Reviewing theory and experimental evidence, they suggest that spontaneous BOLD activity may be more closely aligned with off-line plasticity and homeostatic processes than on-line fluctuations in cognitive content

Brain activity is not only for thinking. Timothy OLaumann, Abraham ZSnyder. Current Opinion in Behavioral Sciences, Volume 40, August 2021, Pages 130-136. https://doi.org/10.1016/j.cobeha.2021.04.002

Abstract: The human brain is a complex organ with multiple competing imperatives. It must perceive and interpret the world, incorporate new information, and maintain its functional integrity over the lifespan. Neural activity is associated with all of these processes. Spontaneous BOLD signals have been invoked as representing neural activity associated with all of these processes. However, their exact role in these processes remains controversial. Here, we review learning machine theory, molecular mechanisms of synaptic plasticity and homeostasis, and recent experimental evidence to suggest that spontaneous BOLD activity may be more closely aligned with off-line plasticity and homeostatic processes than on-line fluctuations in cognitive content.

Introduction

The existence of unceasing spontaneous brain activity has been recognized since at least the 1930s [1]. However, the functions of this activity have remained mysterious [2]. Over the last three decades, blood oxygen level dependent (BOLD) fMRI has become the dominant tool for measurement of brain activity in humans. Soon after the adoption of fMRI, it was observed that fMRI signals exhibit constant fluctuations unrelated to the task [3]. In the context of task fMRI, this activity was conventionally regarded as ‘physiological noise’ [4]. However, it is now clear that this ‘physiological noise’ is temporally correlated within functional systems [5,6]. It is this property of spontaneous brain activity that constitutes the basis of resting state functional connectivity (RSFC) [7]. The existence of this well-structured organization implies that spontaneous brain activity is physiologically consequential.

The meaning of spontaneous BOLD signal fluctuations has been variably interpreted along two different perspectives. According to one view, spontaneous BOLD fluctuations are proposed to reflect unconstrained cognitive processes, for example, retrospection, prospection, reflection, environmental monitoring — the ‘stream of consciousness’ — attendant to our subjective experience. Given the centrality of perception and action to mental life, it is appealing to assume that all observed brain activity is directly related to moment-to-moment cognition and behavior. This perspective has been reinforced by a massive accumulation of PET and fMRI experiments in which brain activity has been imaged with the objective of localizing cognitive operations [8]. More recently, the observation of ‘dynamic’ functional connectivity during wakeful rest and changes in functional connectivity between rest and task states have, at times, been interpreted as reflecting cognition [9,10].

We have previously articulated several problems with the notion that all ongoing BOLD activity directly reflects cognition and behavior (Box 1; [11]): (1) The topography of BOLD fMRI correlations remains largely intact during slow-wave sleep [12] and even anesthesia [13], states in which cognition is presumed to be either absent or greatly attenuated; (2) The extent to which task paradigms modify the correlation structure of spontaneous BOLD signal fluctuations is very limited [14,15,16•]; (3) While unconstrained cognition might be expected to vary from scan to scan within an individual, RSFC remains remarkably consistent across sessions [17,18]. RSFC is also relatively stable within a given scan, discounting fluctuations attributable to drowsiness [11] or arousal [19], which likely relate to fluctuations in BOLD signals, at least partly due to alterations in respiratory behavior and pCO2 [20]. Moreover, brain metabolic activity is high at all times and minimally affected by task performance [21].

Box 1

Evidence that RSFC structure is largely independent of cognitive content

1.

RSFC structure is similar during wake and sleep. For instance, DMN structure is observed through wake, S1, S2, and SWS [12,100].

2.

RSFC structure is present under anesthesia [13,101], although covariance does diminish with increased sedation [102].

3.

RSFC structure is minimally altered by task state [14, 15].

4.

RSFC structure within subject is consistent across sessions [16,11,103].

5.

RSFC structure within subject is similar over long time scales [18,17].

6.

RSFC structure is consistent across subjects at the population level [104,105].

7.

Similar RSFC structure is evident across mammalian species [106, 107, 108].

For all of these reasons, unconstrained cognition does not fully explain ongoing spontaneous activity. An alternative view proposes that spontaneous BOLD activity may more closely relate to mechanisms associated with learning and memory [22•,23]. In the following, we review prior literature supporting the perspective that a substantial fraction of spontaneous brain activity represents homeostatic and consolidative signaling, the function of which is to enable neural plasticity while maintaining the brain's functional integrity through time. We also review recent evidence that BOLD RSFC may be intimately tied to these processes.

Learning machine theory

When considering the role of ongoing neural activity in brain function, it is important to recognize that one of the brain’s primary capacities is its ability to learn new information about its environment. Theoretical considerations, initially formulated by David Marr [24], suggest that any associative learning machine functions optimally if it is allowed to alternate between two states: (1) a learning phase, during which the machine is connected to inputs and connections are enhanced between simultaneously active elements and (2) a restorative phase during which the machine is disconnected from inputs and connections between elements are rebalanced in a manner that increases randomness (entropy) [25,26•]. In multi-layer perceptrons, this principle is expressed as iterative alternation between a forward phase, during which prediction error is evaluated, and a backward phase, during which connection weights are adjusted by back-propagation [27]. The starkest expression of the state alternation principle in living organisms is sleep versus wake. This alternation appears to be necessary: all organisms capable of learning alternate between wake versus sleep states [28]. In vertebrates, events experienced during wake are registered in the hippocampus and the cerebral cortex [29,30]. During slow wave sleep (SWS), reactivation of the same circuits leads to the creation of stable (consolidated) episodic memory [31•].

Understanding how state alternation is implemented in brains requires consideration of the cellular and molecular events underlying synaptic weight modification. Activity-dependent synaptic plasticity plays a crucial role in brain development well before birth [32, 33, 34]. For example, retino-tectal connections have been shown to be sculpted by spontaneous retinal waves during prenatal development of the visual system [35]. Following birth, spontaneous activity continues to refine neural connections using sensory feedback [36,37]. During early life critical periods, experience-dependent synaptic plasticity tunes the response properties of cortical sensory neurons (e.g. ocular dominance columns) [38]. As the brain matures, metabolic ‘brakes’ limit neural plasticity to mechanisms centered on inhibitory interneurons [39, 40, 41]. Although neural plasticity in adults is more restricted, the underlying activity-dependent processes likely follow similar principles.

Molecular mechanisms of activity-dependent synaptic plasticity

Activity-dependent synaptic plasticity is conventionally discussed under the headings of long-term potentiation (LTP) and long-term depression (LTD). But LTP/LTD are deceptively simple terms encompassing a wide range of molecular processes [42,43]. The early phase of LTP (E-LTP) is triggered by Ca2+ influx linked to post-synaptic depolarization, which sets in motion molecular cascades mediated by phosphorylation and dephosphorylation of regulatory molecules (e.g. protein kinase C (PKC) and Ca2+-calmodulin-dependent protein kinase (CamKII)) that govern neurotransmitter receptor trafficking. E-LTP lasts 1−3 hours and is independent of gene expression. The late phase of LTP (L-LTP) begins with the transcription of immediate early genes (IEGs; e.g. Arc, Zif268) that control translational processes, which lead, on a time scale of hours, to structural changes in dendritic spines [44, 45, 46]. Thus, whereas electrophysiological event-related responses may last up to a few hundred millisec and BOLD hemodynamic responses typically evolve over ∼16−20 s, the metabolic traces of the evoked activity persist over much longer time scales. These traces may underlie the observation that fMRI responses to task A are modulated by having performed unrelated task B during the past half hour [47].

The Hebbian principle (‘fire together → wire together’) is often invoked to account for resting state functional connectivity [48,49]. The mechanism underlying Hebbian learning, that is, spike-timing dependent synaptic plasticity (STDP), has been elucidated in considerable detail [50,51]. In brief, neural back-propagation of depolarization induced by a first excitatory stimulus removes the Mg2+ block at NMDA receptors, thereby allowing a second stimulus (if it occurs within a 20–85 ms window) to induce local Ca2+ entry, which initiates the LTP molecular cascade, ultimately reinforcing the association between the paired stimuli. Hebbian mechanisms undoubtedly play a central role in adult learning. Accordingly, it is reasonable to posit that synchronous spontaneous BOLD fluctuations that give rise to RSFC are due to a history of prior co-activation. However, a system dominated by unopposed Hebbian plasticity inevitably becomes either infinitely active or silent.

In contrast to Hebbian plasticity, which adjusts synaptic weights in the same direction as an applied stimulus, the brain also employs various mechanisms of homeostatic plasticity, which adjusts synaptic strengths in the opposite direction to return excitatory/inhibitory (E/I) balance and mean firing rate to prior set points [52]. Homeostatic plasticity includes cell-autonomous mechanisms that directly adjust neuronal excitability to counteract environmental stimuli, as well as multiplicative synaptic scaling, which preserves relative strengths between neighboring synapses, thereby maintaining currently represented information [53•]. These homeostatic mechanisms operate at the level of dendritic branches [45], individual neurons [53•], and large-scale circuits [54], and are active over multiple time scales [55,56]. A correlate of these homeostatic processes is ongoing turnover of synaptic proteins and lipids with half-lives on the order of ‘minutes, hours, days, weeks’ [57]. Modeling experiments suggest that homeostatic regulation of E/I balance plays a crucial role in maintaining the characteristic features of spontaneous brain activity [58]. Importantly, synaptic homeostasis is inseparable from consolidation, the process whereby brief changes in neural activity ultimately lead to stable memory [59,60•]. Thus, it is reasonable to posit that spontaneous activity includes both Hebbian and homeostatic signaling.