What language do neural data speak?

by Pierpaolo Sorrentino, MD, PhD.

In the last twenty years or so, the widespread availability of computational power, as well as the diffusion of biomedical devices such as EEG, PET, fMRIs, MEG etc. provided the scientific community with an unprecedented opportunity: to observe the activity of the whole brain as it unfolds. Needless to say, the excitement was huge. It always seemed reasonable to state that the coordination of multiple brain regions underpins behavior. But now, one could finally test this! An entire field formed in the attempt of describing such coordination among areas. However, after years where evidence accumulated very quickly, it also became evident that a clear pattern was lacking. Results would not replicate [1]. In a nerve-wracking race, contradictory evidence accumulated, failing to converge into an overarching framework. Let alone translating these techniques to practical applications! In striking contrast with the successes of structural imaging, functional imaging was left behind, with only a marginal role to play in the Neurological diagnosis. And all this despite major efforts…how did that happen? what went wrong?

Quite soon, the fact that the statistical analyses were not appropriate for this kind of large dataset became apparent, and this could explain some of the contradictory results. In general, the reliability of the data (i.e. how likely is it that my results would replicate in a different sample?  Is making inference reasonable?) became a major concern in the community [2]. Importantly, a great deal of attention was dedicated to the problem of multiple comparisons, whereby spurious results become more likely if one performs multiple statistical tests at once [3]. However, recently, more fundamental aspects that flawed the field became apparent: the “validity” of the results. That is, provided I found a consistent and reliable statistical pattern in my functional data (as in, the level of activity of two regions appears to be correlated), what would this mean in terms of physiology? Do such patterns really capture some (patho)physiological events taking place within the brain? If so, which ones? Answering these questions is of paramount importance if any external manipulation (i.e. therapeutic intervention) is to be performed. In fact, at the whole-brain level, each signal that we record is generated by a large population of neurons, with an intricate local geometry and a pattern of regional interactions that unfold at the microscopic scale and cannot be observed directly from outside the brain. However, such regional interactions sum up to provide an “average” local activity, and this is how the signal we are typically dealing with are thought of. Does such “average” convey the meaningful part of the physiological processes occurring microscopically? Since we lack detailed knowledge of how brain signals are generated, large-scale functional patterns need to predict some behavior in order to be considered valid in neurophysiological terms. Only if the presence of a certain statistical pattern in the signals predicts a behavioral outcome, then such activity might be meaningful indeed [4,5]. Degeneracy adds another layer of complication that we will not delve into here [6].

This led to the question, what is the best methodology to extract the activity that is present in the data? In other words, when we look at the data, we are trying to see the traces of what internal mechanisms? What is the right language - the right framework - to listen to what the data are saying?

As said, the most straightforward way to analyze activity is to compute correlations. The reasoning behind this reads something like: “well, if every time region A gets active, so does region B….and every time region A goes to rest, so does region B…maybe these regions are communicating somehow”. As one can see, in this context no statement about the mechanism that might provoke such communication is provided. However, using certain signals (e.g. M/EEG) that are directly generated by neuronal activity, one might choose to go one step further, and study the interactions with the paradigm of synchronization [7]. Synchronization over time needs a certain periodicity in the dynamics to be defined, and it is a mathematical framework that captures if two elements (in this context, brain regions) “wait for each other”. If one thinks of the activity of brain regions as being oscillating like pendula, and if - over time - such pendula end up “swinging” together, then one can infer that some interactions are going on. In this context, one is assuming a more physiologically grounded mechanism in order to interpret coherent large-scale activity [8,9]. More recently, however, a further issue rose, since aperiodic, “scale-free” activity was present in the data [10–12]. If one looked at the data with such new lens, it became apparent that short, aperiodic bursts of activation might carry relevant information about the underlying neuronal microscopic activity, and that such fast transients are not to be discarded, as typically done when assuming periodic activity [13,14]. Interestingly, the spreading of such functional bursts follows the structural tracks linking brain areas that are far apart (connectome) [15], suggesting that the spreading of activity might be indeed mediated by neuronal mechanisms, and convey relevant information [16]. The structural properties of the track affect the velocity of spread of such perturbations, corroborating this interpretation [17]. The variability of such bursts (i.e. the number of reconfigurations over time) has even been used to measure the flexibility of the dynamics and applied as a clinical marker [18]. In this new light, a small perturbation that occurs “rarely”, spreading from one region to another, would be a physiologically meaningful part of the signal [19]. With quite a stretch, one could see such perturbations as analogous of photons, as “messenger particles” or “messenger perturbations”, mediating the interactions among brain areas like photons mediate the electromagnetic interaction.

All in all, what is the best way to describe large-scale interactions from data is not yet clear, and multiple approaches might be necessary. However, the circle is now narrowing down. And we are not far from effectively extracting reliable and valid information from the data, in order to address the many theoretical predictions that are waiting to be tested.

All in all, we work night and day to provide our community with answers, and finally bring at the side of the Neurologist the remarkable body of knowledge that has been created, in order to build the health-care we dream of, where Neurology is capable of early diagnoses and etiological therapies.

We’ll try hard, and get there!!

Bibliography

[1] Ioannidis, J. P. A. Why Most Published Research Findings Are False. PLoS Med. 2, e124 (2005).

[2] Zuo, X.-N., Xu, T. & Milham, M. P. Harnessing reliability for neuroscience research. Nat. Hum. Behav. 3, 768–771 (2019).

[3] Chen, X., Lu, B. & Yan, C. Reproducibility of R‐fMRI metrics on the impact of different strategies for multiple comparison correction and sample sizes. Hum. Brain Mapp. 39, 300–318 (2017).

[4] Finn, E. S. & Rosenberg, M. D. Beyond fingerprinting: Choosing predictive connectomes over reliable connectomes. NeuroImage 239, 118254 (2021).

[5] Sorrentino, P. et al. Clinical connectome fingerprints of cognitive decline. doi:10.1101/2020.10.09.332635.

[6] Jirsa, V. Structured Flows on Manifolds as guiding concepts in brain science. in Selbstorganisation – ein Paradigma für die Humanwissenschaften 89–102 (Springer Fachmedien Wiesbaden, 2020). doi:10.1007/978-3-658-29906-4_6.

[7] Buzsaki, G. Neuronal Oscillations in Cortical Networks. Science 304, 1926–1929 (2004).

[8] Stam, C. J. Modern network science of neurological disorders. Nat. Rev. Neurosci. 15, 683–695 (2014).

[9] Engel, A. K., Fries, P. & Singer, W. Dynamic predictions: Oscillations and synchrony in top–down processing. Nat. Rev. Neurosci. 2, 704–716 (2001).

[10] Haldeman, C. & Beggs, J. M. Critical Branching Captures Activity in Living Neural Networks and Maximizes the Number of Metastable States. Phys. Rev. Lett. 94, 058101 (2005).

[11] Esfahlani, F. Z. et al. High-amplitude cofluctuations in cortical activity drive functional connectivity. Proc. Natl. Acad. Sci. 117, 28393–28401 (2020).

[12] Faskowitz, J., Esfahlani, F. Z., Jo, Y., Sporns, O. & Betzel, R. F. Edge-centric functional network representations of human cerebral cortex reveal overlapping system-level architecture. bioRxiv 799924 (2019) doi:10.1101/799924.

[13] Beggs, J. M. & Plenz, D. Neuronal Avalanches Are Diverse and Precise Activity Patterns That Are Stable for Many Hours in Cortical Slice Cultures. J. Neurosci. 24, 5216–5229 (2004).

[14] Shriki, O. et al. Neuronal avalanches in the resting MEG of the human brain. J. Neurosci. 33, 7079–7090 (2013).

[15] Sorrentino, P. et al. The structural connectome constrains fast brain dynamics. eLife 10, e67400 (2021).

[16] Benisty, H. et al. Rapid fluctuations in functional connectivity of cortical networks encode spontaneous behavior. 2021.08.15.456390 https://www.biorxiv.org/content/10.1101/2021.08.15.456390v2 (2021) doi:10.1101/2021.08.15.456390.

[17] Sorrentino, P. et al. On the topochronic map of the human brain dynamics. 2021.07.01.447872 https://www.biorxiv.org/content/10.1101/2021.07.01.447872v3 (2021) doi:10.1101/2021.07.01.447872.

[18] Sorrentino, P. et al. Flexible brain dynamics underpins complex behaviours as observed in Parkinson’s disease. Sci. Rep. 11, 4051 (2021).

[19] Bansal, K. et al. Scale-specific dynamics of high-amplitude bursts in EEG capture behaviorally meaningful variability. NeuroImage 241, 118425 (2021).