Intrinsic timescales as an organizational principle of neural processing across the whole rhesus macaque brain

  1. Ana MG Manea
  2. Anna Zilverstand
  3. Kamil Ugurbil
  4. Sarah R Heilbronner
  5. Jan Zimmermann

  • Curated by eLife

    eLife logo

    Evaluation Summary:

    Neural activity measured in both electrophysiological and functional neuroimaging experiments are often temporally correlated, and the timescales of such correlation in ongoing neural activity, or intrinsic neural timescales, show a hierarchical pattern across the cortical surface. The present study establishes a close link between these timescales and functional connectivity in the brains of non-human primates, suggesting that temporal autocorrelation is an important organizing feature of large-scale neural activity.

    (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. Reviewer #1 agreed to share their name with the authors.)

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

Hierarchical temporal dynamics are a fundamental computational property of the brain; however, there are no whole brain, noninvasive investigations into timescales of neural processing in animal models. To that end, we used the spatial resolution and sensitivity of ultrahigh field functional magnetic resonance imaging (fMRI) performed at 10.5 T to probe timescales across the whole macaque brain. We uncovered within-species consistency between timescales estimated from fMRI and electrophysiology. Crucially, we extended existing electrophysiological hierarchies to whole-brain topographies. Our results validate the complementary use of hemodynamic and electrophysiological intrinsic timescales, establishing a basis for future translational work. Further, with these results in hand, we were able to show that one facet of the high-dimensional functional connectivity (FC) topography of any region in the brain is closely related to hierarchical temporal dynamics. We demonstrated that intrinsic timescales are organized along spatial gradients that closely match FC gradient topographies across the whole brain. We conclude that intrinsic timescales are a unifying organizational principle of neural processing across the whole brain.

Article activity feed

  1. Version published to 10.7554/elife.75540 on eLife
  2. Version published to 10.1101/2021.10.05.463277v3 on bioRxiv
  3. Version published to 10.1101/2021.10.05.463277v2 on bioRxiv
  4. eLife

    Evaluation Summary:

    Neural activity measured in both electrophysiological and functional neuroimaging experiments are often temporally correlated, and the timescales of such correlation in ongoing neural activity, or intrinsic neural timescales, show a hierarchical pattern across the cortical surface. The present study establishes a close link between these timescales and functional connectivity in the brains of non-human primates, suggesting that temporal autocorrelation is an important organizing feature of large-scale neural activity.

    (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. Reviewer #1 agreed to share their name with the authors.)

  5. eLife

    Reviewer #1 (Public Review):

    The concept of Intrinsic neural timescale (INT) has recently emerged as an important dimension and organizing principle for cortical hierarchy, but how it is reflected and measured by the functional MRI has not been thoroughly tested and compared to the INT based on the single neuron activity in the same species. Manea and colleagues measured INT from anesthetized rhesus monkeys using the resting-state fMRI in a high-field (10.5T), and found that they show patterns consistent with the previous electrophysiological measurements and are correlated with the anatomical gradient in the functional connectivity within and between different cortical areas. These results provide robust empirical foundation for broader applications of INT to probe variability and heterogeneity of cortical functions. The analytical methods and statistical models used to measure INT have some weaknesses, and the authors should discuss the effects of anesthesia on the main conclusions.

  6. eLife

    Reviewer #2 (Public Review):

    Manea and colleagues present an analysis of autocorrelation in BOLD timeseries in anesthetized monkeys collected with high-field fMRI. Using a measure (INT) related to autocorrelation timescale (but see concern below), they demonstrate that inter-regional differences in INT follow patterns observed in prior studies measuring autocorrelation (intrinsic) timescales using single-neuron spike train recordings. They demonstrate in frontal and parietal lobes that INT follows topographies of functional connectivity variation. In addition to comparing cortical regions, they observe topography of INT variation within striatum.

    This study will be of broad interest to systems neuroscientists and neuroimagers. Prior studies have characterized intrinsic timescales of resting-state BOLD in human cortex and observed topography related to hierarchy. Establishing this in non-human primate allows closer linking to prior observations in neuronal recordings, and potentially opens up research directions to probe the origins of intrinsic timescales (e.g., through causal perturbation).

    I have two methodological concerns which could be addressed, one related to the INT measure, and the other related to the functional connectivity gradients.

    1. INT measure: They authors put forth the INT measure as related to intrinsic timescale. INT is defined as the integrated area under the autocorrelation function (ACF) up until the first point where the ACF goes below zero. This is different than how intrinsic timescale has been measured in single-neuron spike trains or in prior fMRI studies. While a longer timescale would be expected to increase INT, the problem is that INT (as an integrated area) combines effects of autocorrelation timescale and autocorrelation amplitude.

    - It would be insightful to visualize INT properties at the whole-brain or whole-cortex level (instead of only a single lobe), including (i) INT values themselves, (ii) the lag-one autocorrelation value (reflecting autocorrelation amplitude, and (iii) the zero-crossing lag time used to compute INT.

    - A highly relevant paper (which is not currently cited) is Ito et al. (2020) NeuroImage, "A cortical hierarchy of localized and distributed processes revealed via dissociation of task activations, connectivity changes, and intrinsic timescales". Fig. 5 of Ito shows a cortex-wide map of intrinsic timescale as defined in single-neuron studies (i.e. fitting time constant of decay). Fig. 6 then shows this is related to cortical hierarchy as reflected in the T1w/T2w map (which in principle could be tested by the authors here too). Ito's analysis was performed on the parcellated timeseries, not the voxel level as in the present study, which is a notable methodological difference.

    - That INT combines effects of timescale and amplitude would not be a problem if the autocorrelation amplitude does not vary across brain regions. However, it appears that it does for whatever reasons (neural and/or in fMRI measurement such as SNR). A relevant preprint is by Shinn et al. (2021) bioRxiv, "Spatial and temporal autocorrelation weave human brain networks". In human cortex, again using parcellated timeseries, Fig. 1F there shows systematic variation across cortical parcels of the lag-1 autocorrelation value. In the present study, it is currently unknown whether INT is reflecting regional differences in autocorrelation timescale (as interpreted), amplitude (not considered), or both.

    - Contribution of autocorrelation amplitude to INT may potentially explain why a cortex-wide map of INT does not follow an expected hierarchy as much the more restricted views within one lobe as the current manuscript focuses. For instance, Fig. 1 shows that INT values for somatosensory cortex (Fig. 1A) are larger than association regions (Fig. 1B). Is this potentially due to autocorrelation amplitude being larger in somatosensory cortex?

    - Perhaps some smoothing or parcellation would be required to better tease apart autocorrelation timescale from autocorrelation amplitude.

    2. Functional connectivity gradients: Figures 3 and 4 rely on functional connectivity gradients calculated within a single lobe, against which INT topography is correlated. My concern here is that on such a restricted geometry as a single lobe, a functional connectivity gradient may be reflecting a simpler property, namely the geometry of the restricted cortical sheet. In other words, given the sheet geometry of the frontal lobe, does an anterior-posterior topography fall out naturally as the first gradient (e.g. with distance-dependent falloff) and medial-lateral as second gradient? If so, it is difficult to strongly interpret these results as linking INT to functional connectivity when the gradient is a generic consequence of the sheet geometry. In human neuroimaging such functional gradients are typically calculated at the whole-cortex level which reveals less trivial topographies (e.g. Margulies et al., 2016, PNAS). These results and interpretations should be considered in light of this concern.

  7. eLife

    Reviewer #3 (Public Review):

    This is an influential paper that establishes the utility of fMRI for studying the hierarchy of temporal dynamics across the macaque brain. The authors demonstrate that the time constants of BOLD responses in different cortical regions have the same ranking as those previously discovered with electrophysiological measurements of spiking activity. This paper extends previous studies by providing whole-brain maps of temporal hierarchy, showing a close correspondence with the hierarchies inferred from a variety of functional connectivity as well as anatomical measurements. Overall, this is a strong paper with deep technical and scientific implications for the field. However, there are interpretational and technical concerns that I would like to see addressed.

    Does the calculation of fMRI-based neuronal time constants obscure the unit of time? True comparison with ephys data is not possible without clarifying the relationship of the two quantities compared. In the ephys measurements time constants are in units of seconds and often below 1s. In contrast, BOLD response has a sluggish time course (tens of seconds) due to the properties of the hemodynamic response function. The smoothing of spiking and field-potential activity with the HRF introduces substantial auto-correlation in BOLD and is expected to reduce our ability to distinguish small differences of time constants discovered with ephys. Because the analyses in this paper do not explore the complications caused by the slow and noisy BOLD measurements, it is impossible to know if the observed temporal hierarchy has the same nature and origin as those reported with ephys. I would love to see additional analyses and modeling that clarifies this missing link. If that is not possible, at the very least I would recommend explicit reporting of the units of time constants based on BOLD in the figures, and discussing if the differences of BOLD time constants across regions match the differences of spiking activity time constants in previous publications.

  8. Version published to 10.1101/2021.10.05.463277v1 on bioRxiv