Neural Coding With Bursts: Single Cell Computation vs Population Coding

The neural network can be thought of as the computational hardware of the brain whilst the “neural code” can be thought of as the software. The structure of the former i.e the neural network, continues to be probed via brain histology methods and a lot has been learnt about the connectivity of neurons in the human brain. Although a lot has been revealed regarding the electrical properties of neurons including their communication using action potentials, quite a lot more remains to be known regarding how neural networks actually encode information. From memory recall to carrying out intricate body manoeuvres, the neural code that orchestrates such coordinated activity is yet to be truly unravelled. Some of the suggestions are that information is encoded in the firing rates of individual neurons. To study such firing rates of neurons, a number of techniques have been developed. One of these methods is intracellular electrode recordings. Such recordings have the temporal resolution required to capture firing dynamics of individual neurons.

The advantages of intracellular recordings are that unlike neural population recordings, there is no ambiguity about the source of the action potentials, the identity of each firing neuron is known. The other advantage is that such close monitoring of a neuron allows us to capture the firing rate behaviour of each neuron under investigation. One such firing behaviour is “burst” firing [1]. Burst firing is when a neuron fires a group of action potentials generated in rapid succession, followed by a period of relative quiescence. In some areas of the brain such as the cerebellum, bursting seems to be an important computational property of a specific type of cell; the granule cells [2] and so considerable effort has been made to develop ways that capture the bursting dynamics of individual neurons.

However, although single cell intracellular recordings can capture firing rate dynamics with such high temporal resolution, they cannot capture population dynamics because only a small number of neurons can be recorded simultaneously this way. It is the ensemble of neural activity of neural populations that give rise to the neural code required to execute cognitive functions and other complex tasks [3]. It is for this reason that it important to study the firing activity of populations of neurons simultaneously as opposed to single cell recordings or recording of a handful of cells using intracellular recording methods. However any method that aims to measure population activity has to somehow find a way to offset the emergent action potential source-ambiguity that plagues population recording.

A number of population recording techniques have been developed and each of them implement various techniques to remove source ambiguity in the recordings. These include, fMRI recordings, calcium imaging, 2-photon imaging and the use of multi-site electrode arrays. The fMRI bold signal has been used to infer neuronal brain activity by monitoring the flow of blood to brain regions. Active brain regions need an increased blood supply and so the distribution of brain activity can be imaged this way [4]. The fMRI bold signal however lacks the temporal resolution required to characterise individual neuron activity and their individual contribution to brain activity [5]. This means that although fMRI reveals distribution of brain activity, its temporal resolution is too coarse to make out individual neuronal contributions to the signal [7]. Thus distinct firing patterns of individual neurons cannot be made out and their individual contributions to the output signal cannot be tracked.

Two -photon Calcium imaging is a population recording technique with a higher temporal resolution than fMRI in that single spikes can be resolved [11]. It involves using animal strains that have been genetically modified to encode a calcium indicator that causes the cells to fluoresce when they undergo spiking activity [12]. This way, the activity of a large population can be imaged both in vitro and in vivo. The temporal resolution of calcium imaging is however still not high enough to quantify burst spiking where a neuron’s action potentials follow one another in very quick succession [11]. A number of methods however abound to increase the resolution capability of calcium imaging by reconstructing bursting activity using different iterations of what is essentially the template matching technique for reconstructing spiking activity from calcium traces.

Multi-site electrode arrays (MEA’s) have an even greater temporal resolution than the above mentioned imaging techniques [6]. Like the use of the fMRI bold signal, MEA’s monitors the spiking activity of large neuronal populations. But instead of relying on tracking temporally slow changes in blood flow like fMRI, MEA’s record neuronal action potential spikes. A single electrode in an array records the action potentials from neurons in the nearby vicinity. This way, each electrode in the array has a radius from which it will pick up action potentials. The recording radii of the electrodes will overlap to a degree, and it is this overlap that helps in the triangulation and identification of neurons responsible for spikes. The shape and structure of action potential waveforms is also used in the spike sorting algorithm to help match action potentials to the neurons responsible. High-band filtering removes the lower frequency Local Field Potential (LFP) as well as low frequency pulses from neurons that are too far away from the recording electrodes to be triangulated. Even more analysis is required to be able to identify and distinguish bursts from spikes. From MEA recordings, some techniques involve using a combination of Inter-Spike Intervals (ISI) and expected spike number to determine to statistically distinguish spikes from bursts [8], where the probability of bursting can be expressed as a poisson distribution. Several burst detection algorithms abound [9] and because of the variable mechanisms that give rise to bursts it is not possible to have a singular classification algorithm for detecting them [10].

Before use of the newer computer algorithms such as Principal Component Analysis (PCA), variability in spike sorting was also due to the researcher’s level of experience and objective judgement in choosing what waveform characteristics to take into account during the spike sorting process [13]. Nowadays these features are chosen by algorithms. The idea behind principal component analysis (PCA) is to find an ordered set of orthogonal basis vectors that capture the directions in the data of largest variation [14]. The data are the original spikes from the recorded waveform. Principal components analysis (PCA) performs a linear transformation on the spike recordings and is used to reduce multidimensionality of spike data down to a few dimensions for easier analysis and the redundant vectors dimensions are removed. PCA transforms the data so that as much variation as possible will be crammed into the fewest possible dimensions; allowing the compression of data by ignoring other dimensions. Whilst this method has been shown to have great accuracy in assigning spiking activity to neurons, the shortfall of it is that such dimensionality reduction blurs the distinction between spikes and bursts [15]

The challenges of spike sorting may be met by simultaneous recording of both population activity as well as intracellular recordings of some of the cells in the population [16]. The idea behind this is that ultimately there is no present population recording equipment or method that has the kind of resolution that currently only be achieved by single cell recording [17]. I their paper, Vardi et tal. [16] demonstrate how intracellular recording of a single neuron can give a reliable estimation of macroscopic properties of the neural network. Because of the currently unbeatable resolution of intracellular recording, it makes sense to put effort in developing population recording of in-vivo multi neuron intracellular recording. The development of “auto-patching robots” for the simultaneous whole cell recording of in vivo neurons [18] accomplishes both the temporal and spatial resolution to characterise complex spiking patterns such as bursting; as well as accomplishing from recording from a larger population of neurons using its fully automated robot guided patch clamping arrays. Using such a technique simultaneously with larger population recording techniques like two-photo imaging and or MEA’s will likely give even greater results regarding spike assignment and discrimination without losing out on capturing the cooperative dynamics of the neural network.

References

[1] Wagenaar, D. A., Pine, J., and Potter, S. M. (2006). An extremely rich repertoire of bursting patterns during the development of cortical cultures. BMC Neurosci. 7:11

[2] D’Angelo, E., Nieus, T., Maffei, A., Armano, S., Rossi, P., Taglietti, V., et al. (2001). Theta-frequency bursting and resonance in cerebellar granule cells: experimental evidence and modeling of a slow K+-dependent mechanism. J. Neurosci. 21, 759–770

[3] Stefano Panzeri, Jakob H.Macke, Joachim Gross, Christoph Kayser (2015). Neural population coding: combining insights from microscopic and mass signals. Trends in cognitive sciences. 19 (3): 162-172.

[4] Raichle ME. 2015. The restless brain: how intrinsic activity organizes brain function. Philos Trans R Soc Lond B Biol Sci. 370:20140172

[5] Michael Okun, Nicholas A. Steinmetz, Armin Lak, Martynas Dervinis and Kenneth D. Harris (2019) Distinct Structure of Cortical Population Activity on Fast and Infraslow Timescales. Cerebral cortex. 29: 2196–2210.

[6] Jiangang Du, Ingmar H. Riedel-Kruse, Janna C. Nawroth, Michael L. Roukes (2009) High-Resolution Three-Dimensional Extracellular Recording of Neuronal Activity With Microfabricated Electrode Arrays. Journal of neurophysiology.

[7] Logothetis NK, Pauls J, Augath M, Trinath T, Oeltermann A. Neurophysiological investigation of the basis of the fMRI signal. Nature 412: 150–157, 2001.

[8] Chiappalone, M., Novellino, A., Vajda, I., Vato, A., Martinoia, S., and van Pelt, J. (2005). Burst detection algorithms for the analysis of spatio-temporal patterns in cortical networks of neurons. Neurocomputing 65–66, 653–662. doi: 10.1016/j.neucom.2004.10.094

[9] Martens, M. B., Chiappalone, M., Schubert, D., and Tiesinga, P. H. E. (2014). Separating burst from background spikes in multichannel neuronal recordings using return map analysis. Int. J. Neural Syst. 24:1450012. doi: 10.1142/s0129065714500129

[10] Fleur Zeldenrust1, Wytse J. Wadman and Bernhard Englitz (2018) Neural Coding With Bursts—Current State and Future Perspectives. Front. Comput. Neurosci. doi.org/10.3389/fncom.2018.00048

[11] Tingwei Quan, Xiaohua Lv, Xiuli Liu and Shaoqun Zeng (2016) Reconstruction of burst activity from calcium imaging of neuronal population via Lq minimization and interval screening. Optical Society of America. 7(6):2103-2117

[12] Akinori Mitani and Takaki Komiyama (2018) Real-Time Processing of Two-Photon Calcium Imaging Data Including Lateral Motion Artifact Correction. Front. Neuroinform. doi.org/10.3389/fninf.2018.00098

[13] F. Wood, M.J. Black, C. Vargas-Irwin, M. Fellows, J.P. Donoghue. On the variability of manual spike sorting IEEE Trans. Biomed. Eng., 51:912-918

[14] M.S. Lewicki (1998)A review of methods for spike sorting: the detection and classification of neural action potentialsNetw. Comput. Neural. Syst., 9:R53-R78

[15] Takashi Takekawa, Yoshikazu Isomura and Tomoki Fukai (2012) Spike sorting of heterogeneous neuron types by multimodality-weighted PCA and explicit robust variational Bayes. Front. Neuroinform. doi.org/10.3389/fninf.2012.00005

[16] Roni Vardi, Amir Goldental, Shira Sardi, Anton Sheinin & Ido Kanter (2016) Simultaneous multi-patch-clamp and extracellular-array recordings: Single neuron reflects network activity. Scientific reports. 6: 36228

[17] Marx, V. Neurobiology: rethinking the electrode. Nature methods. 11, 1099–1103 (2014)

[18] Suhasa B Kodandaramaiah, Francisco J Flores et tal. (2018) Multi-neuron intracellular recording in vivo via interacting autopatching robots. eLife. 7:e24656 DOI: 10.7554/eLife.24656

#computationalneuroscience #burstingneuron #Principalcomponentanalysis #twophotoncalciumimaging