henhaobio
Sign in

bioRxivNeurosciencePreprint

Distributed neural dynamics underlie the shift from movement preparation to execution

Ziwei YiniD, Jian K. LiuiD, Katja KornyshevaiD

PDFDOI
In-depth full-text digest available · Jump to digest ↓

Abstract

Understanding how neural population activity supports the transition from movement preparation to execution remains a central challenge in systems neuroscience. Invasive recordings in non-human primates indicate that motor preparation and execution occupy orthogonal subspaces within the primary motor cortex (M1), yet how these neural dynamics evolve across the broader networks and support skilled memory-guided action remains poorly understood. We used source-resolved magnetoencephalography (MEG) during a memory-guided finger-sequence task to track population dynamics across a cortico-hippocampal network. Manifold trajectories and decoding revealed that motor preparation and execution of movement sequences occupy distinct, yet not fully orthogonal, neural subspaces - a separation that generalised across all regions but was most pronounced in motor regions. Crucially, we identified a hierarchical spatiotemporal transition across the brain: the state-shift originated in the hippocampus via theta-band modulations and rippled through the pre-SMA, SMA, and lateral premotor areas before culminating in M1 via broadband activity. This progression follows the functional hierarchy of descending motor pathways, with M1 shifting last, 160 ms before the first button press. Although sequence identity could be decoded from all areas during execution, only the hippocampus and M1 maintained predictive representations of upcoming sequences during preparation. These findings demonstrate that the initiation of complex motor sequences relies on a well-timed integration of hippocampal memory systems with primary motor output, rather than being a process localised to the motor cortices alone.

In-depth summary

Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 4 (19 Oct 2026) · CC BY 4.0

Magnetoencephalography during memory-guided finger sequences shows preparation and execution occupy distinct, partly overlapping subspaces; execution patterns appear first in hippocampus and last in M1, 160 ms before the first press.

Primate recordings place motor preparation and execution in orthogonal subspaces of primary motor cortex (M1), but invasive work has focused on M1 and premotor cortex, and human fMRI has not resolved the timing or geometry of the transition. The work applies geometric and decoding analyses to source-reconstructed MEG across hippocampal, premotor and M1 regions, asking whether the subspaces stay orthogonal for memorised sequences and in what order regions switch.

Key results

  1. SuggestedExecution-specific patterns emerge in a graded order before the first press, earliest in hippocampus and pre-supplementary motor area (pre-SMA) and last in M1, whereas preparatory patterns end after the Go Cue with no regional ordering.

    Onset of execution patterns relative to first press: hippocampus −0.76 ± 0.65 s, pre-SMA −0.68 ± 0.69 s, SMA −0.30 ± 0.27 s, PMd −0.18 ± 0.20 s, PMv −0.16 ± 0.16 s, M1 −0.16 ± 0.14 s; mixed-model slope over hierarchy rank β1 = +0.133, p < .001. Offset of preparation patterns showed no trend (β1 = −0.024, p = .406). · Timing of the transition from preparation to…; Hierarchical transition to execution patterns

  2. ShownPreparatory and execution activity separate in every region tested, including the primary visual cortex (V1) control; covariance-based separation grows along the hippocampus-to-M1 ordering, while distance in PCA space does not differ by region.

    Euclidean phase distance in 3-PC space exceeded the permuted null in all regions (p ≤ .0015) but did not vary by region (Friedman χ²(5) = 2.12, p = .832). Covariance pattern distance exceeded shuffled labels (p < .001) and rose along the hierarchy (β1 = 0.000138, p = .009); t-SNE phase distance was 28.25 in M1 vs 23.14 in hippocampus (β1 = 0.840, p < .001). · Principal component analysis reveals distinct…; MEG covariance pattern separation is driven by…; Primary visual cortex as a non-motor control…

  3. ShownThe two phase subspaces are neither orthogonal nor fully aligned in any region: their principal axes sit at intermediate angles, and subspace overlap is high but below perfect.

    Angles between Gaussian-mixture principal axes ranged from 38.4 ± 22.8° (PMd) to 60.6 ± 24.0° (pre-SMA), differing from both 0° and 90° (p < .001). Alignment index ranged from 0.993 to 0.999 across the six regions, each significantly below 1 (p < .001). V1 had lower alignment than M1 (p = .018) but was also not orthogonal. · Principal component analysis reveals distinct…; Primary visual cortex as a non-motor control…

  4. SuggestedSequence identity is decodable in all motor regions, mainly during execution. During preparation it appears only in M1 and hippocampus, plus the V1 control, where the authors attribute it to the on-screen visual cue.

    Within-phase sliding-window LDA (200 ms windows) beat shuffled labels during preparation only in M1 (−1.64 to −1.2 s, p = .025) and hippocampus (−1.58 to −1.26 s, p = .044); V1 exceeded all regions in preparation (p < .001). Mean accuracy rose along the hierarchy (execution β1 = 0.0036, p < .001; preparation β1 = 0.0009, p = .034). Finger order decoded better than timing. · Decoding sequence identity; Primary visual cortex as a non-motor control…

  5. SuggestedThe switch to execution patterns tracks broadband power in M1 and premotor areas but only low-frequency power in the hippocampus; sequence decoding is not reliably tied to any single band.

    Ridge regression of the execution-decoding time course on band power gave significant clusters near the transition for delta, theta, alpha, beta and gamma in M1 (similar in PMd, PMv, SMA, pre-SMA), but only delta and theta in hippocampus. For sequence-decoder output, coefficients were unstable and mostly not significant against a shuffled baseline. · Timing of the transition from preparation to…; Decoding sequence identity; Hierarchical transition to execution patterns

Methods, data and reproducibility

Reanalysis of 275-channel MEG during memorised five-press finger sequences. Beamformed signals from M1, dorsal and ventral premotor cortex (PMd, PMv), SMA, pre-SMA, hippocampus and a primary visual cortex (V1) control were analysed with linear discriminant analysis (LDA), PCA, demixed PCA, covariance t-SNE, band-power regression and hierarchy-ranked mixed models.

  • Reanalysed MEG (275-channel CTF axial gradiometer, 1,200 Hz) with concurrent EMG from a memory-guided finger-sequence task (Kornysheva et al. 2019); 240 memory-guided trials per participant · human (healthy right-handed participants) · n = 14 participants (16 in original study)
Methods & tools
FieldTrip (v20260214), CTF Omega MEG system, LCMV beamformer, Human Motor Template (HMAT) atlas, NeuroVault sub-cortical masks, Jülich Brain Atlas (hOc1/Area 17), statsmodels MixedLM, linear discriminant analysis, demixed PCA (dPCA), t-SNE, Gaussian mixture model, ridge regression
Reagents & environment
FieldTrip v20260214; linear mixed-effects models fitted in Python with statsmodels MixedLM (REML). No other software versions or environment files stated.
Reproducibility gaps
No analysis code link is given; study data and materials are available only on request by email to the lead author.
Limitations (6)
  • MEG reflects synchronised postsynaptic currents rather than single units, weakens with depth and favours superficial cortex, so spatial blurring and source leakage between nearby regions are possible; strict orthogonality may be hard to resolve in mixed source signals. · stated by the authors
  • The sample of N = 14 limits power to detect subtle effects and higher-order factor interactions. · stated by the authors
  • In this task, planning is entangled with memory retrieval and cue-sequence association; non-memory control tasks would be needed to separate general motor preparation from sequence retrieval. · stated by the authors
  • Transition timing varied substantially between participants, especially in premotor areas and hippocampus, with some participants shifting to execution patterns before the Go Cue. · stated by the authors
  • Sensor data were co-registered to a standard template MRI rather than individual anatomy. This adds spatial uncertainty, notably for the deep hippocampal source that leads the timing sequence. · noted in this summary
  • The hierarchy ranks are an ordinal assumption set by the authors. Pairwise timing differences were significant only for M1 vs pre-SMA and for hippocampus vs M1, PMd, PMv and SMA (PMv and M1 share a −0.16 s mean onset), so M1 switching last rests on the fitted slope. · noted in this summary

Read this if you study motor sequence planning or neural population geometry and want to see how far subspace, manifold and decoding analyses developed for invasive recordings carry over to source-reconstructed human MEG.

Citations
1

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISYewbrey '24Yoon '23Yewbrey '23Ariani '22Tang '21Inagaki '22
78% semantic matchNeuroscience
Rhys Yewbrey, Katja Kornysheva·Journal of Neuroscience·Sep 24, 2024· 18 citations