Interpreting BOLD: a dialogue between cognitive and cellular neuroscience

The computational properties of the human brain arise from an intricate interplay between billions of neurons connected in complex networks. However, our ability to study these networks in healthy human brain is limited by the necessity to use noninvasive technologies. This is in contrast to animal models where a rich, detailed view on the cellular level brain function has become available due to recent advances in microscopic optical imaging and genetics. Thus, a central challenge facing neuroscience today is leveraging these mechanistic insights from animal studies to accurately draw physiological inferences from human noninvasive signals. We will discuss a strategy of addressing this challenge by combining human and animal experiments and computational modeling with the endpoint goal to deliver a quantitative probe for neuronal activity of known cell types in human brain enabling a paradigm shift in human fMRI studies: from a simple mapping of fMRI signal change to the explicit estimation of the relative activity levels of specific neuronal cell types.

Noninvasive neural imaging of the brain has become a primary neuroscience tool. Different techniques have been used but functional magnetic resonance imaging (fMRI) is currently the dominant procedure by which hemodynamic processes are monitored to infer properties of neural activity. A fundamental question concerns the functional relationship between the local hemodynamic changes that are measured and the implications for neuronal function. In a series of studies, we have measured an important metabolic parameter, tissue oxygen concentration, along with activated neural activity in the central visual pathway. Tissue oxygen changes are directly related to the blood oxygen level-dependent (BOLD) signal that is used in fMRI. We have employed a sensor that provides simultaneous co-localized measurements of oxygen concentration and neural activity from single cells to characteristics of the oxygen response including spatial, temporal and scaling parameters. We have also investigated changes in oxygen concentration with single cell, multiple unit, and local field potential activity. Finally, we have used specialized sensors to measure neurometabolic coupling between neural activity, glucose, and lactate in activated visual cortex. Findings of these studies will be described.

Initially regarded as ‘noise’, spontaneous (intrinsic) activity accounts for a large portion of the brain’s metabolic cost. Moreover, it is now widely known that infra-slow (< 0.1 Hz) spontaneous activity, measured using resting state functional magnetic resonance imaging (rs-fMRI) of the blood oxygen level dependent (BOLD) signal, is spatially correlated within resting state networks (RSNs). However, despite these advances, the temporal organisation of spontaneous BOLD fluctuations among RNSs has remained elusive. By studying temporal lags in the resting state BOLD signal, we have recently shown spontaneous BOLD fluctuations consist of remarkably reproducible patterns of whole-brain propagation. Embedded in these propagation patterns are ‘motifs’ which, in turn, give rise to RSNs. Additionally, propagation patterns are markedly altered as a function of state, whether physiological or pathological. Thus, a deeper understanding of the temporal organisation of the BOLD signal may yield insights into the roles spontaneous activity plays in brain function.

The blood oxygenation level-dependent (BOLD) contrast is widely used in functional magnetic resonance imaging (fMRI) studies aimed at investigating neuronal activity. However, the BOLD signal reflects changes in blood volume and oxygenation rather than neuronal activity per se. Therefore, understanding the transformation of microscopic vascular behavior into macroscopic BOLD signals is at the foundation of physiologically informed noninvasive neuroimaging. We present a new method that uses oxygen-sensitive two-photon microscopy to measure the BOLD-relevant microvascular physiology occurring within a typical rodent fMRI voxel and that predicts the BOLD signal from first principles using those measurements. The predictive power of the approach is illustrated by quantifying variations in the BOLD signal induced by the morphological folding of the human cortex. This framework is then used to quantify the contribution of individual vascular compartments and other factors to the BOLD signal for different magnet strengths and pulse sequences. Finally, this method is used to validate and optimize the Calibrated fMRI approach used to recover the cerebral metabolic rate of oxygen CMRO2 in human studies.

Understanding the cellular mechanisms of neurovascular coupling should lead to improved understanding of the spatiotemporal properties of the BOLD response, the dependence of the response on neural activity, and the sensitivities of neurovascular coupling to different disease states. Here, we describe the previously overlooked importance of the vascular endothelium as route for the conduction of vasodilation during functional hyperemia. Dependence of the response on an endothelial pathway introduces new interpretations of earlier pharmacological results, explains mismatches in timing between astrocyte responses and vasodilation at the level of diving arterioles, introduces the possibility of multiple mechanisms with slow and fast components contributing to the non-linearities of the BOLD response, and introduces new explanations for the effects of drugs and disease states on brain function and cognitive decline. Through high-speed imaging of both neural activity and hemodynamics across the cortices of the awake, behaving mouse brain, we are exploring the broad consequences of endothelial involvement in neurovascular coupling, in stimulus-evoked and resting state conditions, during longitudinal disease progression, under different pharmacological conditions, and during early post-natal development. These results are important for both understanding how neurovascular coupling can influence brain health, and for improving interpretation of the BOLD signal in health and disease.

In the last two and a half decades, magnetic resonance methods aimed at imaging neuronal activity have transformed our ability to study the human brain, going from early experiments demonstrating relatively course functional images in the visual cortex to functional mapping with laminar differentiation of neuronal ensembles that perform elementary computations. This development relied on an extensive set of experiments that examined the mechanisms underpinning the functional imaging signals, tackling questions about the spatial specificity of the neurovascular coupling and the connection between hemodynamic and metabolic consequences of neuronal activity and the perturbations on MR detected signals. These studies, conducted on animal models as well as in humans, have provided a rigorous, albeit as of yet incomplete, understanding of the mechanisms underlying the functional mapping signals that reflect neuronal activity, leading to the use of ever increasing magnetic fields to gain accuracy and resolution and to new imaging and image reconstruction methods to map functional activity at the level of cortical columns and layers, and connectivity with ever increasing spatial and temporal resolution.

Functional MRI has many significant disadvantages as a source of information about nervous systems. It does not directly represent neuronal activity; it has coarse spatial and temporal resolution compared to the range of scales of space and time that brains subtend; it is not measured in SI units; experimental recordings are at least 80% noise; etc. Nonetheless, the patterns of between-regional correlation in slowly oscillating fMRI time series have turned out to be robustly replicable and not trivially explained. Graph theoretical models of human fMRI networks, derived from association matrices of pair-wise functional connectivity estimated for all possible pairs of ~300 regional nodes, demonstrate complex topology: small-worldness, hubs, modules, core/periphery, etc. These features are replicable and heritable. The topological and spatial or geometrical organization of fMRI networks is consistent with the theory that their formation is largely determined by the trade-off between a few competitive factors or conservation laws. Hypothetically, an economic trade-off between the biological cost and the topological value of network components could drive the formation of fMRI networks. To test the generality of this and other hypotheses generated by connectomic analysis of ‘resting state’ fMRI data, graph theoretical methods can be used to make comparable measurements in many other neuroscientific datasets. Meta-analysis of large scale libraries (N~1000 primary papers) of fMRI activation studies demonstrated that more expensive topological features (hubs, rich club) were associated with domain-general, ‘higher-order’ cognitive functions; and that high cost / high value network hubs were hotspots for structural brain deficits associated with many different brain disorders (including Alzheimer’s disease and schizophrenia). Many of the complex topological characteristics of large-scale human fMRI networks are qualitatively reproduced at the microscopic scale of functional networks derived from multi-electrode array recordings of growing neuronal cultures in vitro. The economical model of a trade-off between biological cost and topological value has been specifically re-affirmed by analysis of viral tract tracing data (~400 anterograde tracer injection experiments) on the anatomical connectivity of the mouse brain. We conclude that despite the well-known limitations of fMRI, it has emerged as almost uniquely capable of measuring the complex network organisation of human brain function in a way that is physically, neurobiologically, cognitively, and clinically meaningful.