Contribution of animal models toward understanding resting state functional connectivity

Abstract

Functional connectivity, which reflects the spatial and temporal organization of intrinsic activity throughout the brain, is one of the most studied measures in human neuroimaging research. The noninvasive acquisition of resting state functional magnetic resonance imaging (rs-fMRI) allows the characterization of features designated as functional networks, functional connectivity gradients, and time-varying activity patterns that provide insight into the intrinsic functional organization of the brain and potential alterations related to brain dysfunction. Functional connectivity, hence, captures dimensions of the brain’s activity that have enormous potential for both clinical and preclinical research. However, the mechanisms underlying functional connectivity have yet to be fully characterized, hindering interpretation of rs-fMRI studies. As in other branches of neuroscience, the identification of the neurophysiological processes that contribute to functional connectivity largely depends on research conducted on laboratory animals, which provide a platform where specific, multi-dimensional investigations that involve invasive measurements can be carried out. These highly controlled experiments facilitate the interpretation of the temporal correlations observed across the brain. Indeed, information obtained from animal experimentation to date is the basis for our current understanding of the underlying basis for functional brain connectivity. This review presents a compendium of some of the most critical advances in the field based on the efforts made by the animal neuroimaging community.

I. Introduction

Functional magnetic resonance imaging (fMRI) and its sibling resting state fMRI (rs-fMRI) are currently our most powerful tools for noninvasive functional imaging of the whole brain. Together, they have provided insight into systems-level features related to cognition and behavior, and how those features are altered in neurological and psychiatric disorders. Compared to task-based fMRI (Bandettini, Wong et al. 1992, Menon, Ogawa et al. 1992), where the timing of the task or stimulus is known and can be used to identify the brain areas that are activated, rs-fMRI is more difficult to interpret. The blood oxygenation level-dependent (BOLD (Ogawa, Lee et al. 1990)) fluctuations that are produced in response to the brain’s intrinsic activity exhibit spatial and temporal structure that have been summarized into functional connectivity matrices (Allen, Damaraju et al. 2014), functional networks (Yeo, Krienen et al. 2011), and functional connectivity gradients (Margulies, Ghosh et al. 2016), to name a few of the most common interpretations. While rs-fMRI has become immensely popular due to its ability to characterize activity noninvasively throughout the whole brain, the simplicity of the acquisition paradigm, and the lack of reliance on subject cooperation during task performance, it remains poorly understood in the sense that the BOLD fluctuations arise from a combination of neurophysiological processes, and alterations in any single component can affect the downstream calculation of functional connectivity-based features.

BOLD is an integrative signal that reflects a combination of hemodynamic effects (Buxton, Wong et al. 1998, Friston, Mechelli et al. 2000), including total vascular volume and metabolically-driven changes in the ratio of oxygenated to deoxygenated hemoglobin (Silva, Lee et al. 1999, Kida, Kennan et al. 2000), which occur in response to neural activity (Brinker, Bock et al. 1999, Logothetis, Pauls et al. 2001). Differences in functional connectivity observed in separate groups of subjects (healthy vs. patient; fast responder vs. slow responder) could therefore arise from multiple mechanisms (altered anatomical connectivity, altered neural activity, altered vascular properties). Our primary hope of disentangling the BOLD signal lies with multimodal experiments in animal models, where other, often invasive, measurements of particular types of activity can be combined with rs-fMRI. Here we review progress towards a better understanding of the neurophysiology underlying the spontaneous BOLD fluctuations in animal models, with an emphasis on rodents, and highlight areas that are ripe for further exploration.

The simplest approach to understanding how neurophysiological processes give rise to the BOLD fluctuations is to look at localized relationships between neural activity and the resulting hemodynamics. Most of the early studies that provided the first definitive links between neural activity and the BOLD response after presentation of a stimulus took this approach (Brinker, Bock et al. 1999, Logothetis, Pauls et al. 2001), and it has value for understanding the spontaneous BOLD fluctuations as well (Shmuel and Leopold 2008). While it is often assumed that neurovascular coupling is identical whether the activity is intrinsic or task-driven, some studies suggest that the hemodynamic response functions can differ across the two situations (Chen and Glover 2015).

The limitation of focusing on localized relationships between neural activity and the BOLD signal is that the localized approach cannot explain the structured relationships between the BOLD fluctuations from different portions of the brain. Even if we adopt the simplified perspective that correlated BOLD fluctuations reflect correlated neural activity, there are still open issues as to why the neural activity in those brain areas is correlated in the first place. Differences between the structural networks and functional networks derived from the BOLD signal, along with correlations between areas that are known to be weakly connected, are evidence that the simple explanation that correlation reflects communication via a direct white matter connection is inadequate. Going beyond single-site experiments has the potential to reveal not just the neurophysiology of the BOLD signal, but answers to deeper neuroscience questions about the intrinsic functional architecture of the brain.

In this review, we cover mechanisms of localized neurovascular coupling along with investigations that are shedding light into the spatial organization of activity across the brain. For the sake of simplicity, we will often refer to this coordinated activity across different brain areas as functional connectivity, shorthand for the statistical dependencies that are captured by Pearson correlation, independent component analysis (ICA), or other analysis methods.

II. Overview of complementary experimental approaches

Functional neuroimaging in healthy human subjects is confined to noninvasive methods such as electroencephalography (EEG), magnetoencephalography (MEG), fMRI, or positron emission tomography (PET), all of which suffer from a combination of limited spatial resolution, limited temporal resolution, lack of whole brain coverage, and lack of specificity. Luckily, both the hemodynamic response to neural activity and the functional networks measured with rs-fMRI are remarkably similar across species (Fox, Snyder et al. 2005, Vincent, Patel et al. 2007, Mantini, Gerits et al. 2011, Lu, Zou et al. 2012, Belcher, Yen et al. 2013, Barks, Parr et al. 2015, Hsu, Liang et al. 2016, Schroeder, Weiss et al. 2016), making it possible to probe the neurophysiological underpinnings of the BOLD response in animals with a reasonable expectation that the BOLD signal will arise from comparable processes in humans. In animals, tools such as microelectrode arrays, high-resolution optical imaging, and genetic manipulations provide insight into the brain’s functional architecture that cannot be obtained in human subjects (Figure 1).

Figure 1.

Currently, no single model organism can accommodate the full range of experimental approaches. Instead, complementary experimental approaches are applied across model organisms. For each model, some approaches are readily applicable (green), some are applicable to an extent (green with orange outline), some are applicable in rare conditions (red with orange outline), and some are inapplicable (red). Only a few common model organisms are shown; others, such as marmosets, are increasingly available and may help to bridge the gap between rodents and large primates.

A. Optical imaging

One of the most powerful tools for these investigations is optical imaging, which has the unique capability to capture both intrinsic signals primarily linked to hemodynamics and extrinsic fluorescent signals that directly indicate neuronal activity. Optical extrinsic signals (OES) arise from fluorescence introduced using dyes or genetically-encoded indicators. The development of genetically encoded sensors for the last 20 years has had a profound effect on neuronal circuit dissection. These sensors provide the ability to monitor intracellular calcium, transmembrane voltage, neurotransmitters and a variety of chemicals (Palmer, Qin et al. 2011, Oh, Lee et al. 2019, Pal and Tian 2020), and can be introduced in cells with virus injections. Also, thanks to the growing availability of Cre-expressing and reporter mice, sensor expression can be targeted specifically to neuronal subtypes, astrocytes, smooth muscle cells, endothelial cells and pericytes in rodents (Hill, Tong et al. 2015, Daigle, Madisen et al. 2018, Tong, Hill et al. 2021), to provide an unprecedented level of specificity about the role of different components in the generation of the hemodynamic fluctuations detected with rs-fMRI .

Optical measurements are typically limited to the surface of the brain, but with the growing use of wide-field optical imaging, the field of view has become large enough to capture some of the same spatial structure observed with rs-fMRI in superficial brain regions. Wide-field mesoscale calcium imaging allows the imaging of the cerebral cortex and gives insights into the functional parcellation of the cortex in rodents (Vanni, Chan et al. 2017) with a resolution of tens of microns.

Optical intrinsic signals (OIS) can also be mapped using wide-field imaging and have provided critical information about the structure of the vasculature and how it responds to a localized increase in neural activity (Grinvald, Lieke et al. 1986, Malonek and Grinvald 1996, Hillman 2007). In OIS, no extrinsic contrast agent is required. The most widely-used implementations of OIS imaging are sensitive to changes in light absorption by oxy- and deoxyhemoglobin, allowing them to capture rich information about multiple parameters related to neurovascular coupling. Other OIS techniques measure blood flow, as in laser speckle imaging (Briers, Duncan et al. 2013), or changes in scattering related to neural activity (Pan, Lee et al. 2018).

Optical imaging can also resolve vascular structure. High resolution measurements of microscopic structural changes with micrometer resolution across several millimeters of cortex can be obtained using ultra-large field two-photon microscopy (Tsai, Mateo et al. 2015, Lu, Liang et al. 2020). This technique can, for example, capture arteriole dynamics that are coherent in related areas across hemispheres (Mateo, Knutsen et al. 2017).

B. Fiber photometry

One of the main drawbacks of optical imaging is that it cannot reach deep into the brain (Helmchen and Denk 2005). The advancement of optical imaging has increased the penetration depth of high resolution imaging up to 1-1.2 mm below the surface of the brain (Horton, Wang et al. 2013, Liu, Li et al. 2019), but many subcortical structures remain inaccessible. Fiber calcium photometry monitors activity-relevant fluorescence via a fiber optic inserted in tissue where a calcium-sensitive sensor is expressed (Nakai, Ohkura et al. 2001, Adelsberger, Garaschuk et al. 2005, Tian, Hires et al. 2009, Dana, Sun et al. 2019) and is crucial to imaging neuronal activity in areas that are below 1 mm. For instance, the activity of the rat hippocampus, ~2.7 mm below the pial surface, can be measured by first expressing the calcium sensor GCaMP in hippocampal neurons and later measuring their calcium dynamics through fiber optic recordings (Chen, Sobczak et al. 2019). Fiber photometry has lower temporal resolution than electrophysiology and does not allow single cell imaging (Sych, Chernysheva et al. 2019) and is compatible with fMRI. Optical fibers are non-ferromagnetic, minimizing disruption to the MR image, and convey signals that are not affected by the rapidly-changing magnetic fields during image acquisition, conferring artefact-free measurements. Most of the signal captured by the fiber derives from ~10⁵-10⁶ μm³ around a 200 μm fiber, although this depends on fibre geometry (Pisanello, Pisano et al. 2019). Concurrent measurement of the BOLD signal and imaging of neuronal and astrocytic calcium transients was first achieved a decade ago using organic calcium sensors (Schulz, Sydekum et al. 2012) and permits interrogation of the contribution of specific neuronal and astrocytic population to the BOLD signal (He, Wang et al. 2018, Wang, He et al. 2018). The relatively simple fiber photometry technique facilitates combination with other methods along with fMRI, such as optogenetic stimulation (Albers, Wachsmuth et al. 2018).

C. Electrophysiology

While optical imaging has grown more and more powerful in recent years, electrophysiology is still the foundation on which our knowledge of the brain’s inner workings rests. Microelectrode recordings of single cell firing rates, multiunit activity (MUA), and local field potentials (LFPs) have provided many of the circuit characterizations that established the context for the interpretation of rs-fMRI. Simultaneous measurements of neural activity and fMRI have provided insight into the neural basis for the BOLD response (Logothetis, Pauls et al. 2001, Lu, Zuo et al. 2007, Magri, Schridde et al. 2012). While the number of channels available for electrophysiological recording is constantly increasing (M Carandini 2021), simultaneous rs-fMRI and electrophysiology remains limited to a relatively small number of sites, because every additional electrode increases the amount of potential interference between the modalities (Kleinfeld, Luan et al. 2019). The technical challenges of combining electrophysiology with MRI include image distortion near the electrodes and unwanted noise in the electrical recordings from currents induced by gradient switching during image acquisition. Electrophysiology can be more easily combined with amperometric measurements of oxygen concentration, which measure the concentration of free oxygen in the tissue. This approach sacrifices the whole brain coverage of rs-fMRI, but provides spatial and temporal resolution of oxygen signals comparable to that of the electrophysiological signals (Bentley, Li et al. 2016, Ledo, Lourenco et al. 2017). It allows for the examination of local coupling between neural activity and oxygen metabolism as well as the coordination of activity and oxygenation across a limited network of sites.

D. Functional ultrasound

The recent development of functional ultrasound (fUS) provides another tool for probing the neurophysiology of the BOLD fluctuations (Mace, Montaldo et al. 2011, Osmanski, Pezet et al. 2014, Dizeux, Gesnik et al. 2019). In many ways, fUS can serve as a surrogate for BOLD, as it is sensitive to the changes in cerebral blood volume and flow (CBV and CBF) that are involved in the BOLD signal. Compared to optical imaging techniques, fUS is able to image much deeper into the brain; however, skull thinning or removal is usually (but not always (Manwar, Kratkiewicz et al. 2020)) required, acquisition is often limited to a single coronal slice, and some portions of the brain remain inaccessible. fUS of the rodent brain has the advantage of higher spatial and temporal resolution than are typically obtained with rs-fMRI, on the order of 100 μm in plane and up to several ms per image (Osmanski, Pezet et al. 2014, Dizeux, Gesnik et al. 2019), which poses certain advantages for the investigation of cerebral hemodynamics, especially in the cortex (Urban, Mace et al. 2014). fUS is also appealing for studies in unanesthetized animals, as the requirements on animal constraint are less stringent than for MRI, image acquisition is far quieter than MRI, and the imaging system enables more naturalistic environments than the confined space of a preclinical MRI scanner. A recent demonstration of the power of fUS comes from measurements in awake, behaving mice that showed deactivation of the retrosplenial cortex during a sensory task (Ferrier, Tiran et al. 2020). The retrosplenial cortex has been posited as a key area in the default mode network (DMN) in rodents, but because fMRI studies of behavior are rarely performed in rodents due to the challenges involved, there was no direct evidence that it deactivated during task performance as the DMN does in humans (Fox, Snyder et al. 2005). fUS is thus perfectly positioned to bridge the gap between rs-fMRI in anesthetized animals and rs-fMRI in awake humans.

Together, these invasive techniques provide a detailed picture of neuronal, metabolic, and hemodynamic processes across a wide range of spatial and temporal scales that cannot be obtained in humans (Figure 2). The noninvasive nature of MRI makes it the ideal technique to serve as a bridge between these informative experiments in animals and their ultimate application to the interpretation of human neuroimaging studies. In the next sections, we review how animal studies help us to understand the link between brain activity and the BOLD response, and then describe insights into the coordination of activity across areas that are obtained from these experiments.

Figure 2.

The figure illustrates six of the most common imaging modalities to measure brain function in rodents. Typical temporal resolutions are shown on the top. The upper panels represent the tissue components that contribute to the recorded signal in each modality: the vasculature, in functional MRI and in ultrasound; all of the neurons surrounding the electrode, in electrophysiology; specific cell types (by using specific promoters), in fiber-based photometry measurements; and all of the tissue components, in microscopy techniques. The panels on the bottom represent the extent of the brain that can be recorded with each technique. While fMRI and ultrasound allow recording of signals from a large portion or even the whole brain, electrode or optical fiber-based measurements are usually restricted to a singular measuring point, and microscopy is limited to superficial portions of the brain. On the other hand, fMRI and ultrasound provide an integrated vascular-based signal from relatively large voxels, while microscopy can identify singular neurons in their particular environment. The spatial resolution of each technique depends on the particular method employed.

III. Local neural activity and the hemodynamic response

Understanding how local BOLD signal changes are linked to local neural activity is a necessary component of understanding how correlated BOLD fluctuations arise across the brain. Coordination across areas could in principle arise at any segment of the chain of events that constitutes neurovascular coupling, whether from the neural activity itself to a network-level change in metabolic demand to a widespread modulation of vascular tone. A number of reviews of recent work on the links between neural activity and the BOLD response are available (for example, (Drew, Mateo et al. 2020)). Here we provide a succinct summary of current thought on the neurophysiology of the BOLD signal to establish a framework for understanding the coordination of neural activity observed with rs-fMRI.

A. Excitatory neurons and neurovascular coupling.

Neurovascular coupling translates the signal from neurons to microvessels and involves astrocytes, pericytes, endothelial cells and smooth muscle cells. Multimodal experiments have been essential in establishing the coherence of the neurovascular partners during the spontaneous ultra-slow oscillations that are at the core of rs-fMRI (Niessing, Ebisch et al. 2005, Scholvinck, Maier et al. 2010, Magri, Schridde et al. 2012, Ma, Shaik et al. 2016, Mateo, Knutsen et al. 2017, Winder, Echagarruga et al. 2017, Echagarruga, Gheres et al. 2020). For example, simultaneous mesoscale calcium imaging and intrinsic signal imaging has shown that, on average, neuronal activity typically precedes hemodynamic responses in the resting state condition in the absence of stimulation (Ma, Shaik et al. 2016, Wright, Brier et al. 2017). In a recent technical tour de force, whole brain fMRI was combined with wide-field calcium imaging of the cortical mantle (Lake, Ge et al. 2020) to show that one third of the variance in the BOLD signal could be accounted for based on spontaneous excitatory activity. Similarly, simultaneous rs-fMRI and microelectrode recording found that the correlation between LFP power and the BOLD signal was between 0.26 and 0.44, depending on the anesthesia used (Zhang, Pan et al. 2019). The correlation coefficients from rs-fMRI are slightly smaller than the value of ~0.6 inferred from two-photon measurements of arteriole diameter concurrent with the LFP (Mateo, Knutsen et al. 2017). These reports provide strong evidence that spontaneous BOLD oscillations are tied to fluctuations of the brain’s intrinsic excitatory activity.

Furthermore, multimodal experiments provide insight into the types of activity that are reflected in the BOLD signal. The first simultaneous electrophysiological recordings and fMRI in monkeys showed that the BOLD signal is predicted by synaptic inputs and neuronal firing in the cortex in response to visual input. LFPs, believed to result primarily from synaptic inputs within the area, are actually a better predictor of the BOLD signal (Logothetis, Pauls et al. 2001) than the local spiking activity in response to visual stimulation. On the postsynaptic side, the gamma band of the LFP is overall the most efficient of bands to predict both the fMRI signal (Logothetis, Pauls et al. 2001, Niessing, Ebisch et al. 2005, Scholvinck, Maier et al. 2010, Magri, Schridde et al. 2012) and arteriole dilation and oxygenation in the cortex (Mateo, Knutsen et al. 2017, Winder, Echagarruga et al. 2017, Echagarruga, Gheres et al. 2020). The activation of a specific area via optogenetic stimulation of cortical excitatory neurons induces a local BOLD response as well as a BOLD response in contralateral and secondary areas, which defines a functional network (Lee, Durand et al. 2010, Desai, Kahn et al. 2011, Leong, Chan et al. 2016).

On the presynaptic side, locally generated spike rates are reflected in the hyperemic response (Lee, Durand et al. 2010, Scott and Murphy 2012, Kahn, Knoblich et al. 2013). These responses are not abolished by blocking glutamatergic ionotropic neurotransmission (Scott and Murphy 2012, Kahn, Knoblich et al. 2013, Iordanova, Vazquez et al. 2015), which suggests that local excitatory action potentials have both synaptic and non-synaptic excitatory dilatory effects on the vasculature and therefore both contribute to the BOLD fluctuations.

Potential mechanisms linking excitatory activity and hemodynamics.

The release of potassium by neurons during action potentials is a likely candidate to act in neurovascular coupling via generation of a hyperpolarizing pulse by voltage-gated and potassium concentration-sensitive potassium channels. Potassium injections close to a capillary are followed by a hyperemic response in the upstream arteriole, and 80% of this response is abolished by blocking inward-rectifier potassium channel (KIR2.1) in vivo (Longden, Dabertrand et al. 2017). In a parallel study, potassium-induced dilation was strongly affected by genetically ablating KIR2.1 only in endothelial cells in a knock-out mouse. This puts endothelial cells as sensors of neuronal activity in a position to control upstream blood flow and meet the oxygenation needs of the active area.

Cyclooxygenase-2 (COX-2), which is expressed by cortical excitatory neurons (Lecrux, Toussay et al. 2011) is another contributor to neurovascular coupling. In brain slices, there is pharmacological evidence that in response to neuronal activation, COX-2 permits prostaglandin E2 increases that target receptors potentially on smooth muscle (and pericyte) cells that participate in the dilation (Lacroix, Toussay et al. 2015). Using COX-2 inhibitors dramatically decreases the local hemodynamic response to optogenetic stimulation of excitatory neurons without abolishing it. Although optogenetic stimulation offers a way to activate a specific group of cells, the postsynaptic recruitment of inhibitory neurons is much faster (i.e., milliseconds) (Mateo, Avermann et al. 2011) than the hemodynamic responses (i.e., seconds), and as a result, in the absence of synaptic blockers, the hemodynamic responses to excitatory and inhibitory neurons are then mixed and part of the response may come from the postsynaptic activation of vasoactive inhibitory neurons.

B. Inhibitory neurons and neurovascular coupling.

Cortical inhibitory neurons are a diverse population with different expression patterns and morphologies (Gonchar, Wang et al. 2007, Petilla Interneuron Nomenclature, Ascoli et al. 2008, Perrenoud, Geoffroy et al. 2012, Kubota, Karube et al. 2016, Tremblay, Lee et al. 2016), yet all release the neurotransmitter GABA. In the cortex, only 20% of neurons are inhibitory neurons (Kim, Yang et al. 2017, Torres-Gomez, Blonde et al. 2020). They fire at higher rates than excitatory neurons and their activity is modulated across brain states and behavioral tasks (Gentet, Avermann et al. 2010, Niell and Stryker 2010, Gentet, Kremer et al. 2012, Liguz-Lecznar, Urban-Ciecko et al. 2016, Urban-Ciecko and Barth 2016), making them plausible contributors to the spontaneous BOLD fluctuations observed with rs-fMRI.

Studies in brain slices show that relatively slow arteriole dilations can be elicited by the excitation of a single interneuron that expresses vasoactive intestinal peptide (VIP) or nitric oxide synthase (NOS). Neuropeptide Y (NPY) expressing interneurons induced constriction and somatostatin (SOM) expressing interneurons could elicit both constrictions and dilations (Perrenoud, Rossier et al. 2012), indicating a possible role for inhibitory activity in neurovascular coupling corresponding to the effect of their vasoactive components.

The optogenetic stimulation of inhibitory neurons regardless of their subtypes induces large arteriole dilation similar to that of sensory responses, followed by an NPY-mediated undershoot (Uhlirova, Kilic et al. 2016) and CBF increase (Anenberg, Chan et al. 2015, Vazquez, Fukuda et al. 2018). After a blockade of synaptic transmission, the hemodynamic responses to optogenetic stimulation were barely affected, but when using a NO blocker, 75% of the response was abolished, leading to the hypothesis that the hemodynamic effects are mediated directly by NOS neurons (Vazquez, Fukuda et al. 2018, Krawchuk, Ruff et al. 2020). Inactivation of NOS neurons using designer receptors exclusively activated by designer drugs (DREADDs) shows a decoupling of neuronal activity with arteriole diameter (Echagarruga, Gheres et al. 2020) and local oxygenation as seen with OIS (Lee, Boorman et al. 2020). Interestingly, parvalbumin neuron (PV) stimulation induced arteriole constriction (Urban, Rancillac et al. 2012) and negative BOLD signal (Lee, Durand et al. 2010, Lee, Stile et al. 2020). All of the manipulations cited are very artificial and only establish causal links. Optogenetic stimulation of inhibitory neurons led to a biphasic BOLD signal locally at the stimulation site followed by negative downstream BOLD (Moon, Jiang et al. 2021). Of note, the HRF that arises from stimulation of inhibitory neurons appears to be highly variable on the stimulation frequency when compared to the activity of excitatory neurons (Moon, Jiang et al. 2021). In order to understand better the physiological role of inhibitory neurons during spontaneous activity, in the future we need to perturb and/or record the activity of subtypes of inhibitory neurons concurrently with vasodynamics and the release of vasoactive compounds using genetically encoded sensors (Oh, Lee et al. 2019, Pal and Tian 2020).

C. Astrocytes, pericytes, and neurovascular coupling

The study of the role that astrocytes play in neurovascular coupling is a shining example of how the refinement of imaging and stimulation techniques furthers the understanding of the role of cell types in neurovascular coupling (Cauli and Hamel 2018). At the ultra-slow frequencies (<0.3 Hz) detected with rs-fMRI, vascular fluctuations are coherent with neuronal fluctuations but not with astrocytic activity fluctuations (Gu, Chen et al. 2018). Despite the lack of coherence with vascular fluctuations, optogenetic stimulation of astrocytes triggers a positive BOLD signal without changing neuronal activity (Masamoto, Unekawa et al. 2015, Takata, Sugiura et al. 2018), so it is clear that they have the capacity to impact the vasculature. In more physiological conditions, astrocytic calcium in the cortex increases at the soma and in processes in response to sensory stimulations (Nizar, Uhlirova et al. 2013, Tran, Peringod et al. 2018). However, these increases are delayed relative to the onset of vasodilation, therefore calling into question the role of astrocytes in the physiological initiation of the dilation. In the olfactory bulb, on the other hand, calcium transients in astrocytic endfeet precede the vasodilation (Otsu, Couchman et al. 2015). Pathways that bridge the astrocytic activity and vasodynamics are under study and suggest a series of molecular candidates. As an example, astrocytes can act directly onto the pericytes (Mishra, Reynolds et al. 2016).

Brain pericytes are contractile cells (Fernandez-Klett, Offenhauser et al. 2010) that wrap around capillaries and precapillary arterioles (Hartmann, Underly et al. 2015). Optogenetic stimulation of pericytes can trigger vessel constriction and decrease the blood flow in some small vessels (Attwell, Mishra et al. 2016, Nelson, Sagare et al. 2020, Grubb, Lauritzen et al. 2021, Hartmann, Berthiaume et al. 2021). The control of capillary diameter by pericytes in physiological conditions is at the center of an important debate (Hill, Tong et al. 2015, Attwell, Mishra et al. 2016) as their position is ideal to generate small changes in capillary diameter that have a dramatic effect on blood flow (Blinder, Tsai et al. 2013). Pericytes exhibit spontaneous calcium transients that are sensitive to neuronal activity (Rungta, Chaigneau et al. 2018) and they are likely involved in maintaining basal vascular tone (Hartmann, Berthiaume et al. 2021).

While we are beginning to understand how excitatory neurons, inhibitory neurons, astrocytes and pericytes might separately contribute to the BOLD fluctuations, the complex interplay between these different types of cells that exists during normal brain function remains insufficiently understood, and is at the frontier of research into neurovascular coupling.

D. Neurovascular coupling outside of sensory cortical areas.

The majority of the work on neurovascular coupling has been performed in primary sensory cortex, for the simple reason that activity can easily be elicited with the proper sensory stimulus, and the timing of the activity can be precisely controlled. A few studies have examined neurovascular coupling in areas outside of the sensory cortex, with sometimes conflicting results. Particularly in rodent models of epilepsy, localized reductions in the BOLD signal have been observed despite increases in neural activity. David et al. observed deactivation in the striatum during spike wave discharges while activation was present in the somatosensory cortex and the thalamus (David, Guillemain et al. 2008). Similarly, Mishra et al. (Mishra, Ellens et al. 2011) found that cerebral blood volume (CBV), cerebral blood flow (CBF), LFP, MUA and the fMRI signal in somatosensory cortex and the thalamus all increased during spike wave discharges, while in the caudate and putamen, the fMRI signal and hemodynamic metrics decreased despite an increase in LFP and MUA (Mishra, Ellens et al. 2011). In rats with generalized seizures induced by bicuculline injection, BOLD deactivation was observed in the hippocampus while activation was present in the cortex. A thorough examination of neural activity, oxygen metabolism, and blood flow determined that while blood flow and oxygen metabolism increased in both the hippocampus and the cortex, neural activity in the hippocampus during the seizure increased 1000-fold, compared to a 100-fold increase in the cortex, and thus the balance between neural activity and hemodynamics resulted in a negative BOLD signal (Schridde, Khubchandani et al. 2008).

Another study using multielectrode arrays, OIS, and fMRI in the rat examined the neurovascular coupling in different areas of the ascending vibrissa pathway during vibrissa stimulation, and found that some areas exhibited a nonlinear or even inverse relationship between neural activity and hemodynamics (Devonshire, Papadakis et al. 2012). Even non-sensory cortical areas may have neurovascular coupling that differs from the typical model. In head-fixed but awake mice, it was found that locomotion induced neural activity in somatosensory cortex that was accompanied by an increase in CBV, but while neural activity also increased in frontal cortex, the CBV was unchanged (Huo, Smith et al. 2014). As an interesting potential mechanism, locomotion increases extracellular potassium (Rasmussen, Nicholas et al. 2019), which could potentiate hyperpolarization of the vascular endothelium in response to local neuronal activity (Longden, Dabertrand et al. 2017). These studies serve as a warning that blithe application of the same hemodynamic response function in all areas of the brain could provide a distorted view of underlying neural activity. Further investigation of neurovascular coupling in areas outside of the sensory cortex is needed to give guidance about areas and conditions where neurovascular coupling is likely to be altered.

E. Spatial specificity of the hemodynamic response

In addition to the underlying interplay of neurons, astrocytes, vascular components and diverse metabolites that results in the BOLD response, it is important to also consider the spatial specificity of fMRI functional mapping, which is intrinsically limited by the nature of the fMRI signal, i.e. the neurovascular coupling that transfers the neural signal to the cerebrovasculature (Kozberg and Hillman 2016). The perfusion domain consists of a penetrating arteriole that supplies oxygenated blood to a patch of cortex, a convoluted network of capillaries that distributes the oxygenated blood, and a penetrating venule that collects the deoxygenated blood and drains into the venous system (Blinder, Tsai et al. 2013). In rodents, the approximate distance between penetrating venules is ~200 μm (Nishimura, Schaffer et al. 2007), which can reach up to ~1 mm in humans (Cassot, Lauwers et al. 2006, Weber, Keller et al. 2008, Adams, Piserchia et al. 2015, He, Wang et al. 2018, Uludag and Blinder 2018). Numerous studies of the spatial specificity of fMRI have tried to identify the spreading of the fMRI contrast (e.g. the BOLD effect) across the vessel architecture from the active neuronal site. In an attempt to approximate the maximum distance at which one could measure oxygenation changes derived from neuronal activation, Turner et al. (Turner 2002) took into consideration the diameter of a pial vein and its distance to its most distal branches. It was estimated that a 100 mm² cortical area, which could theoretically be drained by a single pial vessel, could produce blood oxygenation changes up to a radius of 4.2 mm, setting an upper limit for the hemodynamic changes derived from activation of a cortical patch (Turner 2002).

Columnar specificity.

Experimental methods to quantify the spatial specificity of fMRI rely on comparing the vascular signals and the true neuronal responses. In humans, the estimation of true neuronal responses relies on knowledge of the underlying neuroanatomy, rather than direct measurements. For instance, a non-invasive approach consists of measuring ocular dominance columns with fMRI and comparing the size of these columns with reference columns in an ex vivo histochemistry-based atlas. In a human fMRI ocular dominance study, the full-width at half-maximum (FWHM) of the hemodynamic PSF was estimated to be 0.86 mm for spin-echo BOLD and 0.99 for gradient-echo BOLD (Chaimow, Yacoub et al. 2018), which suggests that the FWHM of the true hemodynamic PSF lies well below 0.8 mm, as fMRI measures are contaminated, to a certain degree, by motion-derived blurring, among others.

In animal models, it is possible to measure neural activity directly. One study combined high resolution fMRI (270 μm by 279 μm in plane) with a microelectrode array to measure neuronal activity in the somatosensory cortex of non-human primates (Shi, Wu et al. 2017). They showed comparable activation extents of columnar responses, demonstrating the capability of BOLD fMRI to accurately map neuronal activation during stimulation and at rest. Another study used a combination of optically-based CBV measurements, laser doppler flowmetry CBF measurements, and electrophysiological recordings in mice to assess optogenetically-evoked neuronal and hemodynamic responses in the somatosensory cortex (forelimb), to show that the PSF of the hemodynamic response spans approximately 175 μm (Vazquez, Fukuda et al. 2014). The strong resemblance between the general spatio-temporal features of columnar units investigated with fMRI and those measured with electrophysiology supports the use of fMRI to extend the investigation of cortical networks.

The temporal dynamics of the BOLD response can be leveraged to improve spatial specificity. By selectively mapping the initial dip of the hemodynamic response voxel-wise, iso-orientation columns in the cat primary visual cortex could be distinguished with fMRI (Kim, Duong et al. 2000, Fukuda, Moon et al. 2006). A similar approach that differentiates between early and late responses has been used to identify micro vs. macro-vasculature activation, which can greatly reduce the vascular bias observed in evoked laminar fMRI (Kay, Jamison et al. 2019, Kay, Jamison et al. 2020).

At extremely high spatial resolution, it becomes possible to distinguish individual penetrating vessels using high-field T2*-weighted methods for rodent fMRI (Yu, He et al. 2016, He, Wang et al. 2018, Chen, Sobczak et al. 2019). This allows vessel-specific BOLD or CBV signal to be identified from penetrating arterioles and venules (>20-70 μm in rodents) and pushes up against the limits of the theoretically-possible specificity of the BOLD signal (Figure 3). These studies raise a new aspect to be considered when interpreting high-resolution columnar results: the functional identification of single vessels with high-field fMRI mapping and the distinct connectivity patterns identified in venous and arterial networks suggest that vessel distribution is a critical factor in columnar BOLD fMRI.

Figure 3.

Single-vessel mapping with fMRI. A. The 3D contour of a rat brain shows the 2D slice perpendicular to the penetrating vessels, which covers one hemisphere for single-vessel fMRI mapping. B. The integrated multi-gradient-echo (MGE) image shows the arteriole-venule (A-V) map with arteriole voxels as bright dots and venule voxels as dark dots (yellow box: forepaw somatosensory cortex, FP-S1). C. The venule-specific BOLD fMRI and arteriole-specific CBV fMRI functional maps are overlaid on the A-V maps (venule ROI in blue, arteriole ROI in red). Adapted from Yu et al. (Yu, He et al. 2016).

Laminar specificity.

The specificity of the BOLD signal as compared to the size of a cortical column arrives at an estimate of the PSF across the cortical sheet, i.e., tangential plane. Also of interest is the specificity of BOLD in the orthogonal direction, i.e., radial plane, across the layers of the cortex. The cortical thickness can be divided into layers of different cyto- and vaso-architecture (Duvernoy, Delon et al. 1981, Palomero-Gallagher and Zilles 2019). The largest vessels, pial arteries and veins, run tangentially over the surface of the cortex, smaller vessels penetrate into the cortical depths radially, and a slightly but significantly anisotropic capillary mesh exists within the cortical ribbon, the density of which varies across the cortical depth (Weber, Keller et al. 2008, Ji, Ferreira et al. 2021). Instead of measuring across quasi-discrete neurovascular units as in studies of tangential specificity, estimates of radial specificity must disentangle layer-specific input from the complex vascular organization, which includes the draining effect of ascending venules that contaminate the upper layers with blood from deeper territories.

Given the complex neurovascular coupling features and different laminar-specific metabolic demands (Borowsky and Collins 1989, Weber, Keller et al. 2008, Devonshire, Papadakis et al. 2012, Blinder, Tsai et al. 2013, Shih, Chen et al. 2013, Urban, Mace et al. 2014), the interpretation of BOLD fMRI across cortical layers remains one of the most challenging and interesting topics of fMRI. Despite the complexity of the laminar environment in the cortex, multiple studies have proven the capabilities of high-resolution fMRI to reliably track activation changes across the cortical thickness in both bottom-up and top-down tasks (Silva and Koretsky 2002, Goense and Logothetis 2006, Yu, Qian et al. 2014, Poplawsky, Fukuda et al. 2015, Siero, Hendrikse et al. 2015, Huber, Handwerker et al. 2017, Kashyap, Ivanov et al. 2018, Finn, Huber et al. 2019, Sharoh, van Mourik et al. 2019), and combined laminar fMRI with EEG revealed frequency-specific modes of communication at different cortical depths, e.g., superficial layers positively correlated with gamma oscillations but deeper layers more involved with beta activity (Scheeringa, Koopmans et al. 2016), in agreement with laminar LFP studies (Buffalo, Fries et al. 2011). The growing body of literature on high-resolution fMRI with multi-modality studies to target a diverse series of structures in the micro-to-mesoscale levels demonstrates the potential of fMRI to sample brain function with great accuracy and supports the ambition to map laminar, columnar and vascular networks with fMRI.

IV. Coordination of intrinsic activity across the brain

Understanding how neurovascular coupling leads to hemodynamic fluctuations that follow fluctuations in intrinsic neural activity gives us a partial picture of the origin of the coordinated BOLD fluctuations observed with rs-fMRI. To understand how neural activity and hemodynamics are coordinated across areas, multisite, multimodal measurements are essential.

Conceptually, there are a number of ways that activity could be coordinated that are consistent with known neuroanatomical and neurophysiological processes (Figure 4). The simplest explanation is that areas that are directly connected, whether via long range white matter tracts or collateral projections, tend to activate together. Similarly, two areas that receive direct input from a third area, i.e., common input, would exhibit correlated activity with each other and with the third area. These connectivity-based interpretations are the basis for the vast majority of rs-fMRI analyses, yet they may provide an overly-simplistic view of the complex system that is the brain. From the perspective of nonlinear dynamics, the coordination of intrinsic brain activity could arise from interactions between local populations of neurons connected via a vast structural network to produce patterns of “network” activity that are not immediately obvious based on structural connections alone. For example, it is difficult to explain the anticorrelation observed between certain networks based on direct connections alone, but it arises naturally in models based on the full structural network of the brain (Cabral, Hugues et al. 2011).

Figure 4.

Networks of correlated BOLD fluctuations could arise from multiple mechanisms. In the first panel, two bilateral networks (one red, one blue) are driven by separate, direct white matter connections. This scenario is the simplest and functional connectivity is often interpreted to reflect direct connections as a first approximation. In the second panel, correlations in the red network arise via a direct white matter connection. Activity propagates over time along the cortex to the blue areas, which appear to be correlated despite the lack of a direct connection. In the third panel, the red areas and blue areas are part of a larger network. Correlations within each subnetwork and anticorrelation between the subnetworks arises through the interactions of activity over the entire structural network. The fourth panel shows another scenario, where red areas and blue areas are both driven by input from a single area, representing neuromodulation. Differences in receptor type and density may account for different responses to the neuromodulatory input in red and blue networks. It is likely that all of these mechanisms contribute to functional connectivity.

While there is still much to learn about the functional architecture of the brain, a few principles seem likely to govern the brain’s structure and resulting patterns of activity. The minimization of metabolic costs implies certain wiring rules (van den Heuvel and Sporns 2011, Bullmore and Sporns 2012, Betzel, Avena-Koenigsberger et al. 2016, Liang, Hsu et al. 2018) and operational levels of activity (Tagliazucchi, Balenzuela et al. 2012, Haimovici, Tagliazucchi et al. 2013). The vasculature has co-evolved with the metabolic needs of the brain (Ji, Ferreira et al. 2021), and as a result, the densely interconnected core areas of the brain are also highly perfused (Liang, Hsu et al. 2018), and the vasculature may be regulated in such a way as to support functional networks (He, Wang et al. 2018, Bright, Whittaker et al. 2020).

Given the various factors that contribute to coordinated brain activity, how can we decide which ones are dominant? Time-averaged measures like correlation provide a single snapshot of how activity is coordinated over relatively long periods of time, but more recently-developed analysis methods that provide time-varying estimates of activity and/or coordination can give deeper insight into processes that are likely to contribute. Different analysis methods are sensitive to changes in activity that occur at different time scales. For example, point process analysis and coactivation patterns can capture changes that persist for only a single TR (Liu and Duyn 2013, Petridou, Gaudes et al. 2013). Sliding window correlation and brain states obtained from clustering sliding window time courses are sensitive to slower changes in activity determined by the length of the window that is used (Allen, Damaraju et al. 2014, Shakil, Lee et al. 2016). Some methods identify repeated patterns, e.g., quasiperiodic patterns (Majeed, Magnuson et al. 2011, Yousefi, Shin et al. 2018) or common transitions of activity across brain states or between areas (Chen, Langley et al. 2016, Vidaurre, Smith et al. 2017).

In the next section, we review mechanisms for different processes that might guide coordinated activity and evidence that they contribute to the brain’s functional organization, using both time-averaged and time-varying analysis techniques along with multimodal experiments.

A. Functional connectivity and direct white matter connections.

Similarity of structural and functional connectivity.

Correlated activity in two physically separated brain areas is often interpreted as arising from direct white matter connections between those areas. White matter tracts not only connect different cortical regions but also sustain broad subcortical projections from deep brain functional nuclei to the cerebral cortex (Wycoco, Shroff et al. 2013). Many areas that exhibit correlated BOLD fluctuations do have direct connections, and the matrix of structural connectivity within the brain as measured with diffusion-based tractography or viral tracing techniques has substantial similarity to the matrix of functional connectivity; the correlation between structural and functional connectivity is ~0.2-0.4 for an individual in the HCP dataset (Messe 2020). Unimodal areas and particularly visual cortex exhibit tighter coupling between structural and functional connectivity than transmodal areas (Vazquez-Rodriguez, Suarez et al. 2019). Nevertheless, some areas that exhibit strong correlation in rs-fMRI are connected only via indirect pathways, complicating this interpretation of functional connectivity. One contributing factor is that typical MR-based methods of mapping structural connectivity have certain biases and do not capture some types of connections well, which could account for some of the differences between structural and functional connectivity (Reveley, Seth et al. 2015, Schilling, Gao et al. 2018, Yeh, Jones et al. 2020). For example, diffusion based tract-tracing methods are more sensitive to terminations in the gyral crown than in sulci (Schilling, Gao et al. 2018), and long range connections are underestimated (Reveley, Seth et al. 2015).

In mice, injection of anterograde traces in multiple areas of the cortex has allowed the production of a highly detailed structural connectivity atlas (Allen Mouse Brain Connectivity Atlas (2011) (Oh, Harris et al. 2014)). However, even in animal models where viral tracers can be employed instead of MR-based methods, structural connections explain a limited number of the functional connections that are observed. In the rat, a meta-analysis of tracing data combined with rs-fMRI found a group level Spearman’s correlation of 0.48 between the structural and functional connectivity matrices (Diaz-Parra, Osborn et al. 2017). In marmosets, connectivity measured with retrograde tracer injections was correlated with functional connectivity between areas ranging from −0.1 to 0.6 after controlling for the effects of the distance between two brain regions (Hori, Schaeffer et al. 2020). While viral tracers have their own biases, the moderate similarity observed between functional and structural connectivity for any tract-tracing method suggests that direct white matter connections provide only a partial explanation for the coordination of activity across the brain (Kura, Xie et al. 2018). Nevertheless, recent studies that use dense whole-brain tracing combined with rs-fMRI are providing new insights into the potential mechanisms of network organization. The extensive characterization of the mouse structural connectome in combination with mouse rs-fMRI has recently offered a plausible explanation for the way brain networks work and interact with each other (for instance, see (Coletta, Pagani et al. 2020)).

Relationship between BOLD and neural activity in areas connected directly via white matter pathways.

The interhemispheric connections that travel through the corpus callosum connect homologous areas in the left and right hemispheres, which generally exhibit strong, and largely symmetric, functional connectivity as well. These areas have been the target of many experiments involving simultaneous microelectrode recording and rs-fMRI or wide-field optical imaging of neural activity, which clearly show that neural activity in both cerebral hemispheres is correlated with spatial specificity similar to that seen in the hemodynamic response (Lu, Zuo et al. 2007, Mohajerani, McVea et al. 2010, Pan, Thompson et al. 2011, Ma, Shaik et al. 2016). For instance, Ma et al. demonstrated the use of wide-field optical imaging to measure calcium-based neuronal activity (GcaMP fluorescence) concurrently with hemodynamic signals from much of the cerebral cortex in mice (Ma, Shaik et al. 2016), which allowed a reliable examination of the neurovascular coupling in bilateral areas during wakefulness and under anesthesia.

Most electrophysiological recordings are focused on relatively high frequencies (1 Hz and above). In contrast, the BOLD fluctuations fall mostly below 0.1 Hz (Cordes, Haughton et al. 2001). The disparity in frequencies complicates comparison across modalities and is usually addressed by filtering and downsampling the electrophysiological data. However, it is also possible to measure very low frequency electrical signals that are comparable in frequency to and coherent with the BOLD fluctuations (Pan, Thompson et al. 2013). The correlation between these infraslow electrical oscillations and the BOLD signal is at least as strong as the correlation between BOLD and bandlimited power in higher frequency ranges. This correlation between modalities is localized to the area near the electrode and to the comparable area in the other hemisphere, suggesting that infraslow electrical activity could facilitate network communication. In humans, comparable findings have been obtained using EEG (Hiltunen, Kantola et al. 2014, Grooms, Thompson et al. 2017).

Little is known about infraslow electrical activity. One possibility is that the envelope of higher frequency activity is rectified by nonlinear membrane processes and then contributes to the LFP (Ahrens, Levine et al. 2002, Haufler and Pare 2019). Some studies suggest that these slow potentials reflect modulation of cortical excitability (Rockstroh, Muller et al. 1993, Rosler, Heil et al. 1997), which is consistent with findings using electrocorticography (eCoG) in humans that the low frequency modulation of both gamma power and firing rate are correlated across hemispheres (Nir, Mukamel et al. 2008). He et al. found that the slowest eCoG potentials (0.1 - 4 Hz) had a similar correlation structure to the BOLD signal across hemispheres, but that the relationship with gamma band power was state-dependent (He, Snyder et al. 2008). Further support for a neural origin comes from Chan et al., who found that infraslow EEG power was attenuated by the application of drugs that decrease cortical excitability and glutaminergic transmission (Chan, Mohajerani et al. 2015).

In the more traditional EEG frequency bands above 1 Hz, the precise frequencies of neural activity that best predict BOLD correlation depend on the anesthetic condition, but are not necessarily the frequencies most linked to the local BOLD response. For example, under isoflurane, BOLD correlation between left and right somatosensory cortex is best predicted by the correlation of delta and theta band-limited power of local field potentials from the two areas, while the local BOLD response was linked to broadband power, with a particularly high correlation to power in the beta and gamma bands (Pan, Thompson et al. 2011). Lu et al. observed a similar relationship between functional connectivity and power coherence in the delta band in rats under alpha-chloralose anesthesia (Lu, Zuo et al. 2007). In contrast, the highest correlations with the local BOLD signal were observed in high frequency power (gamma band or above) during visual stimulation (Logothetis, Pauls et al. 2001) or during spontaneous activity (Shmuel and Leopold 2008).

There are non-neuronal sources of infraslow electrical activity that confound ECoG measurements (Drew, Mateo et al. 2020). These include astrocytes (Kuga, Sasaki et al. 2011), an intriguing possibility given their links to the vasculature, and changes in the blood-brain barrier potential (Voipio, Tallgren et al. 2003). In the same frequency range, oxygen polarography measurements in four sites from two different networks (visual and DMN) in monkeys showed that correlation peaked at 0.06 Hz and was stronger for sites within the same network than for sites in different networks, providing independent support for a possible infraslow coordination of activity across the brain (Li, Bentley et al. 2015).

Insights from time-varying analysis of directly-connected areas.

Dynamic analysis provides further insight into the relationship between neural activity and BOLD correlation in directly connected areas. In left and right primary somatosensory cortex of rats anesthetized with isoflurane, sliding window correlation of bandlimited LFP power is correlated with sliding window correlation of the BOLD signals from the two areas (Thompson, Merritt et al. 2013), particularly in theta, beta and gamma bands. Compared to Pan et al. (Pan, Thompson et al. 2011), where time-averaged correlation in the theta and delta bands was most predictive of coordinated neural activity, this suggests that analyses that capture the time-variance of the signal might increase sensitivity to high frequency activity. In humans, sliding window correlation shows there are periods when functional connectivity closely matches anatomical connectivity, and others when functional connectivity is dominated by inter-network interactions (Liegeois, Ziegler et al. 2016).

In wide-field optical imaging, the pattern of bilateral fluctuations changes much more quickly than could be captured by BOLD fluctuations, evolving through different bilateral activity patterns over the course of a few hundred milliseconds (Mohajerani, McVea et al. 2010). As in humans, the patterns of spontaneous activity often resemble those evoked by a task or stimulation (Smith, Fox et al. 2009, Mohajerani, Chan et al. 2013). Chan et al. showed that infraslow cortical activity recapitulates similar motifs to those observed at higher frequencies, a vivid demonstration of how activity can nest across temporal scales (Chan, Mohajerani et al. 2015). Parallel work in humans shows that EEG-resolved micro-states that persist for less than one second resemble the spatial patterns of fMRI-defined resting state networks (Custo, Van De Ville et al. 2017, Rajkumar 2021). Similar resting state networks can be detected in OIS measurements of hemodynamics, further support for tight coupling between coordinated neural activity and correlated BOLD signals (Kura, Xie et al. 2018).

Effects of white matter disruption on functional connectivity

One way to test the effects of direct connections on functional connectivity is to manipulate the network structure or function by physical or chemical lesions. For example, chemogenetic inhibition of the anterior cingulate cortex (ACC), a key brain hub, reduces functional connectivity across much of the brain (Peeters, Hinz et al. 2020). Another study in rats found that after callosotomy, both functional connectivity and coherent gamma band power between left and right somatosensory cortex decreased (Magnuson, Thompson et al. 2014). Similarly, in acallosal mice, the bilateral symmetry of activity mapped with voltage sensitive dye was greatly reduced (Mohajerani, McVea et al. 2010). Moreover, the coherence of ultra-slow oscillations of arteriole diameters (0.1 Hz) in related areas across hemispheres depends on the integrity of the corpus callosum (Mateo, Knutsen et al. 2017). These studies support a role for direct connections via the corpus callosum in mediating functional connectivity, but interpretation remains somewhat ambiguous since acallosal mice necessarily adapt for the lack of the corpus callosum during development, and the acute callosotomy procedure used in the rat study is highly invasive and traumatic. Side effects from the surgery may explain why the left and right caudate, which are not directly connected via the corpus callosum, also exhibited reduced functional connectivity after callosotomy. On the other hand, the reduced functional connectivity in the caudate could also be interpreted as a downstream effect of the disruption of coordinated input from the cortex, which would normally drive correlated activity in the caudate.

B. Functional connectivity as a consequence of the structural network

One can expand the question of whether functional connectivity arises from direct white matter connections to include the effects of the entire structural network, including feedback loops. In this case, coordinated BOLD fluctuations could arise from the complex dynamics of activity as local neural populations interact with the rest of the brain via the structural network. This type of network interaction can explain some of the correlation that is observed between regions that are not directly connected themselves, and it can account for anticorrelation between certain areas or networks, which is difficult to understand in terms of direct connections alone.

The effects of network structure are typically examined using brain network models, which combine a structural matrix obtained from diffusion-weighted MRI or viral tracing with neural mass models that describe the activity at each brain node. The activity at each node over time is determined by its prior activity and the input from the nodes to which it is connected. A number of different neural mass models have been used, e.g., for a review, see (Cabral, Kringelbach et al. 2017), all of which require optimal fitting of a number of parameters to find those that give the best fit to measured functional connectivity. Most brain network models recapitulate some aspects of time-averaged functional connectivity but fewer aspects of time-varying functional connectivity (Kashyap and Keilholz 2019). At this stage of their development, brain network models show that particular network interactions between neural populations can give rise to features that are observed in rs-fMRI data, but the similar performance of a wide variety of models and parameterizations means that they are less successful at eliminating particular conceptualizations of how the brain works.

C. Role of intracortical connectivity and propagation

In addition to relatively long range connections via white matter tracts, the cortex is densely interconnected through collateral fibers. These fibers provide another route for coordination of activity across the brain and can account for some of the network structure observed as functional connectivity (Atasoy, Donnelly et al. 2016).

Anatomical structure of cortico-cortical connections.

Analysis of structural connectivity in the mouse has demonstrated that axons connecting two cortical areas of the same hemisphere start in the upper and deeper layers of the cortex, run along layer 6, and terminate either in the middle layers of primary cortices or in superficial layers of higher order cortical areas (Watakabe and Hirokawa 2018). Of the six layers that make up the neocortex (Baillarger 1840, Palomero-Gallagher and Zilles 2019), layer 5 accounts for most intra- and inter-hemispheric intracortical (and corticofugal) outputs, and in turn receives inputs from layers 2/3, 4 and 6. Layer 4 is primarily associated with feed-forward projections, e.g., thalamic afferents, while layers 5 and 6 have been observed to play a role in both feed-forward and feed-back processes (Douglas and Martin 1991, Sherman and Guillery 2011, Tong, Hocke et al. 2013, Harris, Mihalas et al. 2019). In contrast to the primarily vertical orientation of fibers in other layers, the vast majority of fibers in layer 1 are horizontal or oblique (Palomero-Gallagher and Zilles 2019), which suggests a role in cortico-cortical connectivity and information processing.

Analyses of structural connectivity atlases in several species suggest that the degree of intra-hemispheric connectivity primarily depends on the distance between targets, while other features, such as the cyto-architectonics, i.e., the dominating cell types, play an important role in both ipsi- and contra-lateral connections (Ercsey-Ravasz, Markov et al. 2013, Beul, Barbas et al. 2017, Goulas, Uylings et al. 2017, Schmidt, Bakker et al. 2018). An example can be found in mice, where the granularity of the cortex largely conditions the structural connectivity between distant cortical areas (Goulas, Uylings et al. 2017). Interestingly, cholinergic projections from the basal forebrain target cortical laminae in a fiber age-dependent manner (Allaway, Munoz et al. 2020). Anatomical studies, accomplished at least partially ex vivo at the microscopic level, provide the hard-wiring model for potential cortical connections; however, exploring the true functional connectivity requires in vivo techniques such as electrophysiology, to reliably detect neuronal activity within the cortex, or rs-fMRI, to assess the coordination, e.g., temporal correlation, between areas.

Propagation of evoked activity through cortico-cortical connections.

Despite their invasiveness, electrophysiology studies can provide reliable detection of neuronal activity with both laminar- and columnar-specificity, identifying the spreading of information flowing through the cortex. For instance, recording from electrodes placed along the cortical depth in mouse M1 during optogenetic stimulation demonstrated cortical depth-dependent activation upon stimulation of different areas, e.g. the orbital cortex activated neurons in layer 6, secondary motor cortex activated layer 5b neurons, and thalamic and primary somatosensory cortex (S1) stimulation targeted mainly upper layers (Hooks, Mao et al. 2013), demonstrating the differential role of neurons across the cortical depth to maintain connections across distant areas. Cortical columns add an important dimension to the study of functional networks with high specificity. Although cortical columns have been regarded as the smallest unit in the cerebral cortex (Horton and Adams 2005), inter-columnar interactions are critical to shape receptive fields, hence, the autonomy of a single functional column is questionable, as interactions between neighbors are needed for proper operation (Lund, Angelucci et al. 2003). For example, excitatory neurons of the cortex are usually inhibited by interneurons within the same cortical lamina but in a different column. Lateral inhibition from neighboring intra or inter-laminar neurons (Katzel, Zemelman et al. 2011) provides a mechanism to shape or restrict neuronal receptive fields.

The sophisticated organization of the cerebral cortex in laminae and columns suggests that cortical networks spreading across multiple cortical areas result from a complex interplay at a mesoscale (Mitra 2014, Roe 2019). Although this dimension is better explored with multi-site electrodes, broad networks expanded across multiple brain areas require imaging techniques that can cover big fields of view; hence, fMRI has been largely used to characterize inter-areal interactions and novel high-resolution fMRI methods are starting to bridge the gap between high-resolution electrophysiology and large-coverage fMRI mapping. For instance, using a 7T scanner, Huber et al. observed that, in human M1, voluntary movement results in supragranular as well as infragranular activation, while somato-sensation leads to activation of the superficial layers of the cortex (Huber, Handwerker et al. 2017). At 9,4T in rodents, S.G. Kim observed either glomerular, plexiform or granular activation in response to either odor, electrical stimulation of the lateral olfactory bulb or electrical stimulation of the anterior commissure, respectively (Iordanova, Vazquez et al. 2015). Super-high spatio-temporal resolution fMRI of the barrel cortex employing a line-scanning method at 11.7T demonstrated that somato-sensory stimulation to the whisker or forepaw activates layer 4/5 first, peaking later at layer 2/3 and layer 5 (Yu, Qian et al. 2014).

Propagation of spontaneous activity through cortico-cortical connections.

In contrast to the evoked studies, resting-state evaluations with cortical depth-specificity have remained scarce, one of the reasons being the difficulty of imaging the whole brain at high resolution. An elegant study in the auditory cortex of rats showed that the propagation pattern of spontaneous activity differs from that observed during sensory-evoked activation, with spontaneous activity following a bottom-up pattern, spreading slowly across columns, and sensory activity initiating in granular layers and reaching neighbor columns at a higher speed (Sakata and Harris 2009). It is possible that some sub-systems use the same underlying circuit for communication during resting-state and upon stimulation or task performance, e.g., in the visual cortex the columnar behavior during rest resembles the iso-orientation maps obtained during visual stimulation (Vasireddi, Vazquez et al. 2016).

Although the number of resting-state studies with cortical depth-specificity is, at present, insufficient, some reports have shed light on the potential laminar-specific routes supporting resting-state communication. Network specific analysis with high-resolution fMRI in mice suggest that the default mode network (DMN) is especially sustained by supra-granular layers (Whitesell, Liska et al. 2021). Whitesell et al. identified voxels associated with the DMN in resting-state scans, and investigated the structural cortico-cortical connectivity affiliated to those voxels. Fibers starting in layer 2/3 mostly targeted other DMN cortical areas, while fibers starting in layer 5 projected to regions external to the DMN. Similarly, resting-state functional connectivity in area 3b and 1 of the monkey, somato-sensory cortex, appeared stronger between superficial and intermediate layers of the cortex (Mishra, Majumdar et al. 2019). In humans, the superficial layers of M1 correlate, at rest, with S1, and both superficial as well as deep layers communicate with the premotor cortex (Huber, Handwerker et al. 2017). Interestingly, activation of the superficial layers but not the deeper layers of mouse motor cortex triggered network-wide events (Weiler, Wood et al. 2008). The apparent predominance of the supragranular layers to mediate resting-state connectivity may support the use of microscopy-based techniques to investigate brain networks in rodents.

Despite the multiple efforts to characterize resting state networks within the cortex, to date, no consensus has been reached regarding the cortical depths involved in the presumed connections between hubs. Future studies performed at ultra-high resolution without compromising brain coverage will hopefully shed light on the circuit-specific resting-state fMRI correlation features across cortical layers.

Intracortical connections are one potential source for time-lagged relationships and propagation between cortical areas that are observed with some types of rs-fMRI analysis (Majeed, Magnuson et al. 2011, Mitra, Snyder et al. 2015, Yousefi and Keilholz 2021). However, the differences in timing between areas, which are on the order of one second or more, are substantially longer than the time that it takes for synaptic transmission. Another possibility is that the BOLD signal propagation is driven by waves of glial activity. In the hippocampus, there is some evidence that astrocytic calcium waves propagate from cell to cell and are accompanied by changes in infraslow LFPs and blood flow (Kuga, Sasaki et al. 2011).

D. Sources of common input and neuromodulation

If areas A and B both receive strong input from area C, A and B can exhibit correlated activity even if there is no direct connection between them. Thus networks of spatially distinct nodes that exhibit correlated activity may contain areas where activity is correlated by direct connections, areas whose correlated activity reflects a common driver, and downstream areas where parallel pathways of input result in correlation despite lack of a direct connection or direct common input. The ability to distinguish between these mechanisms is important for correct interpretation of altered functional connectivity, e.g., in psychiatric disorders.

One challenge in distinguishing possible sources for functional connectivity within a network is that some studies focus only on the cortex, ignoring diencephalon structures that are known to have widespread projections. Animal fMRI studies are less prone to this type of bias, but reports of subcortical involvement in cortical networks are rare, possibly because the subcortical nuclei that project to specific cortical areas are small, comparable to the size of a voxel. Even in stimulus-based studies where subcortical activity is known to occur, it is often undetectable (Keilholz, Silva et al. 2004, Keilholz, Silva et al. 2006). It has been, however, possible to detect important subcortical contributions to global cortical responses by using whole brain fMRI in rats; an example being the thalamic involvement in widespread cortical suppression coinciding with spontaneous astrocytic activity (Wang, He et al. 2018). Another study showed that optogenetic stimulation of the thalamus increased functional connectivity in multiple cortical networks (Wang, Leong et al. 2019). Low frequency stimulation of the hippocampus increases interhemispheric functional connectivity, and a chemical blockade of hippocampal activity decreases interhemispheric connectivity (Chan, Leong et al. 2017), more evidence of the role that these diencephalon structures play in the coordination of activity across the brain.

Role of neuromodulation in functional connectivity.

The size and position of deep brain nuclei makes it difficult to examine their functional connectivity directly. Instead, researchers often pursue a perturbationsl approach. In animal studies, pharmacological, chemogenetic and optogenetic approaches can be used to manipulate signaling from deep brain nuclei. Chemogenetics allows the reversible manipulation of cells upon systemic administration of a drug by targeted expression of specific receptors in their membranes. The most common example of chemogenetics are DREADDS or designer receptors exclusively activated by designer drugs (Campbell and Marchant 2018). In the case of optogenetics, the putative receptors, also selectively expressed in particular cell populations, are sensitive to light pulses delivered by optical fibers, the most common being Ch R2 (Boyden, Zhang et al. 2005). Both techniques, together with common pharmacological approaches, constitute an essential tool to interfere with the nervous system and study network vulnerability or modulation. A substantial body of literature has shown that manipulation of catecholamine systems results in altered functional connectivity, both in humans and in animals (Shah, Blockx et al. 2015, Shah, Blockx et al. 2016).

Over the years, a number of neuromodulatory nuclei and neurotransmitter systems have been proposed as candidates to drive common input via neuromodulation of cortical networks, including but not limited to the rostral ventrolateral medulla (Golanov and Reis 1996, Drew, Duyn et al. 2008), the basal forebrain (Nair, Klaassen et al. 2018, Turchi, Chang et al. 2018) and cholinergic system, the raphe nucleus (Razoux, Baltes et al. 2013, Shah, Blockx et al. 2016) and serotonin, ventral tegmental area and dopamine (Decot, Namboodiri et al. 2017), and locus coeruleus and norepinephrine (Zerbi, Floriou-Servou et al. 2019). Neuromodulatory systems are often thought of as having widespread and diffuse projections, which would make them an unlikely source of common input for the spatially structured functional networks that are observed with rs-fMRI. However, there is a growing appreciation for the diversity of spatial and temporal structure that these nuclei can provide (van den Brink, Pfeffer et al. 2019, Coletta, Pagani et al. 2020). Differences in the receptor type and/or receptor density across brain areas ensures that the same input can have regionally distinct effects (Lindvall, Bjorklund et al. 1978, Atzori, Cuevas-Olguin et al. 2016, Kim, Jung et al. 2016, Disney and Higley 2020, Sarter and Lustig 2020), even at the laminar level (Palomero-Gallagher and Zilles 2019). Moreover, there is growing evidence that neuromodulatory nuclei themselves are heterogenous, with different populations of cells exhibiting different patterns of input or types of activity (Li, Yu et al. 2018, Totah, Neves et al. 2018, Zaborszky, Gombkoto et al. 2018, Disney and Higley 2020, Sarter and Lustig 2020).

One way in which neuromodulatory nuclei might act is to alter the response to incoming stimuli. For example, by reducing intrinsic fluctuations of activity, the signal-to-noise ratio of activity related to sensory input could be improved. Lottem et al. (Lottem, Lorincz et al. 2016) have shown that serotonin can act in this way, quenching spontaneous fluctuations while sparing sensory-evoked activity. This is consistent with work by Grandjean et al. (Grandjean, Corcoba et al. 2019), showing that optogenetic stimulation of the dorsal raphe nuclei reduced blood volume and neural activity across wide areas of the cortex. Other neurotransmitters act in a similar manner; activation of the cholinergic system also reduces spontaneous activity while leaving evoked activity unaffected (Meir, Katz et al. 2018). In contrast, a recent study showed that chemogenetic activation of the locus coeruleus increased functional connectivity throughout much of the cortex (Zerbi, Floriou-Servou et al. 2019).

It is probably overly simplistic to think of spatial correlations as arising from a single neuromodulatory system. Deep brain nuclei are interconnected and their interactions may be complex. For example, dopamine has a well-characterized role in cortico-striatal and mesolimbic theta oscillations, but norepinephrine levels also play a behaviorally-relevant role in coordination across these areas (Dzirasa, Phillips et al. 2010). In another case, norepinephrine drives activity in midline thalamic areas through a dopaminergic mechanism (Beas, Wright et al. 2018). In humans, a rs-fMRI study found that dopaminergic areas in the midbrain and the serotonergic dorsal raphe nuclei were part of the default mode network, while the locus coeruleus and the remaining serotonergic nuclei were integrated with the executive control network (Bar, de la Cruz et al. 2016). This suggests that neuromodulatory nuclei could together account for the general functional organization of the brain into two anticorrelated networks.

Neuromodulatory nuclei are known to alter cerebral perfusion. Stimulation of the basal forebrain and locus coeruleus both induce vasodilation, which may be mediated partly by interneurons and depend on the state of the brain (Kocharyan, Fernandes et al. 2008, Lecrux and Hamel 2016), offering another method for localized differences in the response to neuromodulatory input. Because components of the neurovascular unit have receptors for neurotransmitters (e.g., norepinephrine), a local increase in neuromodulator could potentially modify neurovascular coupling and affect the fidelity of the hemodynamic response to a given neuronal activation (Lecrux, Sandoe et al. 2017). Thus another contribution to functional connectivity could arise from neurotransmitter-mediated changes in the “signal-to-noise” of the rs-fMRI signal, in terms of its sensitivity to neural activity. While the amplitude of two signals does not affect the correlation between them, an increase in the noise level results in lower correlation, which would appear as a reduction of functional connectivity.

Arousal levels and functional connectivity.

Dynamic analysis might provide more insight into how neuromodulation contributes to functional connectivity. For example, arousal levels are primarily driven by the ascending reticular activating system (ARAS), which arises in the brainstem and then innervates the cortex directly or via the thalamus. It is usually considered to include neurotransmitter-specific pathways from the locus coeruleus, raphe nuclei, and ventral tegmental area, among others. Arousal levels are known to vary over the course of the typical rs-fMRI scan, even in subjects attempting to remain awake. Several studies that look at brain states over the course of a rs-fMRI scan have identified particular states that appear to be associated with drowsiness, as they become more frequent as the scan progresses and occur in conjunction with other features of reduced alertness (Chang, Liu et al. 2013, Chang, Metzger et al. 2013, Allen, Damaraju et al. 2014, Tagliazucchi and Laufs 2014, Chang, Leopold et al. 2016, Falahpour, Chang et al. 2018). The strong contribution of arousal to dynamic functional connectivity ensures that it also plays a role in time-averaged functional connectivity, and suggests that fluctuations in neuromodulatory input related to other neurophysiological processes can influence the spatial structure of the BOLD fluctuations. In fact, the strength of connections between hubs in the human DMN is strongly correlated with the integrity of the arousal system of the brain, e.g. patients suffering from severe disorders of consciousness exhibit weaker connections (Vanhaudenhuyse, Noirhomme et al. 2010). Similarly, a rat model of brainstem coma showed increased functional connectivity as animals regained arousal (Pais-Roldan, Edlow et al. 2019).

Templates of brain activity related to alertness level have been identified for both humans and monkeys (Chang, Leopold et al. 2016, Falahpour, Chang et al. 2018). To date, no such template exists for rodents, partly because the vast majority of rodent studies are conducted under anesthesia. There is evidence that the repertoire of brain states is reduced by anesthesia in the rodent (Hudetz, Liu et al. 2015). However, substantial variability is observed in brain states based on sliding window correlation or coactivation patterns in anesthetized rodents (Keilholz, Magnuson et al. 2013, Zhang, Pan et al. 2020), an indication that fluctuations in arousal level may occur, at least to an extent, under anesthesia. Interestingly, arousal fluctuations tracked as varying pupil dynamics in anesthetized rats were accompanied by correlated rs-fMRI maps (Pais-Roldan, Takahashi et al. 2020). The fact that the brain state fluctuates at the within-subject short-term level indicates that network patterns of rs-fMRI are necessarily complex and subjected to external modulation.

Neuromodulation and repeated spatiotemporal patterns of activity.

Neuromodulatory input might also account for some of the repeated spatiotemporal patterns of activity that are observed in rs-fMRI (Majeed, Magnuson et al. 2009, Majeed, Magnuson et al. 2011, Thompson, Pan et al. 2014, Belloy, Naeyaert et al. 2018, Yousefi, Shin et al. 2018, Abbas, Belloy et al. 2019, Belloy, Billings et al. 2021) (quasiperiodic patterns or QPPs). These patterns, which are remarkably consistent across individuals, involve propagation along the cortex and account for a large portion of the functional connectivity that is observed in humans (Yousefi and Keilholz 2021). Similar patterns have been observed in both neural activity and hemodynamics in mice using optical imaging, and as in humans, the phase of the propagating wave appears to dictate the coactivation patterns of cortical areas (Matsui, Murakami et al. 2016). The repetitive nature of the patterns and the stereotyped timing of activation and deactivation implies that they are a fundamental feature of the brain’s intrinsic organization, rather than a response to external stimuli or internal processing, and direct white matter connections alone are not sufficient to account for the fairly complex patterns of signal propagation that are observed (Kashyap and Keilholz 2019). In rodents, the patterns include deactivation in the reticular formation accompanied by widespread cortical activation (Belloy, Billings et al. 2021). This finding is the inverse of prior work showing cortical deactivation when neuromodulatory nuclei were stimulated. In humans, areas in the brainstem and thalamus lead the changes in cortical activation, which suggests the mechanisms that drive QPPs are similar across species (Yousefi and Keilholz 2021).

In support of a neuromodulatory source for QPPs, QPPs were found to contribute less to functional connectivity in patients with attention deficit hyperactivity disorder (ADHD) than healthy controls, consistent with a reduction in neuromodulatory input (Abbas, Bassil et al. 2019). In rats, the patterns appear to reflect infraslow electrical activity (Pan, Thompson et al. 2013, Thompson, Pan et al. 2014), and the timing varies from a cycle of ~5-6 s in animals sedated with dexmedetomidine to 10 + seconds in animals under isoflurane anesthesia (Thompson, Pan et al. 2014), again consistent with the differing levels of arousal. Moreover, alterations in these patterns were observed in a mouse model of Alzheimer’s, as expected given the early degeneration of neuromodulatory nuclei in this disorder (Belloy, Shah et al. 2018).

It is worth noting that both widespread patterns of neural activity (e.g., QPPs) and the global signal (as described in a later section) may be linked to neuromodulation. In some ways, this is not surprising, since high global signal amplitudes necessarily reflect widespread patterns of strong, coherent activity. However, the question remains of whether the global average signal contains additional information about arousal and neuromodulation that is independent of the contribution from large-scale spatial patterns.

Given the variety of ways that neuromodulatory input from deep brain nuclei can potentially coordinate neural activity, more experiments that capture the BOLD signal along with the local changes in neuromodulator concentration using sensor imaging are badly needed (Sabatini and Tian 2020). Techniques that can capture the effects of neuromodulators throughout the whole brain, rather than in single areas, will be particularly valuable in unraveling the role that these deep brain nuclei play in organizing large scale brain activity.

Other sources of common input.

Other potential sources of common input include autonomic processes like respiration and cardiac pulsation. Both can introduce spatially distributed, correlated noise that overlies the fluctuations related to neural activity, and while in principle this noise can be removed using a number of approaches, in practice this is difficult because both processes also are maintained by active networks in the brain. The primary respiratory centers, for example, are in the medulla and pons, but respiration is also under voluntary control regulated by the cortex, and it is well known that both heart rate and respiration patterns change with the subject’s overall state (for example, breathing and heart rate slow as a person falls asleep). As one example of the interconnectedness of arousal and autonomic functions, heart rate variability covaries with functional connectivity in networks associated with arousal (Chang, Metzger et al. 2013). Moreover, global BOLD fluctuations vary with vascular tone in the extremities, which has suggested that sympathetic innervation of the vasculature in the brain may impact rs-fMRI. The photoplethysmography (PPG) signal from the finger typically drops following EEG K complexes, along with the global fMRI signal (Ozbay, Chang et al. 2018), suggesting that intermittent sympathetic activation changes the rs-fMRI signal.

Many studies in humans have drawn attention to the effect of respiration and cardiac cycles on resting state networks (Birn, Diamond et al. 2006, Chang, Cunningham et al. 2009, Chen, Lewis et al. 2020). In typical animal experiments, the difference in field strength, relative geometry of the coil and chest cavity, and differences in respiratory and cardiac rates could cause the “noise” component from these cycles to be quite different than in humans. In particular, the faster physiological rates in rodents, not accompanied by proportionally faster fMRI acquisition rates, usually lead to signal aliasing (Pais-Roldan, Biswal et al. 2018).

Often forgotten as potential sources of coordinated brain activity are other rhythmic interoceptive inputs. For example, the gut and brain interact continuously, and the digestive system produces rhythmic pulses that could influence spontaneous activity in the brain. Cao et al. (Cao, Lu et al. 2019) showed that stimulation of the stomach resulted in widespread activation in sensory and cingulate cortices. The general tendency to treat the brain as an isolated system may need to be abandoned if we hope to fully understand how spontaneous activity is structured (Stringer, Pachitariu et al. 2019).

The vasculature as source of a common signal.

Another easily-overlooked source of a common signal across widespread brain areas is the vasculature. The hemodynamic response to spontaneous activity is much the same as the response for evoked activity, but other oscillations that are effectively removed by averaging during stimulus-based studies could contribute to widespread coherent fluctuations in the BOLD signal. Vasomotor oscillations are present in arterioles even when they are isolated from neural input (Osol and Halpern 1988) and slow arteriole dilations are conserved in the absence of neuronal activity (Winder, Echagarruga et al. 2017). Spontaneous neuronal activity precedes and can entrain low-frequency vascular oscillations (Mateo, Knutsen et al. 2017, Gu, Chen et al. 2018). Because vasomotion is affected by changes in blood pressure (Koenigsberger, Sauser et al. 2006), systemic blood pressure changes could modulate the apparent connectivity within networks by increasing or reducing the signal to noise ratio of the BOLD fluctuations. Indeed, the low frequency BOLD oscillations increase in amplitude when blood pressure decreases (Kannurpatti, Biswal et al. 2008). The amplitude and frequency of vasomotor oscillations depend on vascular features such as vessel diameter and wall thickness (Koenigsberger, Sauser et al. 2006) and could contribute to parcellations in the frequency of vasomotion observed in fMRI experiments (Mitra, Ogawa et al. 1997).

One of the appealing aspects of vasomotion as the temporal signature of functional connectivity is the relatively symmetrical nature of the vasculature across the midline. If the vasomotor oscillations are also symmetric, that might account for the very strong interhemispheric BOLD correlations that remain when the primary spatiotemporal patterns are removed by regression, and which also appear in areas that are not strongly connected across hemispheres (Yousefi and Keilholz 2021). An intriguing study by Bauer et al. found that optogenetic stimulation of excitatory neurons from different cortical areas resulted in patterns of activity that were less bilaterally symmetric than the hemodynamic-based functional connectivity from the same areas (Baek, Shim et al. 2016). Another study found that the brain of the zebra finch, which lacks a corpus callosum, exhibited strong bilateral functional connectivity (Kundu, Santin et al. 2014). It is possible that the vasomotor oscillations amplify oscillations of neural activity, or, since interstitial oxygen and ionic concentrations might be actively modulated by the vascular oscillations, they might even influence local neuronal activity via slow potential modulation.

E. Global signal

The global signal referred to in fMRI is defined as an average time course across all voxels of the cortex or all of the brain. Although a certain correlation with this average time course can be observed globally, the signal has been shown to reflect the activity of certain networks (Fox, Zhang et al. 2009). The global signal is often removed during fMRI preprocessing because it also contains contributions from physiological noise; however, this almost certainly removes some neuronal signal as well. While a large portion of the global signal comes from non-neuronal contributors, e.g., thermal noise, cardiac or respiratory function, motion, hardware artifacts, etc. (Murphy, Birn et al. 2013) (Power, Laumann et al. 2017, Power, Plitt et al. 2017, Drew, Mateo et al. 2020), several studies have also demonstrated a neurobiological source. For instance, in monkeys, the global fMRI signal correlates with electrophysiological recordings (Scholvinck, Maier et al. 2010), in mice, the phase of the global signal determines hemodynamic and calcium-based neuronal coactivation in the neocortex (Matsui, Murakami et al. 2016) and brain-state fluctuations assessed with external arousal indicators (e.g., eye blinks or pupil size) have demonstrated a strong correlation with the global functional dynamics in monkeys and rats (Chang, Leopold et al. 2016, Pais-Roldan, Takahashi et al. 2020). Interestingly, the global signal appears to be modulated by the activity of the basal forebrain, as its amplitude decreases upon inactivation of the nucleus basilis of meynert specifically within the altered hemisphere (Musch and Honey 2018, Turchi, Chang et al. 2018). Recent studies suggest a relationship between the global signal and the sympathetic tone, as well as with fluctuations in subcortical arousal, as observed during sleep in animals and humans (Fukunaga, Horovitz et al. 2006, Liu, Yanagawa et al. 2015, Ozbay, Chang et al. 2019). The dependence of the global signal on the autonomic system is supported by the fact that autonomically-triggered vasodilation leads to higher fMRI signal (Wise, Ide et al. 2004, de la Cruz, Schumann et al. 2017) and that skin vascular tone, also correlated with the global signal, is regulated by the sympathetic nervous system (van Houdt, Ossenblok et al. 2010, Tong, Hocke et al. 2013, Ozbay, Chang et al. 2018) (Shokri-Kojori, Tomasi et al. 2018).

Unlike in human studies, where the global signal contains major contributions from motion and physiological noise, the global signal in typical experiments on rodents exhibits less contamination from these processes. The majority of rodents are anesthetized and restrained in a stereotaxic head holder for rs-fMRI, which greatly reduces motion. In some studies, rodents are also given a paralytic and mechanically ventilated, which further reduces motion and ensures that contributions from respiratory variation are minimal. Despite these stringent controls, global signal excursions that involve both cortical and subcortical areas can be identified (Figure 5). Further studies on the neural basis of global signal modulations should leverage the well-controlled conditions in the rodent to maximize sensitivity to the underlying processes.

Figure 5. Significant global signal modulation not linked to noise.

In a typical rodent fMRI study, head motion and breathing variability is minimized by the use of anesthesia and artificial ventilation in paralyzed animals. For example, under a mixture of dexmedetomidine and low-dose isoflurane, synchronized patterns can be identified visually in both neocortex and subcortical regions (top), but are not reflected in the neighboring scalp tissue, which exhibits relatively small amplitude changes over time despite having comparable baseline signal intensity. The global signals are unlikely to be created by head motion or physiological variation given the stringent restraints and paralytic agents employed. Thus rodent models may offer a unique platform for the investigation of global neural modulation, relatively free from contamination from physiological noise. The global modulation can also be removed by global signal regression (GSR), shown in the bottom panel.

Besides noise and neuronal activity, the CSF dynamics may also contribute to the global signal, as suggested by the coupling observed between the flow of CSF into the brain (4^th ventricle) and slow-wave sleep (Fultz, Bonmassar et al. 2019). Interestingly, despite a reduction of cerebral blood flow and metabolism of oxygen and glucose, the global BOLD fMRI signal increases during non-REM sleep (McAvoy, Tagliazucchi et al. 2019). Given the strong evidence supporting a certain neurobiological significance of the global signal, its removal in resting-state studies must be justified and taken into consideration during interpretation of resting-state results (Fox, Zhang et al. 2009).

The variety of mechanisms that might mediate the organization of activity across the brain should make it clear that it is unlikely that any single one of them fully accounts for the complex spatial structure and temporal dynamics that are observed. Instead, we believe that the concerted effects of multiple processes contribute to functional connectivity, and that different processes might become the dominant contributors under different conditions (see Figure 6 for one possible model). In this scenario, differences that arise in patient populations may represent a change in the dominant source of coordination, rather than a change in the underlying network structure.

Figure 6.

An illustrative model of how different neurophysiological processes might act in concert to produce patterns of coordinated activity. Direct white matter connections, for example along the midline between the anterior cingulate and the retrosplenial cortex and between homologous areas in the left and right hemisphere, form a framework for ‘fast-pass’ interactions and synchronized oscillations. Less direct connections, for example intracortical connections within each hemisphere, could modify the phase or amplitude of activity in other areas. Common projections from the basal forebrain might drive waves of propagating activity from anterior to posterior areas (Massimini, Huber et al. 2004); input from the locus coeruleus and thalamus (Neske 2015) might modulate the overall level of excitation and slow cortical oscillations. The resulting patterns of activity are filtered through the overlying vascular structure, which may enhance the interhemispheric symmetry.

V. Future directions

The need for technology development

The field of human fMRI has advanced substantially over the last decade, with developments in hardware and software supporting image acquisition at higher spatial and temporal resolution (e.g., following Human Connectome Project protocols (Van Essen, Smith et al. 2013)). These have led to standardized imaging protocols that are widely adopted and which facilitate public data sharing and reanalysis.

In animal rs-fMRI, less progress has been made. Acquisition methods and analysis tools that are routinely used in humans are still not widely available for rs-fMRI studies in animals. The combination of the high field strengths and small size of the rodent brain makes it difficult to design coils suitable for high acceleration factors, and without the potential of clinical adoption, it is difficult to fund the needed development of hardware and software. Moreover, only the most recently-upgraded animal scanners have multichannel capabilities. However, most preclinical MRI systems operate at higher field strengths (7T +) than human scanners, have stronger gradients, and impose less stringent limitations on SAR and gradient switching. There is a need to leverage these advantages along with new technology from human scanners to create an approach that is specifically tailored to neuroimaging of the rodent brain.

One exciting recent development is the use of zero-TE sequences like multiband SWIFT for functional imaging in rodents (Lehto, Idiyatullin et al. 2017, Paasonen, Laakso et al. 2020). With the high magnetic fields of preclinical MRI systems and the close proximity of the rodent brain to susceptibility gradients, particularly in the ear canals, distortion and signal loss are major challenges. Zero-TE sequences minimize these effects, and even more, are robust to susceptibility artifacts caused by implanted electrodes or optical fibers for multi-modal experiments. The challenge is that the source of functional contrast for these sequences remains poorly understood, making translation to humans less straightforward. Still, the animal models provide the ideal opportunity for identifying the source of functional contrast, which may reflect the same underlying processes as the BOLD signal.

Zero-TE sequences could prove to be the key to unlocking another challenge of rs-fMRI in rodents: the use of anesthesia. The vast majority of rodent studies are still performed under anesthesia, with isoflurane and/or dexmedetomidine popular choices. Anesthesia is considered essential for minimizing motion and stress in untrained animals, but also results in a potential confound for translating results to rs-fMRI in awake humans. Many researchers have attempted to move rs-fMRI experiments to awake animals, some with relative success (Peeters, Tindemans et al. 2001, King, Garelick et al. 2005, Brydges, Whalley et al. 2013, Reed, Pira et al. 2013, Chang, Procissi et al. 2016, Stenroos, Paasonen et al. 2018, Han, Chen et al. 2019) but the approach remains challenging and not widely used. For awake studies, animals are typically accustomed to restraint and the noise of the scanner over the course of several days. Low noise sequences like multi-band (MB) SWIFT might reduce the stress imposed by the scanner environment and improve the quality of studies in awake animals. Moreover, MB-SWIFT is relatively insensitive to the body motion of the animal in the scanner (Paasonen, Laakso et al. 2020), unlike EPI-based sequences where even respiratory motion can introduce substantial artifacts (Raj, Paley et al. 2000, Kalthoff, Seehafer et al. 2011). MB-SWIFT could therefore allow quiet scans while relaxing the strenuous constraints on animal movement, a major step forward for awake animal studies.

Meanwhile, the use of anesthesia to minimize stress and motion is likely to continue to predominate in rodent fMRI studies (Becq, Barbier et al. 2020). Recent investigations have been exploring a compromise approach that minimizes the dosage of anesthetic and the resulting vascular effects. For example, a mixture of dexmedetomidine and a low dose of isoflurane appears capable of creating a stable brain state of light anesthesia that can extend for hours (Brynildsen, Hsu et al. 2017), providing flexibility in the experimental approach to acquire more scans from each animal under conditions that are comparable across designs and different subjects.

In terms of analysis, preprocessing pipelines and noise removal are not as well-developed for rodents as for humans (Chuang, Lee et al. 2019). Motion, cardiac and respiratory signals of animals and humans interfere with the fMRI time courses, especially in the brain edges, brainstem and areas near the main blood vessels of the brain (Teichert, Grinband et al. 2010, Birn, Cornejo et al. 2014, Griffanti, Douaud et al. 2017, Keilholz, Pan et al. 2017, Sclocco, Beissner et al. 2018, Yoshikawa, Masaoka et al. 2020). Although the physiological noise is similarly distributed in animals and humans, its distinct temporal features make the identification of physiological noise more complex in rodent fMRI; the respiratory and cardiac cycle in rodents happens at rates approximately five times faster than in humans (Williams, Magnuson et al. 2010, Fleming, Thompson et al. 2011, Pais-Roldan, Biswal et al. 2018), which translates into a poorer sampling of these signals with fMRI and consequently more difficult removal from the data. One approach to denoise animal fMRI data is to modify human preprocessing protocols, e.g., training a FIX classifier (Salimi-Khorshidi, Douaud et al. 2014) for mice (Zerbi, Grandjean et al. 2015). As in human studies, multi-echo sequences have been used to determine which components of the MRI signal do not exhibit BOLD-like T2* dependence and are therefore likely to be noise (Kundu, Santin et al. 2014), although this comes at the expense of slower sampling rates and more complex analysis.

The rodent neuroimaging community has begun to work towards the goal of improved standardization and data sharing by examining rs-fMRI data from mice acquired across multiple institutions (Mandino, Cerri et al. 2019). The time is ripe to create a hub for rodent neuroimaging data, which would encourage exploratory analysis, allow comparisons between different protocols from different labs, make reproducibility analyses feasible, and facilitate collaborations (Mandino, Cerri et al. 2019). Such a database would be highly complementary to the Allen brain atlas, which provides a wealth of information about gene expression and tracer injections that provide valuable context for functional neuroimaging studies. Databases for transgenic animals and disease models could be particularly valuable to the community.

Linking whole brain activation to neural circuits

Neuroscience as a field has traditionally had a heavy focus on circuit-level descriptions of brain activity, many of which have been studied in rodent models using invasive techniques that can identify individual cells or even individual synapses. Functional MRI studies in rodents have begun to bridge the gap between these careful circuit characterizations and the systems level activity of the brain.

Whole brain rs-fMRI is typically used to detect interconnected hubs. Although a series of (mostly cortical) networks have been identified in the human and animal brain, the specific groups of neurons that are recruited for each of those networks (e.g., granule cells, pyramidal cells, etc.) remains to be deciphered. To specifically define the circuit underlying functional connectivity between two remote areas, a required step would be to detect the cortical layers in both territories that exhibit a high temporal correlation. This additional dimension, i.e. the cortical depth in a layer-specific resting-state connectivity analysis, could bring into play new features to characterize the brain function.

Beyond the mesoscale, besides specific cortical layers and cell types, the synapse type between cortical areas may modulate, at some extent, the behavior of brain networks (Covic and Sherman 2011). Emerging studies combining fMRI with calcium imaging, optogenetics and novel stimulation methods like focal infrared neural stimulation may enable the community to dissect, with unprecedented detail, cortical networks as well as the underlying vascular networks in animal models (Chernov and Roe 2014, Albers, Wachsmuth et al. 2018, He, Wang et al. 2018, Wang, He et al. 2018, Chen, Pais-Roldan et al. 2019, Xu, Qian et al. 2019).

Linking brains and behavior

Most broadly, neuroscience aims to link patterns of brain activity to behavior. Functional connectivity has much promise in this regard, with recent studies showing, for example, associations between network structure and changes in attentional states (Kucyi, Daitch et al. 2020, Rosenberg, Scheinost et al. 2020), working memory (Yamashita, Yoshihara et al. 2018, Avery, Yoo et al. 2020), and many other behavioral and psychological phenomena (Woodward and Cascio 2015, Li, Hu et al. 2019, Yan, Chen et al. 2019). A key challenge is to distinguish causal from non-causal relationships (Pearl 2009). Here animal models will be crucial. The interventional experiments necessary to show that specific patterns of connectivity drive particular behaviors cannot be conducted within humans. Behavioral interventions -- inducing a change in behavior and then observing a corresponding change in connectivity -- cannot establish causality. Physiological interventions are necessary, but those available in humans are not specific enough nor powerful enough for the task at hand. While we cannot yet define exactly what the required experiments will look like, it is clear that they will have to be conducted in animal models.

Signatures of specific types of activity or neurophysiological processes.

Given that the BOLD signal is an integrative representation of a wide variety of neurophysiological processes, possibly with additional contributions from the vasculature, and that it is inevitably contaminated with a variety of noise, it is daunting to contemplate the challenge of interpreting functional connectivity, especially when it comes to alterations observed in patients with psychiatric or neurological disorders. Attempting to localize a single cause for a change in functional connectivity might be analogous to trying to listen to a single flute player in a poorly-recorded orchestra. Nevertheless, there are clues that might help to determine which factor dominates in a particular situation.

It is important to use all of the tools at our disposal. To obtain the fullest picture of how the BOLD fluctuations are organized, we must move beyond time-averaged functional connectivity and consider temporal and spectral information as well. As an example, spatial patterns that recur periodically over time are unlikely to reflect transient sensory stimuli or cognitive processes, while short periods of high activity in sensory and motor areas could be related to body motion. The power spectrum of the BOLD signals may help to isolate periodic contributors, whether well-known like respiration and cardiac pulsation, or more obscure, like gastric oscillations. The spatial specificity of changes may imply a source: downstream effects of a change in neuromodulatory input would be expected to be widespread, while effects of a change in connectivity between two areas of interest should be more localized. Spatial specificity might be especially useful in high resolution studies that can resolve laminar activity, since much is known about the distribution of input and output connections in particular layers of some cortical areas.

Similarly, the explosion of machine learning has popularized efforts to learn something about the underlying neural activity from the BOLD fluctuations, but the exercise remains speculative in nature unless there is a way to validate the output. For example, using a recurrent neural network (RNN), one can learn the coefficients of the Wilson-Cowan model to parse firing rates or excitatory and inhibitory currents from the BOLD data (Kashyap and Keilholz 2020). By forcing the RNN to output interpretable parameters, the way is paved for the multimodal experiments that are essential for the ultimate validation.

In all cases, findings from rs-fMRI need to be placed in the context of more specific information obtained with other modalities. With the help of these tools, the interpretation of functional connectivity remains challenging but perhaps not hopeless.

Testing models of the brain

Animal models have always provided a platform where models of brain activity could be tested using targeted but invasive methods such as physical or chemical lesioning, electrical stimulation, or pharmacological manipulations. New advances in selective manipulations using optogenetic stimulation or DREADDs provide an unprecedented degree of selectivity and control, opening new possibilities for understanding how different types of neural or glial cells contribute to the hemodynamic response, or how particular areas are involved in the coordination of activity across the brain (Ryali, Shih et al. 2016). At the same time, new imaging and recording developments, many of which were supported by the BRAIN initiative, have made it possible to study the coordination of neural activity across the brain in ways that were not possible before. These studies, in combination with rs-fMRI, allow us to build new and better models of how the brain works.

The ultimate test of any model is how well it predicts what will happen when some component of the model is perturbed. For example, one could build a brain network model based on a particular neural mass model at each node and parameterized to fit rs-fMRI data, then predict what happens when the activity at a specific node is increased. To test the prediction, one might acquire rs-fMRI data during optogenetic stimulation of the same node, and compare the experimental results to those predicted by the model. The process sounds simpler than it is, given the technical constraints of optogenetic stimulation and the challenges of reproducing the effects of stimulation with a particular neural mass model, but the framework is valid and offers a powerful way to advance our understanding of how the brain works by leveraging interventions available in animal models.

VI. Conclusions

Coordination of activity across distant areas of the brain as measured with functional connectivity has proven to be necessary for normal function, e.g., disruptions in connectivity are often associated with neurological or psychiatric disorders. However, functional connectivity has not been adopted as a clinical tool for routine diagnosis and evaluation, possibly because the source and role of these coupled oscillations remains poorly characterized. A better understanding of the specific processes that contribute to BOLD fluctuations and functional connectivity could move rs-fMRI beyond a biomarker to an informative tool for the diagnosis or evaluation of brain disorders. While data from human subjects provides insight into the implications of functional connectivity for cognitive processes, animal studies are crucial to understand the underlying neurophysiology. At the same time, these studies will improve understanding of the systems-level organization of the brain, bridging between more localized studies of neural circuits and the macroscopic functional architecture of the brain. The identification of particular neurophysiological processes that are reflected in the BOLD fluctuations could improve sensitivity to changes related to cognition and behavior, and provide a more mechanistic explanation of the processes involved.