Refinement of learned skilled movement representation in motor cortex deep output layer

Abstract

The mechanisms underlying the emergence of learned motor skill representation in primary motor cortex (M1) are not well understood. Specifically, how motor representation in the deep output layer 5b (L5b) is shaped by motor learning remains virtually unknown. In rats undergoing motor skill training, we detect a subpopulation of task-recruited L5b neurons that not only become more movement-encoding, but their activities are also more structured and temporally aligned to motor execution with a timescale of refinement in tens-of-milliseconds. Field potentials evoked at L5b in vivo exhibit persistent long-term potentiation (LTP) that parallels motor performance. Intracortical dopamine denervation impairs motor learning, and disrupts the LTP profile as well as the emergent neurodynamical properties of task-recruited L5b neurons. Thus, dopamine-dependent recruitment of L5b neuronal ensembles via synaptic reorganization may allow the motor cortex to generate more temporally structured, movement-encoding output signal from M1 to downstream circuitry that drives increased uniformity and precision of movement during motor learning.

Figure 1. Forelimb reaching for food training and simultaneous recording from L5b neurons in M1.

(a) Schematics of experimental paradigm. Neural activities in L5b were recorded during forelimb food-reaching task training over 7 days by multi-channel recording electrode array (Rec.). FL and HL: forelimb and hindlimb territories of M1; S1: primary somatosensory cortex; V1: primary visual cortex. (b) Six phases of a first reach success trial captured by camera and the forelimb trajectory tracked automatically (see Methods). (c) The evolution of more uniform forelimb trajectories (pink) in first reach success trials. The reference expert trajectory is shown in blue colour. Red asterisk denotes the position of food pellet. (d) The duration of forelimb extension (upper panel), grasping (middle panel) and retraction (lower panel) in first reach success trials. The timing of reaching action shortened significantly in day 1 and exhibited further decrease in day 2 and 3, and remained steady thereafter. Mean±s.d. *P<0.05; **P<0.01; ***P<0.001, one-way repeated measures ANOVA, n=9. (e) Left, delay in first reach attempt quantified by the time interval between food provision and the ‘orient' position of forelimb on days 1 and 7 (120 consecutive trials each from a single representative rat). Right, learning associated shortening in the delay in first reach success attempt. Mean±s.d. of delay in first reach success trials: day 1: 3.44±0.37 s; day 7: 0.96±0.14 s, P=2.51 × 10⁻⁶; *P<0.05; **P<0.01; ***P<0.001, all compared with day 1; one-way ANOVA, 9 rats; first reach failure trials: day 1: 2.31±1.55 s; day 7: 1.45±1.20 s, P=0.012; one-way ANOVA, 9 rats. (f) Training-dependent improvement in first reach success rate (see text). (g) Evaluation of forelimb trajectory spatial variance as the averaged distance integrated over time between the actual trajectories in first reach success trials (pink) and the reference expert trajectory (blue) shown in d. Mean cumulative Euclidean distance±s.d., day 1 session 1: 0.375±0.061 cm; day 1 session 6: 0.265±0.042 cm, P=0.0173; day 2 session 1: 0.252±0.041 cm, P=0.0052; day 2 session 6: 0.205±0.026 cm, P=7.85 × 10⁻⁴; *P<0.05; **P<0.01; ***P<0.001, all compared to day 1 session 1, one-way repeated measures ANOVA, n=9.

Figure 2. Spike sorting and assessment of long-term stability of single-unit recordings by single microwire array over 7 days.

(a) Example of spike sorting from single microwire array in 7 days, showing the superimposed spike waveforms (upper panel) and the inter-spike-interval histogram (ISIH, lower panel), and the corresponding identified clusters in the PCs space (far right panel). Clear isolation of units from a given recording channel is indicated by high, F statistic of MANOVA (F), J3, Dunn validity (Dn) and low Davis-Bouldin (DB) index (see Methods). Note the excluded unit in red, whose spike waveforms changed cross days, and had shifted ISI histogram and cluster location in PCs. (b) Long-term stability of identified single-units shown in a over 7 days. The unit shown in red with drifting of cluster was excluded. (c) Autocorrelograms of the three isolated units and their cross-correlogram (white). The presence of refractory periods in the auto-correlograms and absence of refractoriness in the cross-correlogram indicated spikes with clusters marked in yellow, green and blue were generated by three distinct neurons. The short latency sharp peak in the cross-correlogram (arrow) between the putative pyramidal neuron (yellow, reference of the cross-correlogram) and the interneuron (green) may indicate mono-synaptic activation. (d) Example of units exhibiting stable (top, unit 1 in a) or unstable (bottom, unit 4 in a) spike waveform are shown in d. (e) Gaussian mixture distributions fitted to combinations of the four similarity scores (see Methods) computed from spikes recorded from same neuron (black dots, representing true positive values computed using recording acquired in difference sessions on the same day, see Methods) or distinct ones (grey dots, computed from recordings from different channels simultaneously) on the same day corresponding contours: 50% (red), 95% (blue), 99.9% (orange), 99.97% (black) of the distribution. Red and blues crosses represented recordings classified as stably corresponding to the same or arising from distinct neurons respectively, based on combining the use of multiple similarity scores with quadratic classifiers (green lines). (f) Cross-day stability of single-unit isolation quality assessed by L-ratio and isolation distance (n=158 included units shown in red in e). (g) Cross-day stability of four cluster similarity scores (n=158 included units shown in red in e, day 1 session 1 recordings were used as reference).

Figure 3. Fine-temporal scale refinement of firing-motor output relationship in a subpopulation of L5b PNs.

(a) Three examples of L5b PNs' peri-event time histograms, aligned by the ‘orient' position (time ‘0', white dots). Fifty-six consecutive first reach success and 80 consecutive first reach failure trials are stacked. White arrows indicate the time when food pellets were provided. The correspondent forelimb velocity, are shown overlaid on the top. Neural activities were normalized and expressed as Z score. Neuron A's firing highly correlated with forelimb action but remained unchanged with training. Neuron B's firing became correlated with forelimb action after training. Neuron C's firing did not correlate with forelimb reaching action irrespective of training. (b) Averaged neural activities of the three L5b PNs shown in a during first reach success (red, mean±s.e.m.) and first reach failure (blue, mean±s.e.m.) trials on days 1 and 7. Arrows indicate the time when the neural activity began to diverge (TUD, time until divergence, see Methods). (c) Hierarchical clustering of 131 L5b PNs (recorded from five rats) based on single neuron I_M during motor learning. The dendrogram (upper half) depicts Euclidean distance of single PN I_M vectors across 7 training days, with major subgroups indicated by different colours in the dendrogram. (d) Twenty-seven L5b INs (recorded from five rats) were classified into subgroups by applying the same method of hierarchical clustering of single neuron I_M as shown in c. (e) Summary of training-dependent changes of the optimal time lag of I_M (τ_opt. for the three major types of L5b PNs (recorded from five rats) classified by hierarchical clustering. Frequency histograms of individual neuron τ_opt. at days 1 and 7 are shown on the top and the right respectively. Sixty out of sixty-one of type 2 neurons exhibit a decrease in τ_opt., which are distributed below the dashed line with unit slope.

Figure 4. Learning-dependent changes of population prediction accuracy for forelimb instantaneous velocity.

(a) SVR decoding of forelimb velocity from neural population activities. For each neuron, the firing histogram (bin size: 12.5 ms, left) was aligned to the behavioural event (27 units from an example animal), and the values was linearly normalized to 0–1 range. The actual forelimb instantaneous velocity (top right) was predicted using SVR by the corresponding population spike events. (b) Representative traces of actual forelimb instantaneous velocity (black) and the SVR model predicted forelimb velocity (red) by three types of neurons classified (type 1 neuron: n=4; type 2 neuron: n=12; type 3 neuron: n=11, from an example animal), illustrating the changes in population decoding accuracy during early (day 1) and late (day 7) training sessions. (c) Least squares regression analyses between actual forelimb instantaneous velocity and the SVR model predicted forelimb velocity based on three types of L5b PNs shown in a during early and late training sessions. The Pearson's correlation coefficient (r²) and mean squared deviation (MSD) for each regression are shown. Each data point represents the instantaneous velocity of the forelimb trajectory predicted from neural population activity versus the actual velocity of displacement calculated from high-speed camera recording (in 12.5 ms bins). (d) Summarized result of r² and MSD of predicted and actual forelimb instantaneous velocity by three types of PNs (n=131) recorded from five rats. Upper left panel, day 1: r²=0.513±0.013, day 3: r²=0.505±0.010, P=0.098; day 7: r²=0.520±0.013, P=0.492; bottom left panel, day 1: MSD=0.0069±0.00043, day 3: MSD=0.0073±0.00028, P=0.370; day 7: MSD=0.0072±0.00033, P=0.448. Top middle panel, day 1: r²=0.067±0.007, day 3: r²=0.211±0.014, P=0.006; day 7: r²=0.483±0.016, P=1.23 × 10⁻⁴; Bottom middle panel, day 1: MSD=0.0161±0.00048, day 3: SD=0.0130±0.00034, P=0.008; day 7: MSD=0.0081±0.00035, P=0.002. Top right panel, day 1: r²=0.061±0.006, day 3: r²=0.066±0.015, P=0.587; day 7: r²=0.073±0.008, P=0.312. Bottom right panel, day 1: MSD=0.0169±0.00047, day 3: MSD=0.0171±0.00051, P=0.282; day 7: MSD=0.0167±0.00055, P=0.781, all compared to day 1, one-way repeated measures ANOVA, n=5.

Figure 5. Emergence of correlation structure of L5b PNs during motor learning.

(a) Pairwise cross-correlation matrix of 27 L5b PNs across 7 training days, recorded from one representative rat during first reach success attempts (controlled for trajectory variance, see Methods). Neurons are ordered (PN #) according to the sequence of hierarchical clustering shown in Supplementary Fig. 5b. The squares from top to bottom segregate type 1 to type 3 L5b PNs identified. Increased correlation is evident only among the groups of neurons that show increased I_M and decrease of τ_opt., that is, type 2 neurons. (b) Summary of the preserved overall similarity of cross-day correlation matrix among type1 and type 2 PNs. Each colour-coded element represented the averaged similarity index of cross-day correlation matrices from five rats. (c) Paradigm of rotarod running. Rats were trained to run on the rotarod accelerating from 4 to 40 revolutions per minute over 300 s. Each trial ended when the rat fell off or when 300 s was reached. Each animal received six training sessions every day, and each lasted 10 min with 5-min rest intervals. (d) Latency to fall off the rotarod during training. Animals showed fast improvement in performance on the first two days and maintained throughout the third day (mean±s.e.m. of latency to fall in day 1 session 1: 62.0±19.1 s, day 1 session 6: 133.0±33.4 s, P=0.051; day 2 session 1: 132.1±21.7 s, day 2 session 6: 249.0±34.6 s, P=0.007; day 3 session 1: 255.2±25.7 s, day 3 session 6: 269.5±20.3 s, P=0.681, all compared to day 1 session 1, one-way repeated measures ANONA, 4 rats). (e) Pairwise cross-correlation matrix among the 27 L5b PNs shown in a, but re-ordered for clustering with high correlation coefficient near the diagonal during 3 days' rotarod training (day 8 to day 10).

Figure 6. Illustration of single-trial neural trajectories extracted from population neural activities.

(a) The temporals of ten single-trial neural trajectories extracted from first reach success trials by applying GPFA (see Methods) embedded in the top three orthonormalized latent dimensional space (LD1-3) spanning from 800 ms before to 800 ms after the ‘orient' position, derived from 27 L5b PNs recorded from an representative animal. The black arrows indicate the flow of time series and coloured dots denote different stages of forelimb action. The ellipses indicate the across-trial variability (two s.d. around the mean) at these states. (b) and (d) Top: Neural trajectories of 30 randomly selected first reach success trials with trajectory deviation within mean±s.d. of cumulative Euclidean distance from reference expert trial after DTW, and neural trajectories of 30 randomly selected first reach failure trials, derived from 27 L5b PNs' activities recorded from the same example animal. The flow of time series is colour gradient coded, from blue (start) to red (end). Bottom: the same neural trajectories shown in three individual latent dimensional space (LD1-3). Arrow indicates ‘orient' position. (c) and (e) Analysis of the variance of the neural trajectories in LD1-3, quantified by the diagonal matrix of the covariance ellipsoids, showing progressive reduction in the variance in first reach success trials (c) but not in failure trials (e) during training.

Figure 7. Local dopamine depletion impaired training-induced LTP of synaptic inputs.

(a) Top, averaged traces of field potentials (FPs) evoked in vivo at multiple sites from L1 to L6 of M1, recorded via 20 recording contacts in a linear microprobe (Rec.). The stimulating electrode (Stim.) was placed at L5, which could activate synaptic inputs to the basal dendrites. Arrows indicated typical FPs recorded at target deep layer 5. Bottom, current source density profiles corresponding to the laminar FPs evoked by stimulating at L5. By activating inputs targeting basal dendrites, the early, negative FP recorded in L5 was generated by direct inward current (that is, the sink, yellow/red). This feature was highly consistent among different animal subjects. (b) Top, FPs were recorded from rats undergoing 7 days of forelimb reaching task, followed by 3 days of rotarod running task. Bottom, potentiation of stimulation evoked-FPs slope when activating basal dendritic inputs in L5. Representative traces of the FPs on days 1 and 7 (pre- and post-training) are shown. All bars represent the mean±s.e.m (five rats). (c) Learning-associated shortening in delay of first reach attempt in sham-operated (black, five rats) and 6-OHDA lesioned animals (blue, five rats). Sham group: mean±s.d. of delay on day 1: 4.416±0.783 s, day 3: 3.362±0.588 s, P=0.038; day 7: 1.118±0.216 s, P=0.006, all compared to day 1, one-way repeated measures ANOVA, 5 rats; Lesioned group: mean±s.d. of delay on day 1: 3.663±0.676 s, day 4: 2.631±0.849 s, P=0.065; day 5: 2.376±0.778 s, P=0.048; day 7: 1.587±0.463 s, P=0.019, all compared to day 1, one-way repeated measures ANOVA, 5 rats. (d) Comparison of motor skill performance between sham-operated (black, five rats) and 6-OHDA lesioned (blue, five rats) animals. Data are represented as mean±s.e.m. In contrast to the sham group, the first reach success rate achieved after each day's training by the lesioned animals was not well maintained in the next day. (e) Comparing to sham-operated group (black, five rats), with local dopamine depletion restricted to M1 (blue, five rats), learning-induced potentiation of FPs could not be maintained. Data are represented as mean±s.e.m.

Figure 8. Neural dynamics of L5b PNs after dopamine depletion.

(a) and (b) 95 PNs (a) and 19 INs (b) in L5b recorded from four rats with local dopamine depletion in L5 of M1 forelimb territory, were classified into subgroups by hierarchical clustering of single neuron I_M during motor learning, following the same method as shown in Fig. 3c. (c) Summary of training-dependent changes of the optimal time lag of I_M (τ_opt.) in three types of L5b PNs after dopamine depletion. Statistical quantification indicated that throughout 7 days' training, there was less consistent change and only a slight reduction of averaged τ_opt. in type 2 PNs (type 1: P=0.681, n=11; type 2: P=0.042, n=45; type 3: P=0.852, n=39; paired t-test, 4 rats) compared with intact animals. (d) The pairwise cross-correlation matrix of 27 L5b PNs recorded from the same example rat shows that there was no emergence of consistent functional clusters after 7 days' motor training (cf. Fig. 5a). The averaged correlation values from the PNs of four lesioned rats from day 1 session 1(D1S1) to day 7 session 6 (D7S6) are shown on the right. (e) Top, single-trial neural trajectories of randomly selected first reach success trials (randomly selected 50 trials per day) performed by dopamine-depleted rat (27 L5b PNs recorded from a representative animal). The flow of time series is colour gradient coded, from blue (start) to red (end). Bottom, the same neural trajectories shown in three individual latent dimensional space (LD1-3). Arrow indicates ‘orient' position. Compared with intact animals, reproducible neuronal trajectories did not emerge during training, even in first reach success attempts (cf. Fig. 6b). (f) Analysis of the variance of the neural trajectories in LD1-3 after dopamine depletion (95 L5b PNs from four rats). Statistical quantification confirmed the lack of training-dependent reduction in variance even in first reach success trials at day 7 (cf. Fig. 6c).