Article activity feed

  1. Evaluation Summary:

    This paper will be of interest to electrophysiologists, systems neuroscientists and neural engineers. The authors describe a framework for evaluating the comparison between LFP dynamics and spikes and perform this comparison for several datasets recorded from motor, premotor, and sensory areas of cortex in rhesus macaque monkeys. These results serve as an important benchmark for the information content of LFP recordings, which is relevant to data collection in neuroscientific investigations and to designing brain computer interfaces.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. The reviewers remained anonymous to the authors.)

  2. Reviewer #1 (Public Review):

    Gallego-Carracedo and colleagues investigated the relationship between neural spiking activity and local field potentials (LFP) across three different sensorimotor areas (dorsal premotor (PMd), primary motor (M1), and area 2 of somatosensory cortex (S1)) during a well-trained motor behavior. In contrast to previous studies, where spiking-LFP relationships were studied at the level of single neurons, the authors explore whether and how mesoscopic signals like the LFP are related to population-level patterns in spiking activity (referred to as "latent dynamics"). This is a very interesting and potentially valuable revisiting of LFP-spiking relationships, since increasing evidence has shifted focus away from purely single-neuron-based analyses towards population perspectives. Insights into relationships between LFP and latent dynamics may also inform interpretations of these signals.

    The largest strength of this paper is the large amount of data. The paper includes analyses of datasets from 3 brain areas (7 implanted arrays total) in 4 animals. This reveals that LFP - latent dynamics relationships vary across brain areas and opens possibilities to fully examine relationships of all signals. The wealth of data allows them to clearly show that LFP-latent relationships are frequency-dependent and vary across brain areas.

    The primary weaknesses of the paper are that it skips some important preliminary analyses, does not fully describe/interpret the broad diversity of data they present, and their interpretation of "stable relationship" is somewhat unclear.

    1. Given the frequency-based analyses presented, more detailed characterization of the LFP spectra will greatly benefit the paper. A key question the authors should address is whether the frequency-dependence (and its variance across areas) is related to differences in power spectra across areas. They present an analysis suggesting their results are not simply explained by variance differences across bands, but there are no analyses to address power differences (and deviations from the 1/f "noise" spectrum).

    2. The data reveal some clear differences between subjects and across areas that are not fully elaborated on. The relationship between decoding performance and LFP-latent correlation appears to only be present in M1. The relationships in PMd and area 2 are not quantified or commented on in much detail. Similarly, across all areas there are notable differences in LFP-latent correlations in some frequency bands (primarily the lower frequencies) between animals that is not addressed.

    3. One of the manuscript's primary claims is that LFP-latent correlations are "stable" within areas while being different between areas. These claims are the main basis of their interpretation that these relationships reflect biophysical properties of the cortical networks (e.g. cytoarchitecture). The claim of stable relationships focus on comparing between motor planning and execution task epochs. These task epochs appear to include partially overlapping time windows based on their methodological description, which seems like a potential confound that should be addressed. The time windows used are also different durations, which should be controlled for. Moreover, their results also show that LFP-latent relationships change (mostly disappearing) in inter-trial intervals. If these correlations truly reflect properties of circuit structure, I am unclear on why they would be task-dependent. This interpretational point needs significant clarification.

  3. Reviewer #2 (Public Review):

    In this paper, Gallego-Carracedo, Perich, Chowdhury, Miller, and Gallego set up an important question: In much of systems neuroscience, researchers record spiking data from populations of single neurons or multi-unit channels to estimate neural population state. Applying dimensionality reduction algorithms like PCA to the high dimensional neural population state yields an estimate of the lower-dimensional latent dynamics, which are commonly understood to be a compact representation of the patterns of activity in the brain. Understanding the relationship between these latent dynamics and behaviors, sensory inputs, or cognition, represents a central goal of systems neuroscience.

    In most such experiments, local field potential (LFP), is often also recorded, as it is simple to do so and these complementary signals may also provide scientific utility. In many other studies, often including studies using human participants, only LFP recordings are possible due to constraints of the neural sensors or recording equipment. Understanding the relationships between the LFP signals and latent dynamics thus represents an important bridge for helping to contextualize studies relying on LFP alone. In addition, better understanding the link between the two recording modalities (i.e. understanding the relative information content contained of each), in principle, could help to elucidate the biophysical mechanisms by which LFP arises from the collective spiking activity of neural circuits.

    The authors outline four central hypotheses: 1) That there should be a robust relationship between LFP and latent dyamics, 2) that this relationship should be frequency dependent, 3) that these relationships are similar between preparation and movement (for data recorded in PMd, M1, and S1), and 4) that different areas should have different relationships between LFP and latent dynamics.

    The motivation for this work is strong, and the quality and breadth of the data sets collected and curated is impressive. With a narrow reading of these hypotheses, the analyses presented here support the authors conclusions. The author's explicit goal is to assess whether any relationship exists between LFP and latent dynamics. Their analysis reveals that for certain frequency bands, in certain brain areas, that information in LFP signals is also contained within the manifold of latent activity.

    While the stated goals, hypotheses, and overall presentation of this paper are all clear, the primary analysis method limits the broader interpretability of the results. The main analysis method that the authors use to assess the relationship between LFP and latent dynamics is the distribution of correlation coefficients derived from applying canonical correlations analysis (CCA) between the latent dynamics and individual channels of LFP, for individual frequency bands on those channels. As described, this method produces a metric for how well signals within a specific LFP frequency band on one channel is represented within the manifold of latent dynamics, allowing for a rotation of that manifold.

    This analysis, however, does not say anything, however regarding the information contained in the manifold of latent dynamics that is not present within LFP signals. An analysis capable of revealing these differences would provide a more actionable takeaway for contextualizing what information is lost when an experiment relies on LFP signals alone. For a concrete hypothetical example, if every individual LFP channel (within one frequency band) contained a signal that perfectly correlated with principal component 1 (or any other PC), then the metric would report a distribution clustered tightly around 1. While this metric cannot get any higher, suggesting a high degree of alignment between LFP and latent dynamics, it appears to ignore the fact that in this contrived scenario, no LFP channels have captured any information about PCs 2,3,4,...,n, and the metric tells us nothing about the information lost by only recording LFP. If we don't know what information is lost, it is difficult to know how to apply these results to contextualize other studies based on LFP recordings, which is one of the stated broader motivations for this paper.

    These limitations aside, the authors have carefully shown that there appears to be a frequency dependence between which LFP bands share similar information with the latent dynamics. In addition, they establish that LFP recordings in PMd, M1, and S1 show different relationships with the latent dynamics, and that the degree of LFP correlation with latent dynamics is stable between the movement preparation and execution. This paper is well written, with extraordinary attention to detail and clarity throughout.