Neuronal correlate of BOLD signal fluctuations at rest: Err on the side of the baseline

Functional MRI (fMRI) indirectly measures changes in neuronal activity because the blood oxygenation level-dependent (BOLD) signal is sensitive to changing concentrations of oxyhemoglobin (vs. deoxyhemoglobin) to support functional energy demand (1). Changes in the BOLD signal are usually interpreted from an unspecified baseline state (2). However, there is no true baseline because the brain is never actually at rest (3), measured either in terms of neuronal activity (4) or the energy that activities demand (5). Recently, however, investigators have begun to study the resting state using fMRI, but interpretation remains controversial because of questions about the relationship between the BOLD signal and neuronal activity. In PNAS, Schölvinck et al. (6) tackle this controversy by correlating slow modulations of neuronal activity in the resting state with spontaneous fluctuations in the BOLD signal.

Neuroscientists typically use fMRI with task-based paradigms (or T-fMRI), in which the mean of the baseline state is subtracted from the mean of the stimulated state to unveil activated (or deactivated) regions associated with the task (1). T-fMRI experiments in the human brain generally report small evoked changes in the BOLD signal, which peak within ∼6 s after task onset. The magnitude of the evoked BOLD response varies with the task type (e.g., sensory and cognitive) and with the cortical area (i.e., >1% and <1%, respectively, in primary sensory and high-order areas). However, fMRI is also used to study the brain at rest in the absence of any explicit task. Biswal et al. (7) observed that resting human brain fMRI data contain high-amplitude (∼1%), low-frequency (<0.1 Hz) fluctuations in the spontaneous BOLD signal that are temporally correlated across vast spans of cerebral cortex. In the resting-state fMRI paradigm (or R-fMRI), in which the data are analyzed for spatiotemporal coherence to reveal correlated networks, a preprocessing step is used to regress out contributions from the global BOLD signal fluctuations (8). This process presumably eliminates “noise” from nonneuronal sources (9) and/or uncorrelated neuronal activities. The remaining smaller (or filtered) fluctuations in the spontaneous BOLD signal facilitate detection of network-level correlations (10). These tiny fluctuations in the spontaneous BOLD signal are often assigned to neuronal activity that supports networks (e.g., the default mode) and are believed to be representative of resting brain function rather than the total neuronal activity that characterizes the baseline state (2). If, however, there is a significant correlated neuronal component associated with the global BOLD signal, important information about resting-state brain connectivity is being discarded by this process.

Schölvinck et al. (6) demonstrate that the global component of BOLD signal fluctuations measured at rest is indeed tightly coupled with a slow modulation of neuronal events that appear to be ubiquitous in the cerebral cortex. They thereby recommend caution when arbitrarily removing the global BOLD signal because in doing so a global correlate of the brain's baseline neuronal activity is thrown away, which in turn may affect regions that are defined as either correlated or anticorrelated. Thus, this study has strong ramifications for interpretation of default mode or other networks using R-fMRI data that are assigned on the basis of the removal of the global average fluctuations. In addition, it is also likely that this study may impact interpretation of T-fMRI data in cases in which task-specific BOLD signal increases (or decreases) are detected from the baseline state because the evoked changes (e.g., with cognitive tasks) may be on the order of the spontaneous fluctuations.

Although correlations between slow (<0.1 Hz) modulations of ongoing neuronal activity, as measured by local field potential (LFP) or multiunit activity (MUA), and fluctuations of the resting BOLD signal have been previously reported both locally near the microelectrode and extending over regions of the visual cortex (11), Schölvinck et al. (6) show that these correlations extend over nearly the entire cortical surface with a correlation strength that is not obviously related to the position of the electrode. They simultaneously recorded LFP and cerebral blood volume (CBV)-weighted fMRI signal (see ref. 12 for details) from the resting (awake) primate brain and compared a regional fMRI signal to the slow temporal variations in the power of the LFP in low-, intermediate-, and high-frequency bands (see ref. 13 for frequency distributions of neuronal activities). Slow fluctuations of the spontaneous neuronal activity—in either high- or low- but not intermediate-frequency LFP bands—measured from a single cortical site in one hemisphere exhibited widespread correlations with spontaneous fluctuations in fMRI signals. Global patterns of these spatial correlations were quite similar whether the LFP was measured from the frontal, parietal, or occipital cortices of the primate brain.

How much of the spontaneous fluctuations in the fMRI data can be accounted for by the slow modulation of neuronal events? Schölvinck et al. (6) estimate that a considerable portion of the variance in their fMRI signal is related to the slow modulation of neuronal events represented by the high- and low-frequency LFP bands. However, it is worth speculating that MUA, not measured in this study, could also account for a fraction of the variance in the fMRI signal because of its coupling with LFP both at rest (11) and during task (14). The fMRI signal peaked 5–8 s from the onset of modulations of the power in the high-frequency LFP band (6). This finding is in agreement with prior results from the resting primate brain where the BOLD signal peaked ∼6 s from the neuronal event as measured by LFP or MUA (11). Furthermore, T-fMRI studies in rodent and primate brains showed that both the BOLD and the CBV peak 5–7 s after task onset, and these signals are correlated dynamically with LFP or MUA (14, 15). Together, these results suggest that the neurovascular coupling associated with the slow neuronal component and the global BOLD signal may be similar in R-fMRI and T-fMRI.

A reverberating theme in the study of Schölvinck et al. (6) is that global baseline activity has value in interpretation of fMRI data from both R-fMRI and T-fMRI experiments. However, how large are the slow spontaneous fluctuations in relation to global baseline neuronal activity? What fraction are they of the total? We can begin to address these questions from brain energetic studies by using ¹³C magnetic resonance spectroscopy (MRS), calibrated fMRI, and PET (see ref. 12 for details).

Estimates of total baseline neuronal activity and the demanded energy can be determined from ¹³C MRS, which allows simultaneous measures of energy demand (CMR_O2) in neurons and glia along with rates of neurotransmission (i.e., neuronal activity) as reflected by presynaptic release of glutamate/GABA and recycling through glia (16). [Glutamatergic and GABAergic synapses account for 90% of the total in the adult mammalian cerebral cortex (17).] ¹³C MRS studies in rats established a quantitative relationship between neuronal activity and neuronal energy demand (see ref. 17 for a review of recent studies). When this energy–activity relationship is extrapolated to the awake resting state, it is found that ∼80% of the neuronal energy in the cerebral cortex supports the global neuronal activity at rest (3). Furthermore, it has been suggested that the high global baseline energy—and by inference neuronal signaling—is evenly distributed throughout the human cerebral cortex (18).

We estimate that the spontaneous fluctuations in the global baseline energy are smaller in comparison. No previous studies have investigated this matter specifically because temporal fluctuations of CMR_O2 are difficult to measure directly with PET or ¹³C MRS. However, an estimate may be obtained from calibrated fMRI (see refs. 12 and 19 for details) by using the amplitude of spontaneous fluctuations of the BOLD signal. On the basis of calibrated fMRI measurements (19), approximately 1% amplitude fluctuations in the BOLD signal (7) correspond to, at most, 10% variations in CMR_O2 from baseline. We note that these CMR_O2 fluctuations may be comparable to energetic demand during cognitive tasks (see ref. 20 for details). Thus, it is likely that the slow modulation of the high-frequency LFP band identified by Schölvinck et al. (6) may represent a nonnegligible fraction of the total baseline energy (and by inference neuronal activity).

Neuronal signaling in the cerebral cortex is extremely heterogeneous (13, 21). However, neuroscience studies traditionally focus on neurons that show the largest change in firing (with a task) as opposed to a representative neuronal ensemble symbolizing mixed activities (5, 12). Thus, it is useful to characterize the behavior of a neuronal ensemble by a histogram of their respective signaling speeds and an estimate of the ensemble's relative energy demand (or CMR_O2) can be derived by integrating the histogram (Fig. 1). Estimates of energy demand from neuronal histograms agree reasonably well with energy consumption measured by techniques like calibrated fMRI, ¹³C MRS, and deoxyglucose autoradiography (22). Schölvinck et al. (6) identify the slow modulations of neuronal activity with different LFP frequency bands. It is fascinating to contemplate what fraction of the total neuronal population is responsible for these slow modulations. Is it a small fraction of the population undergoing large variations or a large portion undergoing small deviations?

Fig. 1.

Schema of neuronal histogram and neuronal energetics. Histograms of total neuronal activity represented by distribution of firing rates (ν) of the same neuronal ensemble, composed of many neurons, in low (blue; left) and high (green; right) baseline energy states. The histograms can be converted to CMR_O2 by multiplying the number of neurons firing at a given rate over the entire range of frequencies. See refs. 5 and 22 for details on neuronal histogram and neuronal energetics.

At present, R-fMRI and T-fMRI studies disregard the importance of global baseline activity and its spontaneous fluctuations. Schölvinck et al. (6) show that there is a significant neuronal correlate to these BOLD signal fluctuations. The global baseline neuronal activity and its requisite energy demand are extremely high for the cerebral cortex of the awake human (3, 18), and as discussed above, the spontaneous fluctuations may be on the order of changes generally observed in cognitive tasks. In light of these findings, neither the total baseline neuronal activity nor its fluctuations may be neglected as merely representing uncorrelated neuronal activity and/or nonneuronal factors. As online human databases of T-fMRI (www.brainmap.org) and R-fMRI (www.nitrc.org/projects/fcon_1000) experiments grow, it is ever more vital to err on the side of the baseline for fMRI studies as suggested by Schölvinck et al. (6).