with GROUP (${\text{A}}_{\text{SHORT}},{\text{A}}_{\text{LONG}},{\text{A}}_{\text{SHORT}}+\text{B},{\text{A}}_{\text{LONG}}+\text{B},{\text{A}}_{\text{SHORT}}+{\text{B}}_{\text{DELAY}}$) as the between-subjects factor. We analyzed baseline performance, online learning and retention of skill, error rate and trial duration for skill B using separate one-way ANOVAs with GROUP (${\text{A}}_{\text{SHORT}}+\text{B},{\text{A}}_{\text{LONG}}+\text{B},{\text{A}}_{\text{SHORT}}+{\text{B}}_{\text{DELAY}}$) as the between-subjects factor. In the 20 participants who trained on skill A for ten blocks, we analyzed the average number of trials hit and the cumulative distance error on each trial for block five and and block ten using paired $t$-tests.

We analyzed ICF, ${\text{SICI}}_1$ and ${\text{SICI}}_3$ at the baseline maximum conditioning intensity using separate two-way repeated measures ANOVAs with GROUP (${\text{A}}_{\text{SHORT}},{\text{A}}_{\text{LONG}}$) as the between-subjects factor and TIME (pre, post) as the within-subjects factor. We used Pearson correlation analyses to determine associations between modulation of maximum ICF and retention of skill A.

For the random control group, we analyzed the skill measure using a repeated measures ANOVA with BLOCK (one, five, ten) as the within-subjects factor. We also analyzed ICF, ${\text{SICI}}_1$ and ${\text{SICI}}_3$ at the baseline maximum conditioning intensity using separate repeated measures ANOVAs with TIME (pre, post1, post2) as the within-subjects factor. For the ${\text{A}}_{\text{SHORT}}+{\text{B}}_{\text{DELAY}}$ control group, we analyzed ICF, ${\text{SICI}}_1$ and ${\text{SICI}}_3$ at the baseline maximum conditioning intensity using paired $t$-tests.

Post hoc comparisons were made using two-tailed independent samples and paired $t$-tests and were adjusted when necessary using the Holm-Bonferroni correction. The significance level was set at p < .05. For ANOVAs and $t$-tests, we report partial eta squared $\left({{\eta }^2}_{\text{p}}\right)$ and Cohen’s d as a measure of effect size respectively. Group data are presented as mean ± SEM.

3. Results

3.1. Training at asymptote reduces retrograde and anterograde interference between motor skills

For skill A, there was no difference in baseline performance between groups (${\text{F}}_{\mathrm{4,45}}=.38,\text{p}=.824,{{\eta }^2}_{\text{p}}=.03$; Fig. 2A). There was also no difference in online learning of skill A between groups (${\text{F}}_{\mathrm{4,45}}=.35,\text{p}=.841,{{\eta }^2}_{\text{p}}=.03$; Fig. 2B). In the 20 participants who trained on skill A for ten blocks, we found no difference in the average number of targets hit or the cumulative distance error on each trial between block five $(\text{targets}=3.56\pm .06;\text{distance}=.017\pm .003\text{a}.\text{u}.)$ and block ten $(\text{targets}=3.54\pm .06,{\text{t}}_{19}=.27,\text{p}=.789,\text{d}=-.06;\text{distance}=.017\pm .003\text{a}.\text{u}.,{\text{t}}_{19}=.19,\text{p}=.852,\text{d}=-.04)$. Together, these results indicate that prolonged training (blocks 6–10) in the ${\text{A}}_{\text{LONG}}$ and ${\text{A}}_{\text{LONG}}+\text{B}$ groups was completed at performance asymptote.

Fig. 2 –

Performance of skill A and skill B. Vertical dashed line denotes the separation between the end of training on day one and the retention assessments on day two. (A) The y-axis represents the skill measure and the x-axis shows training and retention epochs. (B) The bar graph shows group average online learning for skill A and skill B, with individual data represented by the open circles. Online learning is the skill measure difference between the last and first block of training on day one. (C) The bar graph shows group average retention for skill A and skill B. Retention is the skill measure difference between the block on day two and the last block on day one. Retention of skill A was worse for ${\text{A}}_{\text{SHORT}}+\text{B}$ compared with ${\text{A}}_{\text{SHORT}}$. Retention of skill A and skill B was greater for ${\text{A}}_{\text{LONG}}+\text{B}$ compared with ${\text{A}}_{\text{SHORT}}+\text{B}$ and ${\text{A}}_{\text{SHORT}}+{\text{B}}_{\text{DELAY}}$. Data are means ± SEM. *p < .05.

When looking at retention of skill A, we found a difference across groups $({\text{F}}_{\mathrm{4,45}}=4.98,\text{p}=.002,{{\eta }^2}_{\text{p}}=.31)$. Indeed, we found that retention of skill A was worse for ${\text{A}}_{\text{SHORT}}+\text{B}$ compared with ${\text{A}}_{\text{SHORT}}\left({\text{t}}_{18}=-2.63,\text{p}=.017,\text{d}=-1.18\right.)$ indicating retrograde interference. In contrast, training at asymptote reduced retrograde interference as retention of skill A was greater for ${\text{A}}_{\text{LONG}}+\text{B}$ compared with ${\text{A}}_{\text{SHORT}}+\text{B}({\text{t}}_{18}=2.82,\text{p}=.034,\text{d}=1.26)$. Importantly, retention of skill A was also greater for ${\text{A}}_{\text{LONG}}+\text{B}$ compared with ${\text{A}}_{\text{SHORT}}+{\text{B}}_{\text{DELAY}}({\text{t}}_{18}=2.72,\text{p}=.028,\text{d}=1.22)$, indicating that it was not simply the additional passage of time that reduced retrograde interference.

There was no difference between groups at baseline in error rate (${\text{F}}_{\mathrm{4,45}}=.35,\text{p}=.841,,{{\eta }^2}_{\text{p}}=.03$; Fig. 3A) or trial duration (${\text{F}}_{\mathrm{4,45}}=.80,\text{p}=.532,{{\eta }^2}_{\text{p}}=.07$; Fig. 3B). There was also no difference between groups in online changes in error rate (${\text{F}}_{\mathrm{4,45}}=.76,\text{p}=.559,{{\eta }^2}_{\text{p}}=.06$; Fig. 3C) or trial duration (${\text{F}}_{\mathrm{4,45}}=.90,\text{p}=.475,{{\eta }^2}_{\text{p}}=.07$; Fig. 3D). However, we found a difference across groups for retention of error rate (${\text{F}}_{\mathrm{4,45}}=5.31,\text{p}=.001,,{{\eta }^2}_{\text{p}}=.32$; Fig. 3E) but not trial duration (${\text{F}}_{\mathrm{4,45}}=2.06,\text{p}=.102,,{{\eta }^2}_{\text{p}}=.16$; Fig. 3F). Retention of error rate was better for ${\text{A}}_{\text{LONG}}+\text{B}$ compared with ${\text{A}}_{\text{SHORT}}+\text{B}\left({\text{t}}_{18}=-3.77,\text{p}=.004\right.,\text{d}=-1.69)$ and ${\text{A}}_{\text{SHORT}}+{\text{B}}_{\text{DELAY}}\left({\text{t}}_{18}=-3.62,\text{p}=.004,\text{d}=-1.62\right)$, indicating that greater accuracy contributed to higher skill in the prolonged training group.

Fig. 3 –

Error rate and trial duration for skill A and skill B. Vertical dashed line denotes the separation between the end of training on day one and the retention assessments on day two. (A) The y-axis represents the error rate and the x-axis shows training and retention epochs. (B) The y-axis represents the trial duration and the x-axis shows training and retention epochs. (C) The bar graph shows group average online changes in error rate for skill A and skill B, with individual data represented by the open circles. Online learning is the error rate difference between the last and first block of training on day one. (D) The bar graph shows group average online changes in trial duration for skill A and skill B. (E) The bar graph shows group average retention of error rate for skill A and skill B. Retention is the error rate difference between the block on day two and the last block on day one. For both skill A and skill B, retention of error rate was greater for ${\text{A}}_{\text{LONG}}+\text{B}$ compared with ${\text{A}}_{\text{SHORT}}+\text{B}$ and ${\text{A}}_{\text{SHORT}}+{\text{B}}_{\text{DELAY}}$. (F) The bar graph shows group average retention of trial duration for skill A and skill B. Data are means ± SEM. *p < .05.

For skill B, there was no difference in baseline performance $({\text{F}}_{\mathrm{2,27}}=.11,\text{p}=.896,{{\eta }^2}_{\text{p}}=.01)$ or online learning ${(\text{F}}_{\mathrm{2,27}}=.36,\text{p}=.703,{{\eta }^2}_{\text{p}}=.03)$ between groups. Similar to skill A, we found a difference in retention of skill B between groups ${(\text{F}}_{\mathrm{2,27}}=4.33,\text{p}=.023,{{\eta }^2}_{\text{p}}=.24)$. Specifically, training at asymptote also reduced anterograde interference denoted by greater retention of skill B in the ${\text{A}}_{\text{LONG}}+\text{B}$ group compared with ${\text{A}}_{\text{SHORT}}+\text{B}\left({\text{t}}_{18}=2.47,\text{p}=.024,\text{d}=1.11\right)$ and ${\text{A}}_{\text{SHORT}}+{\text{B}}_{\text{DELAY}}\left({\text{t}}_{18}=2.63\right.,\text{p}=.034,\text{d}=1.18)$.

There was no difference between groups at baseline in error rate (${\text{F}}_{\mathrm{2,27}}==.13,\text{p}=.881,{{\eta }^2}_{\text{p}}=.01$; Fig. 3A) or trial duration (${\text{F}}_{\mathrm{2,27}}=.60,\text{p}=.555,{{\eta }^2}_{\text{p}}=.04$; Fig. 3B). There was also no difference between groups in online changes in error rate (${\text{F}}_{\mathrm{2,27}}=.67,\text{p}=.519,{{\eta }^2}_{\text{p}}=.05$; Fig. 3C) or trial duration (${\text{F}}_{\mathrm{2,27}}=.40,\text{p}=.678,{{\eta }^2}_{\text{p}}=.03$; Fig. 3D). Similar to skill A, we found a difference across groups for retention of error rate (${\text{F}}_{\mathrm{2,27}}=5.31,\text{p}=.011,{{\eta }^2}_{\text{p}}=.28$; Fig. 3E) but not trial duration (${\text{F}}_{\mathrm{2,27}}=1.18,\text{p}=.321,{{\eta }^2}_{\text{p}}=.08$; Fig. 3F). Retention of error rate was better for ${\text{A}}_{\text{LONG}}+\text{B}$ compared with ${\text{A}}_{\text{SHORT}}+\text{B}\left({\text{t}}_{18}=-3.14\right.,\text{p}=.011,\text{d}=-1.40)$ and ${\text{A}}_{\text{SHORT}}+{\text{B}}_{\text{DELAY}}\left({\text{t}}_{18}=-2.67,\text{p}=.016\right.,\text{d}=-1.20)$, again indicating that greater accuracy contributed to higher skill in the prolonged training group.

3.2. Intracortical excitation is reduced after training at asymptote and is proportional to skill retention

For ICF, our marker of excitatory neurotransmission in M1, there was a main effect of GROUP $({\text{F}}_{\mathrm{1,38}}=5.51,\text{p}=.024,{{\eta }^2}_{\text{p}}=.13)$, no main effect of TIME $\left({\text{F}}_{\mathrm{1,38}}=.78,\text{p}=.383\right.,{{\eta }^2}_{\text{p}}=.02)$, but a GROUP × TIME interaction $({\text{F}}_{\mathrm{1,38}}=20.21,\left.\text{p}<.001,{{\eta }^2}_{\text{p}}=.35\right)$. We found that ICF increased after training in ${\text{A}}_{\text{SHORT}}({\text{t}}_{19}=2.19,\text{p}=.041,\text{d}=.53$; Fig. 4A). In contrast, we found that ICF decreased after training in ${\text{A}}_{\text{LONG}}\left({\text{t}}_{19}=-4.76\right.,\text{p}<.001,\text{d}=-.98)$ and was significantly lower than ${\text{A}}_{\text{SHORT}}{(\text{t}}_{38}=-3.80,\text{p}=.001,\text{d}=-1.20)$. Our correlation analyses indicated that modulation of ICF was proportional the retention of skill A. Specifically, a greater decrease in ICF after training at asymptote was associated with better retention of skill A for ${\text{A}}_{\text{LONG}}(\text{r}=-.71,\text{p}=.023$; Fig. 5) and ${\text{A}}_{\text{LONG}}+\text{B}(\text{r}=-.75,\text{p}=.013)$.

Fig. 4 –

Neurophysiology data. For all panels, the y-axis represents the ratio of the conditioned and nonconditioned motor evoked potential amplitude. The bar graph shows group average, with individual data represented by the open circles. (A) Intracortical facilitation (ICF) for ${\text{A}}_{\text{SHORT}}$ and ${\text{A}}_{\text{LONG.}}$. ICF increased in ${\text{A}}_{\text{SHORT}}$ and decreased in ${\text{A}}_{\text{LONG}}$ after training. (B) Short-interval intracortical inhibition (SICI) at 1 ms for ${\text{A}}_{\text{SHORT}}$ and ${\text{A}}_{\text{LONG}}$. (C) SICI at 3 ms for ${\text{A}}_{\text{SHORT}}$ and ${\text{A}}_{\text{LONG}}$. Data are means ± SEM. *p < .05.

Fig. 5 –

Correlation between modulation of intracortical facilitation (ICF) and retention of skill A. The y-axis represents retention of skill A which is the skill measure difference between the block on day two and the last block on day one skill and the x-axis represents the change in maximum ICF from pre-to post-training. Subjects in the ${\text{A}}_{\text{LONG}}$ and ${\text{A}}_{\text{LONG}}+\text{B}$ groups with larger decreases in ICF after training on skill A had better retention of skill A the following day.

3.3. Intracortical inhibition is reduced after training irrespective of duration

For SICI, our marker of inhibitory neurotransmission in M1, there was a main effect of TIME for ${\text{SICI}}_1\left({\text{F}}_{\mathrm{1,38}}=13.42,\text{p}=.001\right.,\left.{{\eta }^2}_{\text{p}}=.26\right)$ and ${\text{SICI}}_3\left({\text{F}}_{\mathrm{1,38}}=8.94,\text{p}=.005,{{\eta }^2}_{\text{p}}=.19\right)$, indicating a decrease in SICI after training (Fig. 4B and C). There was no main effect of GROUP $({\text{F}}_{\mathrm{1,38}}=.31,\text{p}=.583,{{\eta }^2}_{\text{p}}=.01;{\text{F}}_{\mathrm{1,38}}=.76,\text{p}=.389,{{\eta }^2}_{\text{p}}=.02)$ and no GROUP × TIME interaction ${(\text{F}}_{\mathrm{1,38}}=1.15,\text{p}=.291,{{\eta }^2}_{\text{p}}=.03;{\text{F}}_{\mathrm{1,38}}=2.18,\text{p}=.148,{{\eta }^2}_{\text{p}}=.05)$.

3.4. Motor execution and temporal delay controls

For the random group’s skill measure, there was no main effect of TIME $({\text{F}}_{\mathrm{2,18}}=.19,\text{p}=.827,{{\eta }^2}_{\text{p}}=.02)$, indicating skill was similar to baseline $(-6.07\pm .17)$ after five $(-5.94\pm .24)$ and ten $(-5.89\pm .35)$ blocks of practice. For ICF, there was no main effect of TIME $({\text{F}}_{\mathrm{2,18}}=.06,\text{p}=.941,{{\eta }^2}_{\text{p}}=.01)$, indicating ICF was similar to baseline $(1.45\pm .17)$ after five $(1.45\pm .15)$ and ten $(1.46\pm .19)$ blocks of practice. For ${\text{SICI}}_1$, there was a main effect of TIME $\left({\text{F}}_{\mathrm{2,18}}=7.36,\text{p}=.005,{{\eta }^2}_{\text{p}}=.45\right)$. We found that compared to baseline $(.13\pm .03)$, there was a trend for a decrease in ${\text{SICI}}_1$ after five blocks of practice $\left(.20\pm .06;{\text{t}}_9=2.14\right.,\text{p}=.061,\text{d}=.61)$, which was significant after ten blocks of practice $\left(.26\pm .07;{\text{t}}_9=3.52,\text{p}=.013,\text{d}=1.05\right)$. For ${\text{SICI}}_3$, there was a main effect of TIME ${(\text{F}}_{\mathrm{2,18}}=4.56,\text{p}=.025,{{\eta }^2}_{\text{p}}=.34)$. Similarly to ${\text{SICI}}_1$, we found that compared to baseline $(.30\pm .06)$ there was a trend for a decrease in ${\text{SICI}}_3$ after five blocks of practice $(.44\pm .09;{\text{t}}_9=2.01,\text{p}=.075,\text{d}=.44)$, which was significant after ten blocks of practice $\left(.42\pm .08;{\text{t}}_9=3.87\right.,\text{p}=.008,\text{d}=.53)$. Together, these results indicate that the bidirectional modulation of ICF in the main experiment was due to learning and not simply due to practice. In contrast, the observed decrease in ${\text{SICI}}_1$ and ${\text{SICI}}_3$ after training is consistent with modulation related to motor execution.

Similar to our finding in the ${\text{A}}_{\text{SHORT}}$ group, there was an increase in ICF from pre $(1.44\pm .12)$ to post $\left(1.85\pm .15,{\text{t}}_7=4.95\right.,\text{p}=.002,\text{d}=1.00)$ training in the ${\text{A}}_{\text{SHORT}}+{\text{B}}_{\text{DELAY}}$ group, indicating that it was not simply the additional passage of time that caused the decrease in ICF in ${\text{A}}_{\text{LONG}}$. There was no change in ${\text{SICI}}_1(\text{pre}=.08\pm .03,\text{post}=.15\pm .05,{\text{t}}_7=2.01,\text{p}=.085,\text{d}=.59)$ or ${\text{SICI}}_3(\text{pre}=.24\pm .06,\text{post}=.33\pm .09,{\text{t}}_7=1.61,\text{p}=.152,\text{d}=.23)$.

4. Discussion

Here we show that prolonged training at asymptote reduces the susceptibility of acquired motor skills to bidirectional interference. Excitatory neurotransmission in M1 decreased after training at asymptote, and this reduction was proportional to better skill retention the following day. Together, these findings indicate that training at asymptote improves the capacity to learn multiple motor skills in close succession, eliminating the time-dependency of memory stabilization, and that downregulation of excitatory neurotransmission in M1 may be a marker of this motor learning process.

We observed that training a motor skill at asymptote reduced retrograde interference from a subsequently acquired motor skill. This finding is consistent with previous motor learning studies showing that prolonged training reduces retrograde interference during adaptation to visuomotor rotation (Krakauer et al., 2005) and sequential reaching (Ghilardi et al., 2009). Here we extend this finding to a motor skill which requires sequential force application using a precision grip and engages error-based, reinforcement and use-dependent learning mechanisms (Mawase et al., 2017; Spampinato & Celnik, 2018). Furthermore, by dissecting our skill measure we found that the benefits of prolonged training were specific to accuracy. Interestingly, resistance to retrograde interference with prolonged training does not appear to be specific to motor learning, as similar observations have been described in visual perceptual learning (Shibata et al., 2017). Thus, it is becoming increasingly evident that training at asymptote reduces the susceptibility of motor memories to retrograde interference, an effect that seems to generalize across motor tasks and learning domains.

We also found that prolonged training at asymptote of an initial motor skill reduced anterograde interference of a subsequently learned skill. It is possible that prolonged training at asymptote stabilized the initially acquired skill online, allowing the subsequent skill to be acquired without disruption from a preceding unstable memory in competition for overlapping neural resources (Bang et al., 2019; Cantarero, Tang, et al., 2013). Similar to our retrograde findings, we found that the benefits of prolonged training were specific to accuracy. A recent study showed that a long inter-acquisition interval (~6 h) is usually required to prevent anterograde intereference of a second motor memory by a preceding memory (Lerner et al., 2020). Therefore, our findings indicate that training at asymptote is also a temporally efficient method for reducing anterograde interference of a subsequently acquired motor skill.

Motor skill stabilization occurred following training at performance asymptote. This is evident from our data, as no additional online gains were made despite the prolonged training. It has been proposed that consolidation mechanisms underlying memory stabilization may be initiated online if practice is continued after performance has plateaued (Hauptmann et al., 2005; Krakauer et al., 2005; Yin & Kitazawa, 2001). This online stabilization mechanism is likely distinct to offline consolidation processes occurring minutes to hours post training (Muellbacher et al., 2002; Tunovic et al., 2014), as well as rapid offline consolidation contributing to online gains during early sequence learning (Bonstrup et al., 2019). In the present study, it is possible that consolidation processes were initiated online in the ${\text{A}}_{\text{LONG}}+\text{B}$ group who completed additional practice at performance asymptote, but not the ${\text{A}}_{\text{SHORT}}+\text{B}$ group. Furthermore, interference was still observed in our temporal control group $({\text{A}}_{\text{SHORT}}+{\text{B}}_{\text{DELAY}})$, indicating that in the ${\text{A}}_{\text{LONG}}+\text{B}$ group it was additional practice at plateau and not simply the passage of time that was required for stabilization to occur. It is important to note that our conclusion that the prolonged training group reached performance asymptote is based on comparisons of group average online learning. We acknowledge that some individuals could have exhibited additional online gains between blocks 6–10. Whether being at performance asymptote is required for online stabilization and if tailoring training duration based on when individuals reach performance asymptote leads to more robust effects should be explored in future studies.

Neurophysiological assessments indicated that ICF, an index of excitatory neurotransmission within M1 (Ziemann et al., 2015), was modulated in a training duration-dependent manner. Specifically, ICF increased after a short bout of training and decreased after training at asymptote. Importantly, these effects were specific to learning, as ICF was not modulated after short and long bouts of motor practice without learning. Increased ICF after short training may reflect heightened excitatory neurotransmission within M1 attributed to learning induced LTP (Rioult-Pedotti et al., 1998, 2000; Rosenkranz et al., 2007; Ziemann et al., 2004). However, this excitatory state may create an unstable environment, in which newly acquired motor memories are susceptible to interference. Interestingly, the observed increase in ICF after short training persisted for up to 30 min, during which the acquired motor skill remained unstable. In contrast, ICF decreased after prolonged training at asymptote. After an initial increase in ICF during motor memory acquisition, repetition at performance asymptote may drive a reduction in excitatory activity within M1 to stabilize and protect the encoded memory. Correlation between modulation of ICF and skill retention in the prolonged training groups, where greater decreases in ICF were associated with better retention, supports this idea. Our findings are also consistent with reduced weighting of excitatory processing in primary visual cortex after prolonged visual perceptual training (Shibata et al., 2017). Thus, reduced excitatory neurotransmission in M1 after prolonged training at asymptote may reflect the process of online motor memory stabilization.

We observed decreased SICI, an index of inhibitory neurotransmission within M1 (Ziemann et al., 2015), after training irrespective of training duration. Reduced SICI was observed at both 1 and 3 ms which are proposed to be mechanistically distinct (Stagg et al., 2011), and is consistent with previous studies reporting a reduction in inhibitory tone within M1 during motor learning (Berghuis et al., 2016; Coxon et al., 2014; Floyer-Lea et al., 2006; Kolasinski et al., 2019; Mooney et al., 2019). It has been proposed that downregulation of inhibitory processing may accompany and promote the strengthening of excitatory synapses with learning induced LTP (Feldman, 2009). If this was solely the case, we might have expected to see similar training-duration dependent modulation of SICI as was observed with ICF, whereby decreased SICI would accompany increased ICF after short training and vice-versa after prolonged training. Also, we observed decreased SICI after training on an unlearnable motor task, a finding consistent with previous studies (Jayaram et al., 2011; Spampinato & Celnik, 2017), indicating that SICI is modulated by simple motor practice. Of note, neuroimaging studies have demonstrated learning specific modulation of inhibition (Floyer-Lea et al., 2006; Kolasinski et al., 2019), which may be due to signal differences between assessment modalities (neurotransmitter concentration vs. synaptic efficacy). Nevertheless, our findings of decreased SICI irrespective of training duration and after motor practice without learning is consistent with modulation due to motor execution.

The present study has some limitations. Firstly, the sample sizes for the behavioral component are small (n = 10 per group). However, sample sizes were conservatively determined based on effect sizes from previous motor learning studies demonstrating reduced retrograde interference with prolonged training (Ghilardi et al., 2009; Krakauer et al., 2005). Furthermore, our experimental design allowed us to assess neurophysiological responses in moderately sized groups (n = 20 per group), with all observed changes in ICF and SICI exceeding what would be expected due to measurement noise (Samusyte et al., 2018; Schambra et al., 2015). Secondly, in addition to anterograde interference effects on day one, it is possible that prior recall of skill A may have disrupted recall of skill B on day two. Because we did not collect data from a group who trained solely on skill B, we are unable to determine whether this was the case. However, if present, a detrimental effect of recalling skill A prior to recalling skill B would be expected in all three groups (${\text{A}}_{\text{SHORT}}+\text{B},{\text{A}}_{\text{LONG}}+\text{B}$ and ${\text{A}}_{\text{SHORT}}+{\text{B}}_{\text{DELAY}}$) and is therefore unlikely to confound our finding of better retention of skill B in the prolonged training group. Thirdly, we only collected 12 trials per condition for the TMS assessments, with a minimum of 10 trials for some participants. However, we opted for 12 trials to capture ICF and SICI curves for each participant in a temporally efficient manner. Measuring the curves allowed us to determine the conditioning intensity which produced maximum ICF and SICI for each individual, which has lower measurement error compared with fixed group conditioning intensities (e.g., 80% of motor threshold) (Samusyte et al., 2018).

In summary, we demonstrate that training at asymptote reduces retrograde and anterograde interference during motor skill learning. This phenomenon is associated with an online stabilization process depicted by a reduction of excitatory neurotransmission in M1. Together, these findings have important implications for designing strategies to optimize motor skill learning in healthy individuals and in the context of clinical rehabilitation.

Data and material availability

No part of the study procedures or analyses was pre-registered prior to the research being conducted. We report how we determined our sample size, all data exclusions, all inclusion/exclusion criteria, whether inclusion/exclusion criteria were established prior to data analysis, all manipulations, and all measures for this study. Task and analysis code are available to access at this link: https://github.com/rmooney7/Cortex_Interference_SVIPT. The conditions of our ethics approval do not permit public archiving of any anonymized data. Readers seeking access to the data should contact the corresponding author Pablo Celnik (pcelnik@jhmi.edu). Access will be granted to named individuals in accordance with ethical procedures governing the reuse of data. Specifically, requestors must complete a formal data sharing agreement.