Simultaneously Recording Up to a Million Neurons’ Cortical Dynamics Reveals Unbounded Scaling of Neuronal Quantity and Dimensionality

Summary

This scientific report titled “Simultaneously Recording Up to a Million Neurons’ Cortical Dynamics Reveals Unbounded Scaling of Neuronal Quantity and Dimensionality,” published in the journal Neuron (Volume 112, pages 1694–1709), is authored by Jason Manley, Sihao Lu, Kevin Barber, Jeffrey Demas, Hyewon Kim, David Meyer, Francisca Martínez Traub, and Alipasha Vaziri. The article was released on May 15, 2024, as a result of collaborative research by Rockefeller University and The Kavli Neural Systems Institute. The study delves into the relationship between the dimensionality of neuronal dynamics and the quantity of neurons, revealing significant insights into neuronal computation mechanisms.

Background

Neuron clusters are considered not as independent processors but as interconnected units supporting adaptive and goal-oriented behavioral computations. Due to past technical limitations, records were usually confined to a small number of neurons, leading to the construction of a series of key paradigms based on measurable attributes of individual neurons. Corresponding theoretical frameworks aimed to infer how single-cell responses emerged from large-scale, unobservable neural circuits. With the development of large-scale neuron recording techniques, monitoring the activities of larger neuron clusters in behaving animals has become possible.

The widespread application of dimensionality reduction techniques suggests that neural dynamics can be approximated by low-dimensional “latent” signals, reflecting neural computation. However, can this low-dimensional representation genuinely explain the broad range of brain activities? If not, what is the appropriate recording resolution and scale necessary to capture these activities?

Research Methods

In the study, the authors used a technique called light beads microscopy (LBM) to record the dynamics of up to one million neurons in the mouse cortex. In this technique, the fluorescent activity of neurons in a three-dimensional field of view is captured by a column of 30 axially separated, temporally distinct two-photon excitation points. These light beads densely capture neural activities within the axial range to provide spatially and temporally optimized acquisition, limited only by the fluorescent lifetime of the genetically encoded calcium indicators. The research subjects were transgenic mice expressing gcamp6s or gcamp6f.

Research Results

The study found that the recorded dynamic dimensionality of neuron clusters scaled with the number of neurons in an unbounded power-law relationship. Despite 50% of neuron variability being contained within 16 behavior-related dimensions, these dimensions were unrelated to immediate behavioral or sensory indicators. High-dimensional components exhibited unique temporal structures, spanning time scales from seconds to the temporal resolution limits. Additionally, these high-dimensional activities were almost orthogonal to sensory-induced patterns, indicating that these dimensions addressed neural computations lacking immediate sensory or behavioral relevance.

Specifically, the authors used a shared variance component analysis (SVCA) method to separate all neuronic activities in cortical records into two parts and identified linear dimensions of two groups of neuronal activities, known as Shared Variance Components (SVCs). Using the covariance of these dimensions for reliability testing, the authors found that the SVCA method more accurately isolated signals of potential biological significance than traditional dimensionality reduction techniques.

Through SVCA analysis of the cortex hemisphere records of 146,741 neurons, the authors discovered that dimensionality increased following a power-law to the maximum observable neuron cluster quantity, similarly in affine distributions and under different recording densities. This result indicates that, regardless of sampling strategy, cortical dimensionality is primarily dependent on the number of recorded neurons rather than their spatial distribution within the volume or cortical regions.

Significance

The study’s conclusion highlights the high dimensionality of neuronal computation, emphasizing the importance of large-scale, cell-level recording techniques in revealing the complete neural computing matrix. This implies that understanding the high-dimensional geospatial dynamics of the cortex requires larger-scale observations of neuronal activities. Although over 90% of the reliable neural dimensions are unrelated to the animal’s immediate behavior, this does not mean that these high dimensions are useless to neural functions. The results also show that traditional variance-based methods may significantly underestimate neural dimensionality and stress the need for further large-scale recordings.

By understanding the dimensional characteristics of these high dimensions, scientists can further explore how neural circuits encode and process information. Additionally, the novel techniques and methods used in this study could open up new avenues for future research on how intrinsic states like attention, motivation, hunger, or fear affect neural computation.

This research not only provides new perspectives on neural computation but also demonstrates the need for high dimensionality and large-scale cell recording combinations in future studies. It offers new research directions for understanding how the brain processes complex information.