Motor cortical dynamics are shaped by multiple distinct subspaces during naturalistic behavior

Matthew G. Perich, Sara Conti, Marion Badi, Andrew Bogaard, Beatrice Barra, Kanaka Rajan, Jocelyne Bloch, Grégoire Courtine, Silvestro Micera, Marco Capogrosso, Tomislav MilekovicbioRxiv 2020 · DOI

TL;DR

Motor cortex, traditionally thought to operate with deterministic dynamics for robust movement, can flexibly integrate unexpected sensory feedback by compartmentalizing its activity into distinct subspaces. One subspace handles deterministic motor output, while another, the "feedback subspace," specifically processes incoming sensory information, allowing for rapid error correction without disrupting ongoing motor commands.

What Problem Does This Paper Address?

The motor cortex is crucial for generating smooth, controlled movements. However, a long-standing puzzle is how it achieves both robust, predictable movement (suggesting deterministic dynamics) and the ability to rapidly adapt to unexpected sensory information (suggesting input-driven dynamics). Prior studies often found motor cortex activity to be surprisingly deterministic, even when subjects encountered perturbations. This paper aims to resolve this paradox by investigating how the motor cortex might simultaneously maintain deterministic control for planned actions while also being responsive to salient, unexpected sensory inputs, particularly during error correction.

Methods

The researchers employed a multi-pronged approach:

  • Neural Recordings: They simultaneously recorded neural population activity from the motor and somatosensory cortices of four monkeys performing a naturalistic object interaction task.
  • Trajectory Tangling Metric: This novel metric was used to quantify the "input-drivenness" of neural trajectories. High tangling indicates that similar neural states can lead to divergent future states, suggesting influence from unexpected inputs. Low tangling suggests deterministic dynamics where current states predict future states.
  • Decomposition into Optimal Subspaces (DOS) Algorithm: This algorithm was developed to identify distinct, orthogonal subspaces within the neural manifold. The goal was to isolate dimensions that specifically covary with behavioral outputs (behavioral subspace) and those that covary with somatosensory cortex activity (feedback subspace).
  • Recurrent Neural Network (RNN) Simulations: RNNs were used to model neural dynamics and to establish a baseline for "deterministic dynamics" tangling values, providing a reference for interpreting experimental data.
  • Causal Stimulation Experiments: Electrical stimulation of ascending somatosensory tracts was used to directly activate peripheral sensory pathways. This allowed researchers to causally link specific sensory inputs to activity within the identified subspaces.

Datasets

The study utilized neural recordings from four monkeys performing a naturalistic object interaction task. This task involved reaching, grasping, and pulling an object under controlled conditions, with occasional behavioral errors introduced. No new benchmarks were introduced; rather, the focus was on analyzing existing behavioral and neural data within a novel framework.

Key Insights and Ideas

  • Subspace Compartmentalization: The motor cortex doesn't operate as a monolithic unit. Instead, it segregates different types of dynamics into distinct, lower-dimensional subspaces.
  • Feedback Subspace for Errors: Input-driven dynamics, particularly those associated with error correction, are not spread throughout the motor cortex but are specifically localized to a "feedback subspace."
  • Behavioral Subspace for Determinism: The "behavioral subspace" appears to maintain deterministic dynamics, likely responsible for generating the core motor commands for planned movements.
  • Causal Link to Sensory Input: Electrical stimulation of somatosensory pathways directly and preferentially influenced the feedback subspace, providing strong evidence that this subspace indeed captures peripheral sensory information.
  • Resolving the Paradox: This subspace organization elegantly resolves the paradox of deterministic yet input-driven dynamics. The motor cortex can maintain robust, predictable output via the behavioral subspace while simultaneously processing salient sensory feedback through the feedback subspace, enabling flexible adaptation.

Key Results

  • During behavioral error correction, motor cortical tangling (indicating input-driven dynamics) significantly increased, but this increase was confined to the feedback subspace.
  • The behavioral subspace showed no significant change in tangling during error correction, remaining within the range expected for deterministic dynamics.
  • Electrical stimulation of somatosensory tracts led to a significant increase in the variance captured by the feedback subspace, but not the behavioral subspace, establishing a causal link.
  • The feedback subspace activity was found to lead the behavioral subspace activity by approximately 30ms following stimulation, further supporting its role in processing incoming sensory information.

Limitations and Future Work

  • Linear Subspaces: The study used linear subspaces to decompose the neural manifold. The authors acknowledge that these linear subspaces are likely approximations of a more complex, nonlinear manifold. Future work could explore nonlinear decomposition methods.
  • Inter-Subspace Communication: While the study demonstrates compartmentalization, the precise mechanisms by which these subspaces interact and communicate to generate the overall population dynamics remain an open question. Future research could investigate direct links or indirect communication pathways.
  • Broader Input Integration: The study focused on somatosensory feedback. The motor cortex also receives inputs from other brain regions. Future work could explore how the subspace framework extends to integrating multiple inter-regional inputs.
  • Experimental Tools: The authors suggest that experimental tools enabling precise, real-time manipulation of neural circuits would be valuable for further investigating the role of subspaces in sensorimotor integration.

Summary updated

Abstract

ABSTRACT Behavior relies on continuous influx of sensory information about the body. In primates, motor cortex must integrate somatic feedback to accurately reach and manipulate objects. Yet, prior work demonstrates that motor cortex is well-described with deterministic, rather than input-driven, dynamics. Deterministic dynamics facilitate robust movement generation, but flexible motor output requires rapid responses to unexpected inputs. Here, we resolved this paradox by simultaneously recording neural population activity in motor and somatosensory cortex from four monkeys performing a naturalistic object interaction behavior resulting in occasional errors. Motor cortex was strikingly input-driven surrounding behavioral error correction. Intriguingly, input-driven dynamics were isolated to a subspace of the population activity that putatively captured somatosensory feedback. Using electrical stimulation of ascending somatosensory tracts, we causally verified that this feedback subspace captured peripheral inputs to cortex. Our results demonstrate that cortical activity is compartmentalized within distinct subspaces, enabling flexible integration of salient inputs for robust behavior.
BibTeX
@misc{perich2020motor,
  author = {Matthew G. Perich and Sara Conti and Marion Badi and Andrew Bogaard and Beatrice Barra and Kanaka Rajan and Jocelyne Bloch and Grégoire Courtine and Silvestro Micera and Marco Capogrosso and Tomislav Milekovic},
  title = {{Motor cortical dynamics are shaped by multiple distinct subspaces during naturalistic behavior}},
  year = {2020},
  doi = {10.1101/2020.07.30.228767},
  howpublished = {bioRxiv},
}

Loading related content...

Figure