Structure and variability of delay activity in premotor cortex

Nir Even-Chen, Blue Sheffer, Saurabh Vyas, Stephen I. Ryu, Krishna V. ShenoyPLOS Computational Biology 2019 · DOI

Abstract

Preparatory neural activity in the motor cortex is hypothesized to set the "initial conditions" for the dynamical system that executes movement. However, the population-level organization of these conditions and the impact of neural noise on movement precision remain poorly understood. This study recorded neural population activity from the dorsal premotor cortex (PMd) of rhesus macaques during an instructed-delay reaching task with dense target configurations. We demonstrate that preparatory activity converges onto a structured, low-dimensional manifold where two dimensions represent the spatial endpoint and a third represents movement speed. Crucially, trial-to-trial variability in these neural states mirrors the elliptical pattern of behavioral reach errors, with greater noise observed in the distance dimension than in direction. Furthermore, by employing an arc-length metric, we show that reach distance is as decodable as direction, challenging the long-held belief that the motor system prioritizes directional information. These findings provide a blueprint for the neural representation of movement planning and offer critical insights for the design of high-performance brain-machine interfaces (BMIs).

Introduction

Before a reach is initiated, the brain undergoes a period of neural preparation. The "initial condition hypothesis" suggests that activity during this delay period sets the starting state for the neural dynamics that drive the subsequent movement. Despite the importance of this phase, the internal structure of these preparatory states—specifically how parameters like distance, direction, and speed are organized—has not been fully characterized. Moreover, the relationship between internal neural variability and the precision of the resulting movement is a fundamental question in motor control. Historically, researchers have found it easier to decode movement direction than distance, leading to the assumption that the brain's representation of direction is more robust. This study aims to map the geometry of preparatory activity in the PMd and determine how its structure and noise influence behavioral accuracy.

Methods

Experimental Design

Neural population activity was recorded from the dorsal premotor cortex (PMd) of two rhesus macaques (Monkeys J and R) using 96-electrode Utah arrays. The monkeys performed an instructed-delay reaching task where a target appears on a screen, but movement must be inhibited until a "go cue" is provided. To create a high-resolution map of the neural space, the researchers utilized dense target configurations, including rings (24–36 targets), horizontal/vertical lines (24 targets), and concentric circles (48 targets).

Data Analysis

The study utilized large-scale datasets comprising over 25,000 successful trials. Key technical approaches included:

  • Dimensionality Reduction: Principal Components Analysis (PCA) was applied to the firing rates of nearly 100 neurons to identify the "latent" dimensions that capture the most variance in the population activity.
  • Decoding and Regression: Linear regression was used to map neural dimensions to reach kinematics ($x, y$ coordinates and speed). Support Vector Machines (SVMs) were used to classify the intended target from the neural state.
  • Arc-Length Metric: To evaluate decoding performance fairly across different distances, the researchers used "arc-length"—the physical distance between the intended and decoded points along the curve of the target layout—rather than simple angular error.

Results

The Low-Dimensional Map of Movement Plans

The results reveal that preparatory activity in the PMd is highly structured. Within 300–400 ms of target onset, neural activity converges onto a stable preparatory state. This state can be described by a low-dimensional manifold: two dimensions represent the spatial endpoint ($x, y$), and a third, independent dimension represents the intended maximum speed of the reach. This suggests that the brain plans the spatial and temporal components of a movement using distinct neural channels.

Neural Noise Mirrors Behavioral Variability

A significant finding of this study is the alignment between neural variability and behavioral errors. When reaching without visual feedback, primates tend to be less accurate in distance than in direction, resulting in elliptical error distributions. The researchers found that the trial-to-trial variability of neural states in the PMd has the exact same elliptical shape. This indicates that the "noise" in the brain's initial plan is a primary source of the physical errors made during movement.

Robust Encoding of Distance

The study challenges the notion that direction is "better" encoded than distance. While raw classification accuracy for direction appeared to increase as targets moved further away, this was shown to be a mathematical artifact of target spacing. When using the arc-length metric, decoding error remained constant across all distances. The mean classification error for distance (approximately 2.6–3.2 cm) was comparable to directional decoding when properly normalized. This proves that the PMd provides a balanced and robust representation of the entire reach vector.

Discussion

This study provides a detailed characterization of preparatory activity in the premotor cortex, showing that it forms a structured map of upcoming movement parameters. The discovery that neural variability in the PMd mirrors behavioral error patterns suggests a direct link between the precision of the neural "plan" and the accuracy of the resulting action.

Limitations and Future Directions

While the correlation between neural noise and behavioral error is strong, it does not definitively prove causation. Future experiments involving microstimulation to perturb specific neural dimensions during the delay period could establish whether PMd noise directly causes reach errors. Additionally, while the low-dimensional structure holds for simple center-out reaches, further research is needed to determine if this organization persists during more complex tasks, such as obstacle avoidance. Finally, disambiguating whether PMd encodes abstract kinematics or specific muscle dynamics remains a key challenge for future postural studies.

Implications for BMI Design

These findings have significant implications for the development of brain-machine interfaces. By demonstrating that distance is as decodable as direction and identifying the low-dimensional structure of movement planning, this research provides a framework for building more accurate and intuitive neural prosthetics.

The dataset used in this study is publicly available via Stanford University's servers.

Summary updated

BibTeX
@article{evenchen2019structure,
  author = {Nir Even-Chen and Blue Sheffer and Saurabh Vyas and Stephen I. Ryu and Krishna V. Shenoy},
  title = {{Structure and variability of delay activity in premotor cortex}},
  journal = {PLOS Computational Biology},
  year = {2019},
  doi = {10.1371/journal.pcbi.1006808},
}
Figure