Temporal dynamics in fMRI resting-state activity

In a significant new study, Mitra et al. (1) demonstrate the existence of reproducible temporal patterns of spontaneous activity from human functional magnetic resonance imaging (fMRI) recordings. This finding and the novel methods used to demonstrate it bring the question of the role of temporally patterned activity into the domain of human cognition.

The Brain as a Dynamical Machine

What the brain does is ultimately simple: it takes in sensory information, transforms it into an abstract code of spikes, and uses it to generate motor patterns. This spike code thus constitutes a mental representation of the world, which interacts with memories, expectations, motivations, and other internal states of the animal to generate a series of behaviors that are adaptive and intelligent, and maximize the survival of the individual and the spread of its genes.

Incoming information is encoded in all sensory modalities through temporally modulated patterns of activity, with timescales driven by variations in the world and modified by the dynamical properties of sensory receptors. These inputs are filtered, sorted, and integrated into multimodal, task-relevant representations that likely lose their direct temporal correspondence with instantaneous environmental changes. The importance of timing reemerges, however, in the precise choreography of motor output through which different types of motor units are recruited or turned off to generate smooth behavior. Therefore, the brain transforms timed inputs into timed output.

It is natural, then, to postulate that the internal code of the brain is also a temporal one, whereby dynamic patterns of action potentials represent the computations that create and manipulate abstract categories. These internal dynamics should serve to modulate and integrate time-varying incoming information. Given that sensory receptors often operate with sensitivity at the physical limit (2, 3), one would expect that any ongoing temporal dynamics should be able to receive and preserve this exquisitely precise information. Such a code could form a common currency for different sensory systems and could be harnessed to, or directly drive, different sets of behavioral outputs. Intuitively, one can therefore imagine that the brain acts as a temporal machine that organizes neuronal activity in time. In fact, the biophysical properties of cortical neurons, with strong nonlinear input–output properties near the action potential threshold, and the fact that most excitatory synapses are weak and depressing, make nonsynchronous activity difficult to propagate (4, 5). One then expects to find series of neurons that fire synchronously but at the same time follow a precise temporal pattern, one that would only be evident in the entire activity of the system. The identification and analysis of these precise firing patterns, and an understanding of how they are implemented in circuits, how they are controlled and modulated, and how they encode represented variables could constitute the “cracking of the neural code.”

Evidence for Temporal Sequences of Activity in Experimental Animals

However, is there any evidence for such temporal patterns in internal brain states? Indeed, there have been many experimental findings in the last decades showing precise temporal sequences of spiking in the mammalian brain. Pioneering work was carried out by the group of Moshe Abeles (6), describing precise firing sequences of spikes in the frontal cortex of monkeys during the performance of behavioral tasks. These patterns, recorded with individual electrodes, could stretch over seconds yet had a precision of a few milliseconds, a remarkable property given the short time constants (normally less than tens of milliseconds) associated with most biophysical and synaptic mechanisms. Similar precise firing patterns have been reported in the visual cortex of monkeys, using voltage-sensitive dye imaging (7), in the auditory cortex of rats using multielectrode recordings (8), and in the visual cortex of mouse using calcium imaging (9, 10). Precise firing patterns have even been observed in cortical brain slices (11) and cultures (12), demonstrating that this may be a basic emergent property of any neural circuit. Moreover, these patterns appear to exist at different temporal scales, ranging from milliseconds in the case of monkey recordings to seconds in the case of calcium imaging mouse patterns, representing a broad continuum of temporal structures.

Nevertheless, the existence of these precise firing patterns is controversial. In particular, it has been questioned whether these patterns simply arise by chance, once one considers that there are many neurons generating many spikes. An apparently precise firing sequence could simply be a series of selected spikes, chosen by the experimenter out of a large number of spikes with random timing (13). To test this null hypothesis, one needs a rigorous statistical model to tabulate the number of precise patterns that could arise by chance (14). Although the work cited above (6–12) provided statistical analysis of the data, some skeptics were not convinced. In all truth, it is difficult to provide a perfect statistical model, given that interactions between neurons are likely nonlinear, so the spike-shuffling methods normally used to assess the statistical significance of a temporal pattern, albeit rigorously performed, could be still inadequate to simulate the appearance of patterns from higher order statistical kernels. At the same time, in both brain slices and in vivo, precise firing patterns can be initiated by thalamic or sensory stimulation (9, 10, 15). So to question their reality becomes a moot point when they can be triggered at will, at least for some of these experimental preparations. Moreover, in these slice experiments, the temporal spiking patterns were organized by the switching of cortical neurons between UP and DOWN states (UDSs) (11), so they also have a biophysical signature. Interestingly, UDSs, or similar bistable states, are found in many brain areas under spontaneous activity (16). Although still poorly understood, UDSs have been associated with slow-wave sleep, quiet wakefulness, attentional states, and circuit-level synchronizations (17).

Significantly, in several cases, the temporal dynamics of the activity of a population of neurons can be linked quite deterministically to the behavioral choice of the animal. One example is the decision to swim or to crawl in leeches, which correlates well with the temporal dynamics of population activity in the ganglia of the animal (18). Similarly, decisions made by a mouse during spatial navigation in a virtual reality maze can be recapitulated and even predicted from the multidimensional analysis of the activity of a population of neurons in its parietal cortex (19). Therefore, it is quite likely that these temporal patterns are not only real but are of major functional importance.

Temporal Synchronous Patterns in fMRI Resting-State Activity

Thus, animal work has shown that temporal patterns exist at different scales and can be associated in some cases with UDSs, although the issue of their function has not yet been resolved. With the work of Mitra et al., the investigation of the existence and potential role of temporal dynamics has now moved to human subjects, using fMRI data.

The presence of patterned spontaneous activity in the human brain is not news. Since the first EEG recordings by Berger in 1924, we know that ongoing intrinsic activity is prevalent in all brain areas. In fact, there is not a moment in which any part of the brain is not active. Moreover, the introduction of fMRI enabled the discovery of spontaneous fluctuations in the blood oxygen level-dependent signal that represented an ongoing “resting-state” or “default-network” pattern of activity that involves many brain areas in a concerted fashion (20). This resting-state activity is not random but has spatiotemporal coherence across areas, appears intrinsic to the human brain regardless of behavior, differs between children and adult subjects, and occurs in similar frequency bands as UDSs (21).

Within this context, Mitra et al. have analyzed fMRI signals from over 1,000 subjects and describe specific common temporal dynamics in the resting-state activity. They call these temporal patterns “lag threads,” as a nod to the “threads” of activity in computer processors. To identify these patterns, they perform a dimensionality reduction analysis on the time delays between responses in different areas, applying principal component analysis to the measured lags. This novel technique in principle captures repeated patterns of activity occurring in a non–time-locked fashion. The identified patterns are statistically significant in the sense that they are remarkably conserved between two large subpopulations of patients and are not due to trivial causes like conserved temporal delay between cortical areas. However, like all dimensionality reduction procedures, the analysis method has caveats and can be viewed at this point as a heuristic first approximation to the problem of the identification of spontaneous dynamics. It likely underestimates the temporal structure of the resting-state signals: the exact dimensionality of the set of repeated patterns is surely influenced by details of the analysis; overlapping patterns may not be well separated; and higher-order temporal interactions between areas are not captured, nor are the possibilities of branching or reentrant dynamics. Therefore, it is possible that the actual temporal structure of the resting-state activity could be richer. In fact, it is likely that this will be the first of many studies on the nature and significance of these resting dynamical patterns: Mitra et al. open the door to a new field.

Indeed, the paper raises more questions than it answers. Besides the true spatiotemporal nature and dominance of these patterns, the circuit, synaptic, or cellular mechanisms that generate them are still unknown. Are they related to the UDSs or to the precise firing sequences described in animal preparations? What are the relationships among activity patterns across these very different spatiotemporal scales, ranging from single cells to fMRI voxels and from milliseconds to seconds? Currently, the data do not permit the identification of lag threads in individual subjects; the possibility of finding and interpreting individual differences is intriguing. Furthermore, it is still unclear how they interact with sensory or behavioral evoked patterns. What is their functional relevance, if any?

Functional Significance of Temporal Patterns in Spontaneous Activity

As is often the case in discovery-driven experimental science, these lag threads are a phenomenon looking for a function. Even if one accepts the existence of these temporal patterns, might they be just the “noise in the machine”? The discovery of reproducible spatiotemporal structure now makes that possibility less likely. Moreover, this activity presumably exerts a major metabolic toll on the brain and therefore on the body; from an energetic standpoint alone, it seems unlikely that these temporal resting-state patterns are an epiphenomenon of spurious activity from a heavily connected brain matrix.

So if we assume they have a function, what could it be? One set of possibilities is that they reflect an ongoing local temporal substrate of an internal neural code. If this is the case, they could be related to the implementation of mental states, such as thoughts or percepts. Such activity could encode computational priors, carrying learned information about the world with which to generate internal predictions to be compared against evoked sensory activity. Or perhaps they could be related to motor programs and represent an evolutionary encephalization of fixed action patterns (22). A second set of hypotheses is that the implemented function could be global, such as a signal that percolates through different brain areas, informing them of a common physiological state, for example, behavioral arousal, attention or motivational states, or even cognitive binding signals, such as the ones that must exist to implement the concept of self or the experience of consciousness. A third set of hypotheses could be imagined whereby the patterned spontaneous activity serves a more humble function, such as providing a homeostatic signal for circuit housekeeping, consolidation of recently acquired learning via synaptic plasticity, or even metabolic reactivation of dormant circuits. Finally, it is possible, as the authors suggest, that altered lag thread dynamics could be involved in pathophysiological mechanisms of mental or neurological diseases. These are some of the exciting avenues of future research opened by Mitra et al. As newer techniques are applied to neuroscience, the dynamical nature of brain function will become more evident and the meaning of temporally patterned spontaneous activity will be revealed.